CN115718263A - Attention-based lithium ion battery calendar aging prediction model and method - Google Patents
Attention-based lithium ion battery calendar aging prediction model and method Download PDFInfo
- Publication number
- CN115718263A CN115718263A CN202310023441.2A CN202310023441A CN115718263A CN 115718263 A CN115718263 A CN 115718263A CN 202310023441 A CN202310023441 A CN 202310023441A CN 115718263 A CN115718263 A CN 115718263A
- Authority
- CN
- China
- Prior art keywords
- attention
- model
- prediction
- module
- driven
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 230000032683 aging Effects 0.000 title claims abstract description 61
- HBBGRARXTFLTSG-UHFFFAOYSA-N Lithium ion Chemical compound [Li+] HBBGRARXTFLTSG-UHFFFAOYSA-N 0.000 title claims abstract description 26
- 229910001416 lithium ion Inorganic materials 0.000 title claims abstract description 26
- 238000000034 method Methods 0.000 title claims abstract description 24
- 230000015654 memory Effects 0.000 claims abstract description 13
- 230000007246 mechanism Effects 0.000 claims abstract description 12
- 238000012544 monitoring process Methods 0.000 claims abstract description 7
- 238000012360 testing method Methods 0.000 claims description 57
- 238000012549 training Methods 0.000 claims description 25
- 125000003821 2-(trimethylsilyl)ethoxymethyl group Chemical group [H]C([H])([H])[Si](C([H])([H])[H])(C([H])([H])[H])C([H])([H])C(OC([H])([H])[*])([H])[H] 0.000 claims description 24
- 238000004626 scanning electron microscopy Methods 0.000 claims description 24
- 230000006870 function Effects 0.000 claims description 20
- 239000013598 vector Substances 0.000 claims description 12
- 238000004458 analytical method Methods 0.000 claims description 11
- 238000002679 ablation Methods 0.000 claims description 10
- 230000015556 catabolic process Effects 0.000 claims description 10
- 238000006731 degradation reaction Methods 0.000 claims description 10
- 239000011159 matrix material Substances 0.000 claims description 9
- 238000011156 evaluation Methods 0.000 claims description 6
- 238000007781 pre-processing Methods 0.000 claims description 6
- 238000002474 experimental method Methods 0.000 claims description 4
- 238000012795 verification Methods 0.000 claims description 4
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 claims description 3
- 230000001419 dependent effect Effects 0.000 claims description 3
- 238000013507 mapping Methods 0.000 claims description 3
- 238000010200 validation analysis Methods 0.000 claims description 3
- 230000000694 effects Effects 0.000 claims description 2
- 238000005842 biochemical reaction Methods 0.000 claims 1
- 238000005259 measurement Methods 0.000 claims 1
- 230000036541 health Effects 0.000 abstract description 7
- 238000013528 artificial neural network Methods 0.000 abstract description 4
- 230000001149 cognitive effect Effects 0.000 abstract 1
- 230000002035 prolonged effect Effects 0.000 abstract 1
- 238000010586 diagram Methods 0.000 description 7
- 230000006872 improvement Effects 0.000 description 5
- 230000008569 process Effects 0.000 description 4
- 230000009286 beneficial effect Effects 0.000 description 2
- 238000006243 chemical reaction Methods 0.000 description 2
- 230000000295 complement effect Effects 0.000 description 2
- 230000001351 cycling effect Effects 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 230000003071 parasitic effect Effects 0.000 description 2
- 238000012545 processing Methods 0.000 description 2
- 230000000306 recurrent effect Effects 0.000 description 2
- 230000006403 short-term memory Effects 0.000 description 2
- 238000013526 transfer learning Methods 0.000 description 2
- 230000001133 acceleration Effects 0.000 description 1
- 230000006978 adaptation Effects 0.000 description 1
- 230000006399 behavior Effects 0.000 description 1
- 230000002457 bidirectional effect Effects 0.000 description 1
- 230000019771 cognition Effects 0.000 description 1
- 238000013527 convolutional neural network Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 230000009977 dual effect Effects 0.000 description 1
- 230000007787 long-term memory Effects 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000013508 migration Methods 0.000 description 1
- 230000005012 migration Effects 0.000 description 1
- 230000003278 mimic effect Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 230000004044 response Effects 0.000 description 1
- 238000013179 statistical model Methods 0.000 description 1
Images
Classifications
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02E—REDUCTION OF GREENHOUSE GAS [GHG] EMISSIONS, RELATED TO ENERGY GENERATION, TRANSMISSION OR DISTRIBUTION
- Y02E60/00—Enabling technologies; Technologies with a potential or indirect contribution to GHG emissions mitigation
- Y02E60/10—Energy storage using batteries
Landscapes
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses an Attention-based lithium ion battery Calendar Aging prediction model and method (KDACAF), which comprises a semi-experience module (SEM module), a Knowledge-driven Attention module, a Data-driven Attention module and a long-time and short-time memory module (LSTM module). According to the lithium ion battery calendar aging prediction model and method based on attention, the knowledge-driven attention module takes a semi-empirical module based on the Alonenius law as a front end, electrochemical priori knowledge in the battery field is integrated into a data-driven neural network, the attention mechanism is applied to lithium ion battery calendar aging prediction by taking human cognitive decision mechanism as reference, monitoring and management of the health state of the battery are facilitated, and the service life of the battery is prolonged.
Description
Technical Field
The invention relates to a battery technology, in particular to a lithium ion battery calendar aging prediction model and a lithium ion battery calendar aging prediction method based on attention.
Background
Lithium ion batteries have been widely used in many industrial electronic applications, such as Electric Vehicles (EVs) and smart grids, due to their advantages in energy density and low self-discharge rate. However, effective health management is a key and challenging problem for the wider application of lithium ion batteries. In practical applications such as electric vehicles, batteries are degraded by calendar (calendar) and cycling (cycling). Since over 70% of automotive battery life is spent in storage conditions, there is an urgent need for an effective battery health monitoring and management solution in the calendar degradation mode.
In the battery calendar mode, the rate of decay of the battery capacity may be significantly affected by several factors, including the storage temperature and the battery State of charge (SoC). Since battery calendar degradation is a highly nonlinear and strongly coupled process, developing an appropriate model to diagnose/describe battery capacity degradation behavior under different storage conditions while considering the effects of storage temperature and SoC is crucial to battery health monitoring and management. At present, battery calendar aging prediction models can be divided into two categories, namely a knowledge-driven model and a data-driven model.
For knowledge-driven models, some knowledge that can reflect the battery degradation mechanism will be coupled into the model to account for the dynamics of battery aging. On the other hand, based on some knowledge information, such as the einlin acceleration equation or arrhenius law, a semi-empirical model is made another common knowledge-driven model to capture the dynamics of battery calendar aging.
With the rapid development of machine learning and cloud computing technologies, a data-driven model has become another common tool, and enables battery health estimation and prediction. This type of model can be further divided into traditional statistical models and deep learning based models. The former has a relatively small capacity, such as Support Vector Regression (SVR), gaussian Process Regression (GPR), and the like. The latter adopts a large-capacity neural network with a deep structure, such as a convolutional neural network, a recurrent neural network, a deep belief network, and a plurality of mixed models combining the recurrent neural network and the transfer learning.
Currently, one challenge for battery calendar aging prediction by data-driven models is: without the guidance of the battery electrochemical experience knowledge, the pure data-driven model mainly learns the battery aging information from the training data, and the generalization is poor. Since battery electrochemical knowledge can support calendar aging modeling, incorporating battery electrochemical empirical knowledge into a data-driven model should have significant potential in improving predictive performance, especially under brand new operating conditions lacking historical data. However, existing data-driven models have limited ability to incorporate a priori knowledge and statistical rules from different modalities, i.e., to process multimodal input. Recently proposed attention mechanisms have partially addressed the multi-modal processing problem. The attention mechanism is to mimic the cognitive process of a human being, i.e. to selectively focus on one or several things, while ignoring other things like self-attention, global/soft attention, local/hard attention, etc. The forecasting field also makes multiple attempts on the attention mechanism, such as proposing a hybrid attention-long short-term memory (LSTM) model for photovoltaic power forecasting, combining the attention mechanism and a bidirectional LSTM for power load forecasting, and the like.
Even though these attention-based predictive models have emerged, we believe that effective improvements and performance enhancements can still be achieved because these models are based on purely data-driven models with less consideration for incorporating domain or expertise into the model.
Disclosure of Invention
In response to the above problems, an attention-based lithium ion battery calendar aging prediction model (KDACAF) is designed in this application, which contains two attention modules, namely knowledge-driven attention and data-driven attention. The knowledge-driven attention module takes a semi-empirical model as a front end and fully utilizes the battery aging electrochemical empirical knowledge. Because electrochemical knowledge can guide battery calendar aging prediction modeling, introducing priori knowledge in KDACAF brings remarkable performance improvement for model prediction. And because the proposed knowledge-data driven attention model consists of data driven and semi-empirical modules, the application mainly compares and evaluates the performance of the model with other classical data driven models and semi-empirical models.
In order to achieve the purpose, the invention provides an attention-based lithium ion battery calendar aging prediction model, which comprises a semi-experience module, a knowledge-driven attention module, a data-driven attention module and a long-time and short-time memory module;
the knowledge-driven attention module takes a semi-empirical model as a front end, and the semi-empirical model is based on the arrhenius law.
The invention also includes an experimental platform for testing the effectiveness of KDACAF, comprising a hotcell for controlling the ambient temperature of a storage battery, a battery test device for maintaining a predetermined storage state of charge (SoC) level for the battery, a computer for monitoring and storing battery aging data.
The attention-based method for the lithium ion battery calendar aging prediction model comprises the following steps:
s1, collecting and preprocessing data;
s2, establishing a calendar aging prediction model;
s21, establishing a problem description and a semi-empirical model for calendar aging prediction;
s22, KDACAF structure;
s23, memorizing the loss function of the module and KDACAF in terms of length and time;
s3, experiments and analysis
S31, comparing and testing;
s32, ablation testing;
and S33, convergence analysis.
The step S1 specifically includes the following steps:
data preprocessing and evaluation index
Performing cubic spline interpolation on the training set to make each training setThe battery capacity sequences all have a resolution of 1 hour, so that the length of each sequence is 11521, and in addition, in order to ensure the time resolution consistency of KDACAF on the training set, the verification set and the test set, all the capacity sequences in the training set are sparsely sampled once every 30 days, namely, all the capacity sequences in the KDACAF training period
Of (2) is arbitrary twice successively
And
the time interval between is still 30 days, i.e.
Hours, which is the same as the time resolution of the validation set and test set; the Maximum Absolute Error (MAE) and Root Mean Square Error (RMSE) were used for evaluation, i.e.:
step S21 specifically includes the following steps:
s211, problem description of calendar aging prediction
Calendar aging prediction of lithium ion batteries aims at predicting the variation of their capacity with storage time, the interpretation variables for this task are formatted as the following matrix:
wherein,
all the interpretation information of the present prediction task is represented,
the sequential order number representing the capacity sequence,
indicating a sequence of capacities
The number of the measured values is,
to represent
The storage time of the battery in (1),
which is indicative of the temperature at which the battery is stored,
which represents the storage of the SoC,
is a lag interval;
for capacity prediction for a single step, the target variable is
Thus, the calendar aging prediction task is abstracted to the following mapping:
wherein,
representing data-driven or knowledge-driven predictive models, in which
Does not necessarily require the use of
All of the information in (1). For example: using only semi-empirical models
Storage time, temperature and SoC in (1), while pure data-driven models tend to use
All of the information in (a).
S212, semi-empirical model for calendar aging prediction
According to arrhenius' law, the degradation of calendar capacity is ultimately expressed in a particular semi-empirical form:
wherein,
and
are all the dependent items of the SoC,
which represents the constant of the gas,
a low power parameter representing a duration dependency.
Considering that the influence of the storage SoC on parasitic chemical reactions may lead to a degradation of the battery capacity, the linear and exponential dependence of the SoC is considered, i.e.
And
in the form of (a), the following semi-empirical model was constructed.
All of the SEMs mentioned above can be respectively and simply expressed as
，
，
，
Wherein
Respectively representing all parameters to be identified, namely:
step S22 KDACAF structure
KDACAF takes a semi-empirical model as a basis from which three branches are extracted, namely a prediction branch, a fitting branch and a characteristic branch. Predicting branches through the four SEM pairs
Preliminary predictions were made and the fit branch was used to solve the regression results of the SEMs on past volume sequences (which served as input to the knowledge-driven attention module), and the feature branch constructed feature vectors based on the four SEMs described above (which served as input to the data-driven attention module).
Step S22 specifically includes the following steps:
s221, semi-empirical module and prediction branch
The semi-empirical module is a collection of four SEMs set forth in S212.
In the case of a predicted branch, the branch is predicted,
the four SEM's input to the semi-empirical module were obtained separately
The regression results, i.e., the preliminary prediction results, of (1) are as follows:
that is, for each SEM there is:
s222, fitting branch and knowledge-driven attention module
To fit the results of past capacity sequences, first, one will fit
The interpretation matrix is divided into two parts:
namely:
in the formula,
a sequence of the recent capacity is represented,
each column in (1) represents
Influence factors of the corresponding elements in (1); in the fitting branch
Respectively input into the four SEM devices to obtain
The regression results of (1):
the present knowledge-driven attention module is dedicated to calendar aging prediction tasks, where targets are predicted
Attention to each SEM (i.e., for its preliminary predicted results)
Note of (d) was designed based on the goodness of fit of each SEM model to the past real capacity, i.e.;
first, the knowledge-driven attention model is
And
the score function between is defined as follows:
can then obtain
Attention to each SEMIs composed of
Then, get to
The following points of (1) are accurately predicted:
s223, feature branch and data-driven attention module
To expand the model capacity of KDACAF to better capture the multimodalities of the calendar aging process, data-driven attention was constructed, the details of which are as follows:
to construct feature vectors from semi-empirical modules, we will
Again, the four SEMs were entered and four eigenvectors were constructed from the intermediate variables of these SEMs, respectively, as follows:
wherein,
representation is based on
The intermediate variables in the inner.
Predicting a target in a data-driven attention module
Attention to each SEM (i.e., preliminary prediction of its outcome)
Attention of (d) is designed based on the historical capacity sequence and the constructed features, and the scoring function is defined as:
wherein
A scoring function representing the data-driven attention,
is the first
A trainable matrix corresponding to the SEM;
then, can obtain
Attention number to each SEM, note
I.e. by
Then, get to
Another accurate prediction of:
step S23 long-and-short term memory module and KDACAF loss function
The comprehensive two-part attention mechanism can be obtained
The intermediate prediction of (2):
wherein
Showing two attention module pairs
The inter-prediction of (2) is performed,
and
all are learnable weights;
then, the historical capacity is sequenced
And
the concatenation is as follows vector:
finally, the vectors are input to a long-short term memory module comprising an LSTM layer, and the LSTM output is scaled
In front of and behind
Is expressed as
；
The loss function of KDACAF is set as follows:
in the formula,
for training the number of samples, adopting an Adam training method to train KDACAF;
step S31 specifically includes the following steps:
s311, comparison in test Set \822456;
s312, comparison in test Set \8225j.
Therefore, the invention has the following beneficial effects by adopting the structure:
1. by taking electrochemical knowledge as a key basis of a knowledge-driven attention module, the KDACAF realizes accurate battery calendar aging prediction based on knowledge-data dual drive. Ablation experiments show that the introduction of knowledge in the electrochemical field obviously improves the prediction performance of KDACAF.
2. Multiple comparison tests show that KDACAF is superior to the current most advanced knowledge-driven and data-driven battery calendar aging prediction model.
The technical solution of the present invention is further described in detail by the accompanying drawings and embodiments.
Drawings
FIG. 1 is a block diagram of the present invention;
FIG. 2 is a block diagram of the experimental platform structure of the present invention;
FIG. 3 is a data processing flow diagram of the present invention;
FIG. 4 is a long and short term memory module diagram of the present invention;
FIG. 5 is a graph showing the relationship between the prediction results and the corresponding prediction errors in the test set \822458;
FIG. 6 is a graph of the relationship between the prediction results and the corresponding prediction errors in the test set \8225; according to the present invention;
FIG. 7 is a graph of the convergence and stability analysis of the present invention on a test set \8224;
FIG. 8 is a graph of the convergence and stability analysis of the present invention on test set \8225.
Detailed Description
The present invention will be further described with reference to the accompanying drawings, and it should be noted that the present embodiment is based on the technical scheme, and a detailed implementation manner and a specific operation process are provided, but the protection scope of the present invention is not limited to the present embodiment.
Fig. 1 is a schematic structural diagram of an embodiment of the present invention, and as shown in fig. 1, the structure of the present invention includes a semi-empirical module, a knowledge-driven attention module, a data-driven attention module, and a long-time and short-time memory module;
the knowledge-driven attention module takes a semi-empirical model as a front end, and the semi-empirical model is based on the arrhenius law.
The invention also includes an experimental platform for testing the validity of KDACAF, comprising a hotroom for controlling the ambient temperature of a storage battery, a battery testing device for maintaining a predetermined storage state of charge (SoC) level for the battery, a computer for monitoring and storing battery aging data.
The attention-based method for the lithium ion battery calendar aging prediction model comprises the following steps:
s1, collecting and preprocessing data;
TABLE 1 calendar aging data set
The step S1 specifically includes the following steps:
data preprocessing and evaluation index
Before KDACAF training, 27 capacity sequences are divided into four subsets, namely sequences 13, 14, 16, 17, 22, 23, 25 and 26 are used as training sets within 8000 hours, sequences 15, 18, 24 and 27 are used as verification sets within 8000 hours, and Con. #5, 6, 8 and 9 are used as test sets with time stamps exceeding 8000 hours, and Con. #1, 2, 3, 4 and 7 are used as test sets set \822425.
Carrying out cubic spline interpolation on a training set to ensure that each battery capacity sequence in the training set has 1 hour resolution, the length of each sequence is 11521, and in addition, in order to ensure the time resolution consistency of KDACAF on the training set, the verification set and the test set, all capacity sequences in the training set are sparsely sampled once every 30 days, namely all capacity sequences in the KDACAF training period
Of (2) is arbitrary twice successively
And
the time interval between is still 30 days, i.e.
Hours, which is the same as the time resolution of the validation set and test set; the Maximum Absolute Error (MAE) and Root Mean Square Error (RMSE) were used for evaluation, i.e.:
s2, establishing a calendar aging prediction model;
s21, establishing a problem description and a semi-empirical model for calendar aging prediction;
step S21 specifically includes the following steps:
s211, problem description of calendar aging prediction
Calendar aging prediction of lithium ion batteries aims at predicting the variation of their capacity with storage time, the interpretation variables for this task are formatted as the following matrix:
wherein,
it is meant that all of the interpretation information,
the sequential order number representing the capacity sequence,
represents the second of a capacity sequence
The number of the measured values is,
to represent
The storage time of the battery in (1),
which is indicative of the temperature at which the battery is stored,
representing the stored state of charge (SoC),
is a lag interval;
for capacity prediction for a single step, the target variable is
Thus, the calendar aging prediction task is abstracted as the following mapping:
wherein,
representing data-driven or knowledge-driven predictive models, in which
Does not necessarily require the use of
All of the information in (1). For example: using only semi-empirical models
Storage time, temperature and SoC in (1), while pure data-driven models tend to use
All of the information in (a).
S212, semi-empirical model of calendar aging prediction
According to arrhenius' law, the degradation of calendar capacity is ultimately expressed in a particular semi-empirical form:
wherein,
and
are all the dependent items of the SoC,
which represents the constant of the gas,
a low power parameter representing a duration dependency.
Considering that the influence of the storage SoC on parasitic chemical reactions may lead to degradation of the battery capacity, linear and exponential dependencies of the SoC are considered, i.e.
And
the following semi-empirical model was constructed.
All of the SEMs mentioned above can be respectively and simply expressed as
，
，
，
Wherein
Respectively representing all parameters to be identified, namely:
s22, KDACAF structure;
KDACAF takes a semi-empirical model as a basis from which three branches are extracted, namely a prediction branch, a fitting branch and a characteristic branch. Predicting branches through the four SEM pairs
Preliminary predictions were made and the fit branch was used to solve the regression results of the SEMs on past volume sequences (which served as input to the knowledge-driven attention module), and the feature branch constructed feature vectors based on the four SEMs described above (which served as input to the data-driven attention module).
Step S22 specifically includes the following steps:
s221, semi-empirical module and predicted branch
The semi-empirical module is a set of four SEMs set forth in S212.
In the case of a predicted branch, the branch is predicted,
is input to the above-mentioned semi-empirical moduleSEM, respectively obtaining
The regression results, i.e., the preliminary prediction results, of (1) are as follows:
that is, for each SEM there is:
s222, fitting branch and knowledge-driven attention module
To fit the results of past capacity sequences, first, one will fit
The interpretation matrix is divided into two parts:
namely:
in the formula,
a sequence of the recent capacity is represented,
each column in (1) represents
Influence factors of the corresponding elements in (1); in the fitting branch
Are respectively provided withThe input of the above four SEM, obtain
The regression results of (1):
the present knowledge-driven attention module is dedicated to calendar aging prediction tasks, where targets are predicted
Attention to each SEM (i.e., for its preliminary predicted results)
Note of (d) was designed based on the goodness of fit of each SEM model to the real capacity in the past, i.e.;
first, the knowledge-driven attention model is
And
the score function between is defined as follows:
can then obtain
Attention to each SEM, note
Then, get to
The following points of (1) are accurately predicted:
s223, feature branch and data-driven attention module
To expand the model capacity of KDACAF to better capture the multimodalities of the calendar aging process, data-driven attention was constructed as follows:
to construct feature vectors from semi-empirical modules, we will
Again, the four SEMs were entered and four eigenvectors were constructed from the intermediate variables of these SEMs, respectively, as follows:
wherein,
representation is based on
The intermediate variables in the inner.
Predicting a target in a data-driven attention module
Attention to each SEM (i.e., preliminary prediction of its results
Attention of (d) is designed based on the historical capacity sequence and the constructed features, and the scoring function is defined as:
wherein
A scoring function representing the data-driven attention,
is the first
A trainable matrix corresponding to each SEM;
then, can be obtained
Attention number to each SEM, note
I.e. by
Then, get to
Another accurate prediction of:
s23, memorizing the loss function of the module and KDACAF in terms of length and time;
the comprehensive two-part attention mechanism can be obtained
The intermediate prediction of (2):
wherein
Representing two attention module pairs
The inter-prediction of (2) is performed,
and
all are learnable weights;
then, the historical capacity is sequenced
And
the concatenation is the following vector:
finally, the vector is input to a long-short-term memory module comprising an LSTM layer, and the output of the LSTM is scaled to
After as
Is expressed as
；
The loss function of KDACAF is set as follows:
in the formula,
for training the number of samples, adopting an Adam training method to train KDACAF;
s3, experiments and analysis
S31, comparing and testing;
step S31 specifically includes the following steps:
s311, comparison in test Set \822456;
s312, comparison in test Set \8225.
S32, ablation testing;
and S33, convergence analysis.
Testing a knowledge-driven method and a data-driven method respectively based on the following models;
models participating in comparison: four SEM (parameters of which are determined by a biophysical optimization algorithm BBO), SVR, GPR, deep LSTM (DLSTM) constructed by stacking a plurality of LSTM layers, LSTM + fully-connected layer + transfer learning (LSTM-FC-TL) mixed model. Where SVR, GPR and DLSTM are data driven, all SEMs are knowledge driven and KDACAF is knowledge-data jointly driven.
Table 2 shows the results of the tests performed on test Set \8224;
as can be seen from Table 2: (1) Overall, the four SEMs performed worse than the others, indicating that the relatively small capacity of knowledge-driven SEMs limits their approximation power and regression goodness. (2) Each SEM behaves differently in four cases,
performance in Con. #5 and #6 was better than Con. #8 and #9;
and
performs better in Con. #5 and #8 than in Con. #6 and #9;
performed better in Con. #8 and #9 than in Con. #5 and # 6. This phenomenon indicates that knowledge-driven SEMs have limited ability to handle multiple modalities of capacity prediction tasks under different conditions, i.e., no SEM can perform well under four test conditions simultaneously. In contrast, SVR, DLSTM, GPR, LSTM-FC, and KDACAF perform relatively consistently under four conditions. (3) LSTM-FC performed better than SVR, but slightly worse than DLSTM, due to the deeper structure (and greater model capacity) of DLSTM than LSTM-FC. (4) KDACAF has the best performance in both MAE and RMSE termsMeaning that a priori knowledge and data statistics are complementary to some extent and that combining them into one model (i.e., KDACAF) can yield a more significant performance improvement than data-driven and knowledge-driven models.
In addition, FIG. 5 is a diagram showing the relationship between the prediction result and the corresponding prediction error in the test set \8224andFIG. 6 is a diagram showing the relationship between the prediction result and the corresponding prediction error in the test set \8225.
Comparison in test Set \8225
Table 3 shows the performance of all models in test Set \8225;)
As can be seen from table 3, KDACAF has the most satisfactory performance, i.e. the lowest MAE and RMSE under all conditions. The average performance of each model in test set \8224; is also listed in the last column of table 3 for comparison and analysis. The results were: (1) Although all models performed worse on test set 8225than test set 822458, the performance gap between KDACAF and test set 8224and 8225was minimal, showing the highest generalization and versatility. (2) Again, SEM performs worse than other models due to the limited ability of the models to capture multimodalities. (3) The performance of LSTM-FC-TL is the second best of all models because the migration learning mechanism makes LSTM-FC-TL very generalizable under new conditions. (4) For the five best models of SVR, DLSTM, GPR, LSTM-FC-TL, KDACAF they performed the worst on Con. #1 because Con. #1 had the lowest similarity to #5, #6, # 8. Furthermore, these models performed better in Con. #4 than in Con. #2, which means that reducing the temperature from 25C to 10C is a more important contributor than reducing SoC from 50% to 20% during battery calendar aging, so there is a greater mode mismatch when applying the model to Con. #2 than in Con. # 4. Ablation testing:
table 4 is the ablation test results table
To verify the validity and necessity of each module, we performed ablation tests, and the results are listed in table 4, where K, D, L, and S in table 4 represent the knowledge-driven attention module, the data-driven attention module, the long-short memory module, and the semi-empirical module, respectively. Table 4 the results show that: (1) Both attention modules are crucial to the improvement of the predictive performance, i.e. "D + L + S" and "K + L + S" both achieve a significantly lower MAE and RMSE than "L + S". (2) The knowledge-driven attention and data-driven attention modules complement each other, i.e., "K + D + S" and "K + D + L + S" perform much better than "D + L + S" and "K + L + S". (3) Knowledge-driven attention is more powerful than data-driven attention, that is, "K + L + S" behaves slightly better than "D + L + S". (4) The long-time and short-time memory module is added to the existing frame work to bring further improvement, namely the L + S and the K + D + L + S respectively perform better than the S and the K + D + S.
TABLE 5 SEM ablation test results table
To verify the contribution and necessity of each SEM, we performed another ablation test, with the results listed in table 5, from table 5: (1) All SEM contributions to test set \822452in the following order:
. (2) The contributions of all SEMs to the test set 8225while performance followed the following sequence:
. (3) Test set 822424and test set 8225were compared, and the fluctuations in their contributions were found to be significant. This phenomenon may be due to the relatively small capacity of SEMs, which make their adaptation to each condition very different, and therefore each SEM may not perform well in all conditions (i.e., modes). I.e., each SEM is at oneWith a certain degree of preference for certain specific conditions. Therefore, combining them into a single model, i.e., KDACAF, can be considered a relatively ideal combination of these four SEMs, with the contribution and necessity of all SEMs in KDACAF being realized by the attention module.
Convergence analysis of KDACAF
The training of the KDACAF example shown in fig. 3 was performed 1000 more times, with the final MAE and RMSE on test set 822424and 8225, demonstrated by boxplots in fig. 7 and 8, respectively: KDACAF has better convergence and stability, i.e. the fluctuation variance of final MAE and RMSE is very low, indicating that KDACAF has been trained to a better stage. The convergence analysis shows that the KDACAF jointly driven by knowledge and data has the advantages of a large-capacity model and a small-capacity model, namely good prediction performance and high training stability.
The invention adopts the attention-based lithium ion battery calendar aging prediction model with the structure, applies the attention mechanism to the lithium ion battery calendar aging prediction, namely KDACAF, and is beneficial to the monitoring and management of the battery. KDACAF takes battery electrochemical experience knowledge as the key basis, namely the knowledge-driven attention module, and realizes effective complementation of domain knowledge and data. Ablation tests show that the introduction of domain knowledge significantly improves the prediction performance of KDACAF. In KDACAF, a priori knowledge from SEM plays a more important role than the statistical rules of the data contained in the dataset. Case tests on actual calendar aging data show that KDACAF is superior in predicting and generalizing to unseen (completely new) test conditions. Compared with SEM on test set 8225, the MAE and RMSE were reduced by 5.78% and 3.57%, respectively, indicating that the designed KDACAF performance was superior. For battery health prediction, the lower the prediction error rate that can be achieved by a method, the better the prediction performance of the method. In practical applications, the acceptable standard error rate may vary according to different requirements. For example, some automotive companies recommend 2%, while some energy system companies recommend 3%. Since the health of the battery is critical to ensure the efficiency and safety of the battery, it is worth exploring a lower acceptable error rate to expand the application range of the battery.
Finally, it should be noted that: the above embodiments are only for illustrating the technical solutions of the present invention and not for limiting the same, and although the present invention is described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that: modifications and equivalents may be made to the invention without departing from the spirit and scope of the invention.
Claims (8)
1. A lithium ion battery calendar aging prediction model based on knowledge-data joint driving attention is characterized in that: the system comprises a semi-experience module, a knowledge driving attention module, a data driving attention module and a long-time and short-time memory module;
the knowledge-driven attention module takes a semi-empirical model as a front end, and the semi-empirical model is based on the arrhenius law.
2. The lithium ion battery calendar aging prediction model based on knowledge-data joint driving attention of claim 1, characterized in that: also included is an experimental platform for testing the effectiveness of KDACAF, said experimental platform comprising a hotroom for controlling the ambient temperature of the storage battery, battery test equipment for maintaining a predetermined storage state of charge level for the battery, a computer for monitoring and storing battery aging data.
3. A method of attention-based lithium ion battery calendar aging prediction model comprises the following steps:
s1, collecting and preprocessing data;
s2, establishing a calendar aging prediction model;
s21, establishing a problem description and a semi-empirical model of calendar aging prediction;
s22, structure of KDACAF;
s23, loss functions of the long-time and short-time memory module and the KDACAF;
s3, experiments and analysis
S31, comparing and testing;
s32, ablation testing;
and S33, convergence analysis.
4. The method of claim 3, wherein the model is based on a prediction model of calendar aging of lithium ion battery of attention:
the step S1 specifically includes the following steps:
data preprocessing and evaluation index
Carrying out cubic spline interpolation on a training set to ensure that each battery capacity sequence in the training set has 1 hour resolution, the length of each sequence is 11521, and in addition, in order to ensure the time resolution consistency of KDACAF on the training set, the verification set and the test set, all capacity sequences in the training set are sparsely sampled once every 30 days, namely all capacity sequences in the KDACAF training period
Any two consecutive times of
And
the time interval between is still 30 days, i.e.
Hours, which is the same as the time resolution of the validation set and test set; the maximum absolute error and the root mean square error are used for evaluation, namely:
5. the method of claim 3, wherein the model is based on a prediction model of calendar aging of lithium ion battery of attention:
step S21 specifically includes the following steps:
s211, problem description of calendar aging prediction
Calendar aging prediction of lithium ion batteries aims at predicting the variation of their capacity with storage time, the interpretation variables for this task are formatted as the following matrix:
wherein,
it is meant that all of the interpretation information,
the sequential order number representing the capacity sequence,
indicating a sequence of capacities
The measured value of the number of the first measurement,
represent
The storage time of the battery in (1),
which is indicative of the temperature at which the battery is stored,
indicating storage chargeIn the electrical state of the electric motor, the electric motor is in a closed state,
is a lag interval;
for capacity prediction for a single step, the target variable is
Thus, the calendar aging prediction task is abstracted as the following mapping:
wherein,
a predictive model representing data-driven or knowledge-driven;
s212, semi-empirical model of calendar aging prediction
According to arrhenius' law, the degradation of calendar capacity is ultimately expressed in a particular semi-empirical form:
wherein,
and
are all the dependent items of the SoC,
which represents the constant of the gas,
a low power parameter representing a duration dependency;
consider a memory SoC pair registerThe effects of biochemical reactions can lead to degradation of battery capacity, taking into account the linear and exponential dependence of SoC, i.e.
And
in the form of (a), the following semi-empirical model is constructed:
all the SEMs mentioned above can be respectively simplified and expressed as
，
，
，
Wherein
Respectively representing all parameters to be identified, namely:
6. the method of claim 3, wherein the model is based on a prediction model of calendar aging of lithium ion battery of attention:
the structure of the step S22 KDACAF specifically comprises the following steps:
s221, semi-empirical module and prediction branch
The semi-empirical module is a set of four SEMs set forth in S212;
in the case of a predicted branch, the branch is predicted,
the four SEM's input to the semi-empirical module were obtained separately
The regression results, i.e., the preliminary prediction results of (1), are as follows:
that is, for each SEM there are:
s222, fitting branch and knowledge-driven attention module
To fit the results of past capacity sequences, first, one will fit
The interpretation matrix is divided into two parts:
namely:
in the formula,
a sequence of recent capacities is indicated,
each column in (1) represents
Influence factors of corresponding elements in the Chinese character; in the fitting branch
Respectively input into the four SEM to obtain
The regression results of (1):
the present knowledge-driven attention module is dedicated to calendar aging prediction tasks, where targets are predicted
The attention to each SEM was designed based on the goodness of fit of each SEM model to the past real volume;
first, the knowledge-driven attention model is
And
the score function between is defined as follows:
wherein,
a scoring function representing a knowledge-driven attention model;
can then obtain
Attention to each SEM, note
Then, get to
The following points of (1) are accurately predicted:
s223, feature branching and data-driven attention module
In order to expand the model capacity of KDACAF to better capture the multimodalities of the calendar aging process, data-driven attention was constructed as follows:
to construct feature vectors from semi-empirical modules, we will
Again, the four SEMs were entered and four eigenvectors were constructed from the intermediate variables of these SEMs, respectively, as follows:
wherein,
the representation is based on
The constructed features of intermediate variables within;
predicting a target in a data-driven attention module
The attention to each SEM was designed based on the historical capacity sequence and the constructed features, and the score function was defined as:
wherein
A scoring function representing the data-driven attention,
is the first
A trainable matrix corresponding to the SEM;
then, can be obtained
Attention number to each SEM, note
I.e. by
Then, get to
Another accurate prediction of:
7. the method of claim 3, wherein the model is based on a prediction model of calendar aging of lithium ion battery of attention:
step S23 is a loss function of the long-time and short-time memory module and the KDACAF, and specifically includes the following steps:
the comprehensive two-part attention mechanism can be obtained
The intermediate prediction of (2):
wherein
Representing two attention module pairs
The inter-prediction of (2) is performed,
and
all are learnable weights;
then, the historical capacity is sequenced
And
the concatenation is as follows vector:
finally, the vector is input to a long-short-term memory module comprising an LSTM layer, and the output of the LSTM is scaled to
In front of and behind
Is expressed as
；
The loss function of KDACAF is set as follows:
in the formula,
to train the number of samples, the Adam training method was used to train KDACAF.
8. The method of claim 3, wherein the model is based on a prediction model of calendar aging of lithium ion battery of attention:
step S31 specifically includes the following steps:
s311, comparison in a test Set of \822458;
s312, comparison in test Set \8225j.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310023441.2A CN115718263B (en) | 2023-01-09 | 2023-01-09 | Attention-based lithium ion battery calendar aging prediction model and method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310023441.2A CN115718263B (en) | 2023-01-09 | 2023-01-09 | Attention-based lithium ion battery calendar aging prediction model and method |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN115718263A true CN115718263A (en) | 2023-02-28 |
| CN115718263B CN115718263B (en) | 2023-04-07 |
Family
ID=85257877
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310023441.2A Active CN115718263B (en) | 2023-01-09 | 2023-01-09 | Attention-based lithium ion battery calendar aging prediction model and method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115718263B (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117269805A (en) * | 2023-11-23 | 2023-12-22 | 中国人民解放军国防科技大学 | Vehicle-mounted lithium battery pack health state evaluation model training and predicting method and device |
| CN117630683A (en) * | 2024-01-25 | 2024-03-01 | 北京科技大学 | Multi-scale fusion GRU network-based automobile battery SOC multi-step prediction method and system |
Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20140214348A1 (en) * | 2013-01-29 | 2014-07-31 | Mitsubishi Electric Research Laboratories, Inc. | Method for Estimating a State of Charge of Batteries |
| CN111710376A (en) * | 2020-05-13 | 2020-09-25 | 中国科学院计算机网络信息中心 | Block calculation load balancing method and system for macromolecules and cluster systems |
| CN112098874A (en) * | 2020-08-21 | 2020-12-18 | 杭州电子科技大学 | Lithium ion battery electric quantity prediction method considering aging condition |
| US20210049460A1 (en) * | 2019-08-15 | 2021-02-18 | Noodle Analytics, Inc. | Deep probabilistic decision machines |
| CN114325450A (en) * | 2022-01-11 | 2022-04-12 | 青岛科技大学 | Lithium ion battery health state prediction method based on CNN-BilSTM-AT hybrid model |
| CN114596010A (en) * | 2022-05-09 | 2022-06-07 | 西南石油大学 | BiGRU network drilling condition identification method combined with attention mechanism |
| CN114743600A (en) * | 2022-04-15 | 2022-07-12 | 大连理工大学 | Gate-controlled attention mechanism-based deep learning prediction method for target-ligand binding affinity |
| CN115047350A (en) * | 2022-06-24 | 2022-09-13 | 哈尔滨工业大学 | Digital-analog linkage based lithium ion battery remaining service life prediction method |
| CN115409263A (en) * | 2022-08-30 | 2022-11-29 | 北京航空航天大学 | Method for predicting remaining life of lithium battery based on gating and attention mechanism |
-
2023
- 2023-01-09 CN CN202310023441.2A patent/CN115718263B/en active Active
Patent Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20140214348A1 (en) * | 2013-01-29 | 2014-07-31 | Mitsubishi Electric Research Laboratories, Inc. | Method for Estimating a State of Charge of Batteries |
| US20210049460A1 (en) * | 2019-08-15 | 2021-02-18 | Noodle Analytics, Inc. | Deep probabilistic decision machines |
| CN111710376A (en) * | 2020-05-13 | 2020-09-25 | 中国科学院计算机网络信息中心 | Block calculation load balancing method and system for macromolecules and cluster systems |
| CN112098874A (en) * | 2020-08-21 | 2020-12-18 | 杭州电子科技大学 | Lithium ion battery electric quantity prediction method considering aging condition |
| CN114325450A (en) * | 2022-01-11 | 2022-04-12 | 青岛科技大学 | Lithium ion battery health state prediction method based on CNN-BilSTM-AT hybrid model |
| CN114743600A (en) * | 2022-04-15 | 2022-07-12 | 大连理工大学 | Gate-controlled attention mechanism-based deep learning prediction method for target-ligand binding affinity |
| CN114596010A (en) * | 2022-05-09 | 2022-06-07 | 西南石油大学 | BiGRU network drilling condition identification method combined with attention mechanism |
| CN115047350A (en) * | 2022-06-24 | 2022-09-13 | 哈尔滨工业大学 | Digital-analog linkage based lithium ion battery remaining service life prediction method |
| CN115409263A (en) * | 2022-08-30 | 2022-11-29 | 北京航空航天大学 | Method for predicting remaining life of lithium battery based on gating and attention mechanism |
Non-Patent Citations (1)
| Title |
|---|
| 赵珈卉: "锂离子电池状态估计与剩余寿命预测方法综述", 《发电技术》 * |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117269805A (en) * | 2023-11-23 | 2023-12-22 | 中国人民解放军国防科技大学 | Vehicle-mounted lithium battery pack health state evaluation model training and predicting method and device |
| CN117630683A (en) * | 2024-01-25 | 2024-03-01 | 北京科技大学 | Multi-scale fusion GRU network-based automobile battery SOC multi-step prediction method and system |
| CN117630683B (en) * | 2024-01-25 | 2024-03-29 | 北京科技大学 | Multi-scale fusion GRU network-based automobile battery SOC multi-step prediction method and system |
Also Published As
| Publication number | Publication date |
|---|---|
| CN115718263B (en) | 2023-04-07 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Tian et al. | Deep neural network battery charging curve prediction using 30 points collected in 10 min | |
| US20220114421A1 (en) | Method for estimating the capacity of lithium battery based on convolution long-short-term memory neural network | |
| Zhao et al. | A compact methodology via a recurrent neural network for accurate equivalent circuit type modeling of lithium-ion batteries | |
| CN115718263B (en) | Attention-based lithium ion battery calendar aging prediction model and method | |
| Cui et al. | A dynamic spatial-temporal attention-based GRU model with healthy features for state-of-health estimation of lithium-ion batteries | |
| Zhou et al. | Battery health prognosis using improved temporal convolutional network modeling | |
| CN111856287B (en) | Lithium battery health state detection method based on stacked residual causal convolutional neural network | |
| Mamo et al. | Long short-term memory with attention mechanism for state of charge estimation of lithium-ion batteries | |
| CN116609668B (en) | Lithium ion battery health state and residual life prediction method | |
| CN110941929A (en) | Battery health state assessment method based on ARMA and Elman neural network combined modeling | |
| Huang et al. | A comprehensively optimized lithium-ion battery state-of-health estimator based on Local Coulomb Counting Curve | |
| Wu et al. | Attention-based encoder-decoder networks for state of charge estimation of lithium-ion battery | |
| Hu et al. | Lithium-ion battery calendar health prognostics based on knowledge-data-driven attention | |
| CN115201686B (en) | Lithium ion battery health state assessment method under incomplete charge and discharge data | |
| CN112731183B (en) | Improved ELM-based lithium ion battery life prediction method | |
| CN114966436A (en) | Lithium battery state of charge prediction method, device, equipment and readable storage medium | |
| Tian et al. | State-of-charge estimation for lithium-ion batteries based on attentional sequence-to-sequence architecture | |
| CN114578234A (en) | Lithium ion battery degradation and capacity prediction model considering causality characteristics | |
| CN115994441A (en) | Big data cloud platform online battery life prediction method based on mechanism information | |
| CN116794547A (en) | Lithium ion battery residual service life prediction method based on AFSA-GRU | |
| Shen et al. | Transfer learning-based state of charge and state of health estimation for Li-ion batteries: A review | |
| Jiang et al. | Multiple health indicators assisting data-driven prediction of the later service life for lithium-ion batteries | |
| Qian et al. | A CNN-SAM-LSTM hybrid neural network for multi-state estimation of lithium-ion batteries under dynamical operating conditions | |
| Vilsen et al. | Transfer learning for adapting battery state-of-health estimation from laboratory to field operation | |
| Schmitt et al. | State-of-health estimation by virtual experiments using recurrent decoder–encoder based lithium-ion digital battery twins trained on unstructured battery data |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |