Power battery attenuation trajectory prediction method and system oriented to multiple fast charging systems

Abstract

The invention discloses a power battery attenuation trajectory prediction method and system oriented to multiple fast charging systems, and the method comprises the steps: carrying out the statistics of the duration of a battery in each voltage sub-box, and achieving the conversion from a time-voltage sequence to a voltage-duration sequence; reconstructing a voltage curve, and determining voltage segmentation points based on time equal distribution; constructing an aging feature set under each voltage interval based on a voltage interval division method of equal time distribution; based on the constructed aging feature library, performing feature screening to select key features; and on the basis of the aging feature set, constructing a capacity attenuation curve prediction model based on a transform model. The method can adapt to various battery types and charging working conditions, and realizes prediction from historical aging characteristics to future capacity tracks under different working conditions in combination with a deep learning algorithm.

Description

Technical Field

[0001] The present invention belongs to the technical field of power battery management systems, and specifically relates to a method and system for predicting power battery attenuation trajectories for multiple fast-charging modes. Background Art

[0002] Power lithium-ion batteries, due to their excellent energy density and long service life, are used as the core power source for electric vehicles and are a key factor in driving technological advancement and market expansion. However, over the long term, batteries gradually degrade due to side reactions. This degradation directly affects the range of electric vehicles and increases charging time. Battery state of health (SOH) is a key indicator for assessing the service status of batteries, quantifying the degree of degradation of the battery's current performance compared to its new state. Because battery state of health is affected by multiple factors, such as temperature and the number of charge and discharge cycles, it often exhibits a nonlinear decay process. Therefore, health assessment is subject to uncertainty. Accurate battery health assessment and prediction are crucial for ensuring the long-term safe and efficient operation of vehicles.

[0003] From the perspective of modeling methods, existing health state degradation trajectory prediction technologies can be divided into two categories: model-based prediction methods and data-driven prediction methods. Model-based prediction methods aim to predict the battery degradation path by constructing a simulation model that accurately describes the battery aging behavior, which mainly includes equivalent circuit models and electrochemical models. The electrochemical model is a mechanism based on electrochemical reactions. Due to its ability to reflect the internal information of the battery, it is often used to analyze battery aging. However, this method has high model complexity, strict requirements on the quality and quantity of historical data, and may be limited by computing resources in practical applications. Compared with the electrochemical model, the equivalent circuit model simulates the electrochemical characteristics of the battery by using resistors, inductors, capacitors and other circuit elements to characterize the voltage characteristics of the battery. Its model structure is simpler, but in order to identify the key parameters related to the model, the model often needs to set specific current conditions for identification, so it is limited in practical applications.

[0004] Data-driven prediction methods treat the battery as a "black box." Through data mining, they directly map historically measured or indirectly acquired battery parameters to the battery's future capacity, thereby predicting the battery's future trajectory. Data-driven prediction methods primarily rely on historical training battery data to construct a degradation model. During prediction, the model predicts the available capacity or SOH for the next cycle (i+1) based on the battery's historical cycle data (cycles 1 to i). This process is repeated until a predetermined number of cycles are reached. The sequence of predicted capacities or SOH from cycle i+1 to the predetermined cycle period represents the battery's future degradation trajectory. With the rapid development of natural language processing technology, state-of-the-art machine translation systems currently employ an encoder-decoder framework for sequence-to-sequence prediction. This approach eliminates the need for repeated iterative computations and enables efficient sequence-to-sequence prediction directly from the model. In recent years, the encoder-decoder framework has also been applied by researchers both domestically and internationally to lithium-ion battery degradation prediction. However, this approach does not provide a detailed explanation of the basis for selecting aging features, and its applicability under different operating conditions is lacking clear discussion. Compared with model-based methods, data-driven methods rely only on aging data during operation. Without the need to accurately understand the internal aging mechanism of the battery, they can learn the underlying relationships between the data and explore the laws of battery aging evolution, thereby achieving accurate predictions of future states. Therefore, they are more versatile and more suitable for practical applications. However, most of the current data-driven methods rely only on historical capacity information, resulting in limited information and an inability to effectively capture the aging changes within the battery. Therefore, how to accurately characterize the battery aging changes under different operating conditions and more accurately predict the battery capacity trajectory based on this is the main challenge currently faced. Summary of the Invention

[0005] For different charging modes, the existing unified development model is difficult to capture the differences in battery response under different operating conditions, and thus cannot accurately track aging evolution; and the development of prediction models separately will lead to a large number of models, with the disadvantages of low prediction accuracy or increased development and maintenance costs.

[0006] On the one hand, the present invention provides a method for predicting the decay trajectory of a power battery for multiple fast charging modes, the method comprising: Count the duration of the battery in each voltage bin and realize the conversion of time-voltage series to voltage-duration series; Reconstruct the voltage curve and determine the voltage segmentation points based on time equalization; Based on the voltage interval division method with equal time distribution, the aging feature set under each voltage interval is constructed; Based on the constructed aging feature library, feature screening is performed to select key features; Based on the aging feature set, a capacity decay curve prediction model based on the transformer model is constructed.

[0007] Furthermore, the method of counting the duration of the battery at each voltage bin realizes the conversion of the time-voltage sequence to the voltage-duration sequence. To count the duration of the battery at each voltage, the non-monotonic time-voltage sequence is converted into a monotonic voltage-duration sequence to reconstruct the voltage curve. Specifically, the method includes: For non-monotonic Voltage sequence, its sampling time Sampling interval , perform binning on the monotone discrete sequence shown in formula (1): ; Where, and Respectively represent the minimum voltage and maximum voltage during the sampling period; It is the voltage interval, which can be set by yourself; The duration of the battery under each discrete voltage bin is statistically analyzed to obtain the duration distribution shown in formula (2): ; After obtaining the voltage-duration sequence according to the above formula, draw the bin voltage-duration curve; the distribution characteristics of the duration are similar to the IC curve, formula (3): ; Based on the obtained bin voltage-duration curve, the voltage of each bin is The duration under the fixed sub-box voltage point is integrated to obtain the cumulative time , this process is expressed as formula (4): ; Formula (4) transforms the non-monotonic time-voltage curve into a monotonic bin voltage-accumulated time curve. By inversely solving this function through formula (5), we can obtain the reconstructed voltage curve based on the accumulated time: .

[0008] Furthermore, the reconstructing of the voltage curve and determining the voltage segmentation points based on time equalization specifically include: The charging time of the entire multi-stage constant current charging stage is equally divided into three stages, and the segmented time of each stage is as well as , the voltage dividing point corresponds to and , then add the charging starting point and CV stage , forming 5 voltage values , and combine the voltage segmentation points to form 10 voltage intervals; Then, the original data is divided into some intervals based on the segmented voltage points ( ) and the data sequence of the corresponding interval, and verified that the key voltage dividing point is highly concentrated, and the different charging modes are and Take the average value respectively, which can meet the requirements of different standards and determine is a fixed voltage point.

[0009] Furthermore, the voltage interval division method based on equal time distribution constructs an aging feature set in each voltage interval, specifically including: Combining the direct information, implicit electrochemical information and battery electrical characteristics carried by the voltage and current signals monitored during battery aging, a feature set consisting of direct characterization parameters, electrochemical mechanism parameters and electrical performance parameters is constructed; the voltage sequence in each interval of the direct characterization parameters and the current sequence at the corresponding moment are expressed as Equations (6) and (7): ; ; Where, Subscript Indicates any one of the voltage ranges; superscript Indicates the number of cycles; Represents the current sequence at the corresponding sampling point; Indicates the number of sampling points; The corresponding capacity sequence and power sequence can be expressed as Equations (8) and (9): ; ; The electrochemical mechanism parameters mainly include IC curve, Δ Q ( V ) curve and Δ E ( V ) curve introduces the feature set, and the corresponding sequence is expressed by Equation (10), Equation (11) and Equation (12): ; ; ; Where, and Respectively represent the IC curve in the corresponding interval, Curves and curve; The electrical performance parameters of the battery include ohmic internal resistance , polarization internal resistance And the average voltage of each voltage interval It is expressed as formula (13): ; Where, Indicates the capacity at the final moment of charging; Indicates the capacity at the start of charging.

[0010] Furthermore, after obtaining the parameter sequence under each voltage interval, statistical analysis is carried out, and the statistical analysis method includes maximum value, minimum value, mean value, variance, kurtosis, and skewness.

[0011] Furthermore, the feature screening based on the constructed aging feature library is performed to select key features, specifically including: The constructed feature set is screened, and the screening of the constructed feature set is divided into three steps: eliminating low-contribution features based on the variance contribution rate of the principal component analysis method; directly using the Pearson correlation coefficient to judge and delete redundant features; and retaining features that are highly correlated with aging characteristics through the distance correlation coefficient.

[0012] Furthermore, the variance contribution rate based on principal component analysis eliminates low contribution features, specifically including: of observations dimensional monomer aging feature space Normalize and calculate the covariance matrix of the standardized aging feature library and the matrix after eigenvalue decomposition U , by setting the variance contribution rate of the principal component Take out the largest The eigenvalues ​​corresponding to the eigenvectors , and get The principal component , and then use the variance contribution rate of the principal component Delete the principal components in the original feature library Features with low variance contribution; The method of directly using the Pearson correlation coefficient to judge and delete redundant features specifically includes setting up an empty feature library that does not contain redundant features , based on the Pearson correlation coefficient of the original feature library Sort and select the most relevant features to put into , this feature is subjected to Pearson analysis with other features and redundant features are deleted, and the loop is iterated until No assessment required; The method of retaining features highly correlated with aging characteristics by distance correlation coefficient specifically includes setting a feature set and battery capacity , calculate the distance matrix and , then and Calculate the centralized distance matrix, calculate the distance covariance and distance variance, and finally obtain the distance correlation coefficient , formula (14): ; Sort the distance correlations and select features with higher values ​​as the preferred features to be retained.

[0013] Furthermore, based on the aging feature set, a capacity decay curve prediction model based on the transformer model is constructed, specifically including: The prediction model input is Historical cycle data samples are extracted for each cycle data. key features, forming a Matrix A Transformer model based on the seq2seq framework, i.e., an encoder and decoder, is used to predict future capacity decay. The encoder and decoder parameters include parameter s, which balances the amount of data with the degree of battery degradation. , the parameter of the future prediction sequence length ; Output is out and from The cycle starts with a step size of A sequence of battery capacity decay that increases up to future cycle numbers; Input module introduces position encoding , construct the final input matrix : ; The encoder module processes the input sequence, which includes two layers: a multi-head self-attention mechanism and a feedforward neural network layer. The multi-head self-attention mechanism generates the output as Equation (16): ; Where, and is the learned weight matrix; and They are query and key-value matrices respectively; The decoder module generates the target sequence, including a multi-head self-attention mechanism with a mask, and processes the decoder's own input sequence. The formula is modified to Equation (17): ; Where, and are the input query matrix, key matrix and value matrix respectively, is a mask matrix used to mask the attention weights of future positions; The multi-head encoder-decoder attention mechanism allows each decoder position to pay attention to all positions in the encoder output sequence. This process is expressed as Equation (18): ; Where, the query matrix The key matrix from the previous layer output of the decoder Sum Matrix From encoder output.

[0014] Furthermore, the feedforward neural network layer, the encoder and the decoder have the same structure, consisting of a fully connected layer and an activation function.

[0015] On the other hand, the present invention also provides a power battery attenuation trajectory prediction system for multiple fast charging modes, the system comprising: The conversion module is used to count the duration of the battery under each voltage bin and realize the conversion of time-voltage series to voltage-duration series; A reconstruction module is used to reconstruct the voltage curve and determine the voltage segmentation points based on time equalization; A construction module is used to construct an aging feature set in each voltage interval based on a voltage interval division method with equal time distribution; The selection module performs feature screening and selects key features based on the constructed aging feature library; The prediction module is used to build a capacity decay curve prediction model based on the transformer model based on the aging feature set.

[0016] The technical effects achieved by the present invention are: The present invention can establish a voltage interval division method based on the reconstructed voltage curve, decouple the aging differences caused by different working conditions, generate a feature subset that can characterize the aging characteristics of each charging mode, and be adaptable to a variety of battery types and charging conditions. In combination with a deep learning algorithm, it can realize the prediction of historical aging characteristics to future capacity trajectory under different working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a schematic flow chart of the prediction method of the present invention; Figure 2 This is a charging curve diagram of an embodiment of the present invention under the 7C (40%)-3C standard; Figure 3 This is a bar graph of the duration of each sub-box voltage statistics of the present invention; Figure 4 This is a graph of the voltage-duration of the sub-box of the present invention; Figure 5It is a capacity prediction curve diagram of batteries with different lifespans according to the present invention; Figure 6 This is the error distribution of the capacity sequence prediction of the present invention. DETAILED DESCRIPTION

[0018] Because most current encoder-decoder architectures rely solely on historical capacity information, the amount of information is limited and cannot effectively capture internal battery aging changes. Furthermore, developing separate prediction models for different charging modes would result in a large number of models, increasing development and maintenance costs. Existing unified development models struggle to capture differences in battery response under different operating conditions, and thus cannot accurately track aging evolution, resulting in low prediction accuracy. This invention aims to establish a voltage interval partitioning method based on reconstructed voltage curves, decouple the aging differences caused by different operating conditions, generate a feature subset that can characterize the aging characteristics of each charging mode, and combine it with a deep learning algorithm to achieve predictions from historical aging characteristics to future capacity trajectories under different operating conditions.

[0019] The present invention will be described in detail below with reference to the accompanying drawings.

[0020] As shown in Figure 1, a flowchart of a method for predicting the decay trajectory of power batteries for various fast-charging systems is provided. The method is performed in the following steps: Step 1: Count the duration of the battery at each voltage and reconstruct the voltage curve by converting the non-monotonic time-voltage series into a monotonic voltage-duration series.

[0021] For non-monotonic Voltage sequence, its sampling time Sampling interval , perform binning on the monotone discrete sequence as shown in formula (1): ; Where, and Respectively represent the minimum voltage and maximum voltage during the sampling period; It is the voltage interval, which can be set by yourself, here it is 1mV.

[0022] By counting the duration of the battery in each discrete voltage bin, we can obtain the duration distribution shown in formula (2): ; After obtaining the voltage-duration sequence according to the above formula, the bin voltage-duration curve is plotted. It can be seen that the distribution characteristics of these durations are similar to the IC curve, which can be explained by formula (3): ; In this example, the conversion process under a typical charging system is demonstrated. Figure 2 shows the charging curve under the 7C (40%) to 3C system. The voltage range from 2V to 3.6V is binned at 1mV intervals, totaling 1600 voltage bin intervals. The duration of each bin voltage is statistically analyzed, and the results are shown in the bar chart of Figure 3. To more clearly observe the time distribution information under each voltage, a bin voltage-duration curve is plotted, as shown in Figure 4.

[0023] Based on the obtained bin voltage-duration curve, the voltage of each bin is The duration under the fixed sub-box voltage point is integrated to obtain the cumulative time , the process can be expressed as formula (4): ; This formula transforms the non-monotonic time-voltage curve into a monotonic bin voltage-accumulated time curve. By inversely solving the function through formula (5), the reconstructed voltage curve based on the accumulated time is obtained: ; Step 2: Determine the voltage segmentation points based on the reconstructed voltage curve using the time equalization principle.

[0024] The voltage segmentation points of the reconstructed voltage curve are determined based on the principle of equal time distribution. The charging time of the entire multi-stage constant current charging stage is evenly divided into three stages, and the segmentation time of each stage is as well as , the voltage dividing point corresponds to and , then add the charging starting point and CV stage , forming 5 voltage values , and combine the voltage segmentation points to form 10 voltage intervals.

[0025] Then, the original data is divided into some intervals based on the segmented voltage points ( ) and the data sequence of the corresponding interval, and verified that the key voltage dividing point is highly concentrated, and the different charging modes are and Take the average value respectively, which can meet the requirements of different standards and determine is a fixed voltage point.

[0026] In this embodiment, under the 7C (40%)-3C standard, and Determined to be 3.43V, 3.46V and 3.6V, the starting charging voltage , the end voltage of the CV phase , divide part of the original data into intervals As shown in Figure 2, for the statistics of all charging modes, the five dividing points are obtained by taking the average value. 2V, 3.43V, 3.49V, 3.6V (constant voltage stage starting point) and 3.6V (constant voltage stage end point).

[0027] Step 3: Based on the divided voltage intervals, construct an aging feature set that characterizes battery degradation.

[0028] Combining the direct information, implicit electrochemical information, and battery electrical characteristics carried by the voltage and current signals monitored during battery aging, a feature set consisting of direct characterization parameters, electrochemical mechanism parameters, and electrical performance parameters is constructed. The voltage sequence in each interval of the direct characterization parameters and the current sequence at the corresponding moment can be expressed as Equations (6) and (7): ; ; Where, Subscript Indicates any one of the voltage ranges; superscript Indicates the number of cycles; Represents the current sequence at the corresponding sampling point; Indicates the number of sampling points.

[0029] The corresponding capacity sequence and power sequence can be expressed as Equations (8) and (9): ; ; The electrochemical mechanism parameters mainly include IC curve, Δ Q ( V ) curve and Δ E ( V ) curve introduces the feature set, and the corresponding sequence is expressed by Equation (10), Equation (11) and Equation (12): ; ; ; Where, and Respectively represent the IC curve in the corresponding interval, Curves and curve.

[0030] The electrical performance parameters of the battery mainly include ohmic internal resistance , polarization internal resistance And the average voltage of each voltage interval It can be expressed as formula (13): ; Where, Indicates the capacity at the final moment of charging; Indicates the capacity at the start of charging.

[0031] After fully obtaining the parameter sequences under different charging intervals, statistical analysis was carried out using six statistical methods: maximum value, minimum value, mean, variance, kurtosis, and skewness.

[0032] In this embodiment, there are There are 10 voltage intervals, 3 categories and 10 data types are introduced, and 338 features are finally generated after 6 types of statistics.

[0033] Step 4: Based on the constructed large aging feature library, perform feature screening to select key features.

[0034] Screening the constructed feature set is mainly divided into three steps: eliminating low-contribution features based on the variance contribution rate of the principal component analysis method; directly using the Pearson correlation coefficient to judge and delete redundant features; and retaining features that are highly correlated with aging characteristics through the distance correlation coefficient.

[0035] The first step of feature screening is to of observations dimensional monomer aging feature space Normalize and calculate the covariance matrix of the standardized aging feature library and the matrix after eigenvalue decomposition U , by setting the variance contribution rate of the principal component Take out the largest The eigenvalues ​​corresponding to the eigenvectors , and get The principal component , and then use the variance contribution rate of the principal component Delete the principal components in the original feature library Features with low variance contribution.

[0036] The second step requires setting up an empty feature library that does not contain redundant features , based on the Pearson correlation coefficient of the original feature library Sort and select the most relevant features to put into , the feature is subjected to Pearson analysis with other features and redundant features are deleted, and the loop is iterated until No assessment required.

[0037] The third step requires setting the feature set and battery capacity , calculate the distance matrix and , then and Calculate the centralized distance matrix, calculate the distance covariance and distance variance, and finally obtain the distance correlation coefficient As shown in formula (14): ; Sort the distance correlations and select features with higher values ​​as the preferred features to be retained.

[0038] In this embodiment, the first step is based on the variance contribution rate of the principal component analysis method, Set to 90%, The distance correlation coefficient was set to 1 / 338, and a total of 18 features were eliminated. In the second step, 145 features were eliminated by analyzing the linear correlation between the features, and 175 features were retained. In the third step, 29 features were finally retained by the distance correlation coefficient. The feature subset is shown in Table 1: Table 1 Optimal feature subsets screened based on distance correlation coefficient ; Step 5: Construct a capacity decay curve prediction model based on the transformer model with the feature subset as input.

[0039] The model input is Historical cycle data samples are extracted for each cycle data. n key features, forming a Matrix A Transformer model based on the seq2seq framework, i.e., an encoder and decoder, is used to predict future capacity decay. The encoder and decoder parameters include parameter s, which balances the amount of data with the degree of battery degradation. , the parameter q of the future prediction sequence length; the output is The cycle starts with a step size of A sequence of battery capacity decay that increases incrementally until a certain number of cycles in the future.

[0040] Input module introduces position encoding , construct the final input matrix : ; The encoder module processes the input sequence, including two layers of multi-head self-attention mechanism and feedforward neural network. The multi-head self-attention mechanism generates the output as formula (16): ; Where, and is the learned weight matrix; and They are query and key-value matrices respectively.

[0041] The decoder module generates the target sequence, including a multi-head self-attention mechanism with a mask, and processes the decoder's own input sequence. The formula is modified to Equation (17): ; Where, and are the input query matrix, key matrix and value matrix respectively, is a mask matrix used to mask the attention weights of future positions.

[0042] The multi-head encoder-decoder attention mechanism allows each decoder position to pay attention to all positions in the encoder output sequence. This process is expressed as Equation (18): ; Where, the query matrix The key matrix from the previous layer output of the decoder Sum Matrix From encoder output.

[0043] The feedforward neural network layer, the encoder and decoder have the same structure, consisting of a fully connected layer and an activation function.

[0044] In this embodiment, the parameters The prediction step parameter p is set to 1, the prediction step parameter p is set to 5, and q is set to 40. The training set, validation set, and test set are divided into a ratio of 8:1:1. After the samples are classified by cycle life, they are sampled proportionally. The hyperparameter settings of the Transformer model are detailed in Table 2. The maximum training epoch is 200, and the learning rate is 2e -5 , the drop rate is set to 0.3.

[0045] Table 2 Transformer algorithm hyperparameter values Hyperparameters Parameter value Number of layers [1, 2, 4, 6] High-dimensional vector dimensions [128, 256, 512] Number of heads in the multi-head self-attention mechanism [4, 8] Feedforward neural network dimensions [1024, 2048] Batch size [64, 128] RMSE, MAXE, RMSPE, and MAXPE were used to evaluate the model's prediction performance. The capacity prediction curves for batteries with different lifespans and the error distribution of capacity series predictions are shown in Figures 5 and 6. This proves that the model has good prediction effects for batteries with different lifespans and different operating conditions, and has wide applicability.

[0046] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by any person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims. The information disclosed in the background technology section of this article is only intended to deepen the understanding of the overall background technology of the present invention, and should not be regarded as an admission or any form of implication that the information constitutes prior art already known to those skilled in the art.

Claims

1. A method for predicting power battery attenuation trajectories for multiple fast-charging systems, characterized in that: The method comprises: Count the duration of the battery in each voltage bin and realize the conversion of time-voltage series to voltage-duration series; Reconstruct the voltage curve and determine the voltage segmentation points based on time equalization; Based on the voltage interval division method with equal time distribution, the aging feature set under each voltage interval is constructed; Based on the constructed aging feature library, feature screening is performed to select key features; Based on the aging feature set, a capacity decay curve prediction model based on the transformer model is constructed.

2. The method for predicting power battery attenuation trajectories for multiple fast-charging modes according to claim 1, characterized in that: The method of counting the duration of the battery at each voltage bin realizes the conversion of the time-voltage sequence to the voltage-duration sequence. In order to count the duration of the battery at each voltage, the non-monotonic time-voltage sequence is converted into a monotonic voltage-duration sequence to reconstruct the voltage curve. Specifically, the method includes: For non-monotonic Voltage sequence, its sampling time Sampling interval , perform binning on the monotone discrete sequence shown in formula (1): ; Where, and Respectively represent the minimum voltage and maximum voltage during the sampling period; It is the voltage interval, which can be set by yourself; The duration of the battery under each discrete voltage bin is statistically analyzed to obtain the duration distribution shown in formula (2): ; After obtaining the voltage-duration sequence according to the above formula, draw the bin voltage-duration curve; the distribution characteristics of the duration are similar to the IC curve, formula (3): ; Based on the obtained bin voltage-duration curve, the voltage of each bin is The duration under the fixed sub-box voltage point is integrated to obtain the cumulative time , this process is expressed as formula (4): ; Formula (4) transforms the non-monotonic time-voltage curve into a monotonic bin voltage-accumulated time curve. By inversely solving this function through formula (5), we can obtain the reconstructed voltage curve based on the accumulated time: 。 3. The power battery attenuation trajectory prediction method for multiple fast charging systems according to claim 1 is characterized in that: The reconstructed voltage curve determines the voltage segmentation points based on time equalization, specifically including: The charging time of the entire multi-stage constant current charging stage is equally divided into three stages, and the segmented time of each stage is as well as , the voltage dividing point corresponds to and , then add the charging starting point and CV stage , forming 5 voltage values , and combine the voltage segmentation points to form 10 voltage intervals; Then divide the original data into some intervals based on the segmented voltage points And the data sequence of the corresponding interval, and through verification, the key voltage dividing point is highly concentrated, and the different charging modes are and Take the average value respectively, which can meet the requirements of different standards and determine is a fixed voltage point.

4. The method for predicting power battery attenuation trajectories for multiple fast-charging modes according to claim 1, characterized in that: The voltage interval division method based on equal time distribution is used to construct an aging feature set in each voltage interval, specifically including: Combining the direct information, implicit electrochemical information and battery electrical characteristics carried by the voltage and current signals monitored during battery aging, a feature set consisting of direct characterization parameters, electrochemical mechanism parameters and electrical performance parameters is constructed; the voltage sequence in each interval of the direct characterization parameters and the current sequence at the corresponding moment are expressed as Equations (6) and (7): ; ; Where, Subscript Indicates any one of the voltage ranges; superscript Indicates the number of cycles; Represents the current sequence at the corresponding sampling point; Indicates the number of sampling points; The corresponding capacity sequence and power sequence can be expressed as Equations (8) and (9): ; ; The electrochemical mechanism parameters mainly include IC curve, Δ Q ( V ) curve and Δ E ( V ) curve introduces the feature set, and the corresponding sequence is expressed by Equation (10), Equation (11) and Equation (12): ; ; ; Where, and Respectively represent the IC curve in the corresponding interval, Curves and curve; The electrical performance parameters of the battery include ohmic internal resistance , polarization internal resistance And the average voltage of each voltage interval It is expressed as formula (13): ; Where, Indicates the capacity at the final moment of charging; Indicates the capacity at the start of charging.

5. The method for predicting power battery attenuation trajectories for multiple fast-charging modes according to claim 4 is characterized in that: After obtaining the parameter sequence under each voltage interval, statistical analysis is carried out, and the statistical analysis method includes maximum value, minimum value, mean value, variance, kurtosis, and skewness.

6. The method for predicting power battery attenuation trajectories for multiple fast-charging modes according to claim 1, characterized in that: The feature screening based on the constructed aging feature library is performed to select key features, specifically including: The constructed feature set is screened, and the screening of the constructed feature set is divided into three steps: eliminating low-contribution features based on the variance contribution rate of the principal component analysis method; directly using the Pearson correlation coefficient to judge and delete redundant features; and retaining features that are highly correlated with aging characteristics through the distance correlation coefficient.

7. The method for predicting power battery attenuation trajectories for multiple fast-charging modes according to claim 6, characterized in that: The variance contribution rate based on principal component analysis method eliminates low contribution features, specifically including: of observations dimensional monomer aging feature space Normalize and calculate the covariance matrix of the standardized aging feature library and the matrix after eigenvalue decomposition U , by setting the variance contribution rate of the principal component Take out the largest The eigenvalues ​​corresponding to the eigenvectors , and get The principal component , and then use the variance contribution rate of the principal component Delete the principal components in the original feature library Features with low variance contribution; The method of directly using the Pearson correlation coefficient to judge and delete redundant features specifically includes setting up an empty feature library that does not contain redundant features , based on the Pearson correlation coefficient of the original feature library Sort and select the most relevant features to put into , this feature is subjected to Pearson analysis with other features and redundant features are deleted, and the loop is iterated until No assessment required; The method of retaining features highly correlated with aging characteristics by distance correlation coefficient specifically includes setting a feature set and battery capacity , calculate the distance matrix and , then and Calculate the centralized distance matrix, calculate the distance covariance and distance variance, and finally obtain the distance correlation coefficient , formula (14): ; Sort the distance correlations and select features with higher values ​​as the preferred features to be retained.

8. The method for predicting power battery attenuation trajectories for multiple fast-charging modes according to claim 1 is characterized in that: The capacity attenuation curve prediction model based on the transformer model is constructed based on the aging feature set, specifically including: The prediction model input is Historical cycle data samples are extracted for each cycle data. n key features, forming a Matrix ; The Transformer model based on the seq2seq framework, i.e., the encoder and decoder, is used to predict future capacity decay. The encoder and decoder parameters include parameter s, which balances the amount of data with the degree of battery degradation. , the parameter q of the future prediction sequence length; the output is The cycle starts with a step size of A sequence of battery capacity decay that increases up to future cycle numbers; Input module introduces position encoding , construct the final input matrix : ; The encoder module processes the input sequence, which includes two layers: a multi-head self-attention mechanism and a feedforward neural network layer. The multi-head self-attention mechanism generates the output as Equation (16): ; Where, and is the learned weight matrix; and They are query and key-value matrices respectively; The decoder module generates the target sequence, including a multi-head self-attention mechanism with a mask, and processes the decoder's own input sequence. The formula is modified to Equation (17): ; Where, and are the input query matrix, key matrix and value matrix respectively, is a mask matrix used to mask the attention weights of future positions; The multi-head encoder-decoder attention mechanism allows each decoder position to pay attention to all positions in the encoder output sequence. This process is expressed as Equation (18): ; Where, the query matrix The key matrix from the previous layer output of the decoder Sum Matrix From encoder output.

9. The method for predicting power battery attenuation trajectories for multiple fast-charging modes according to claim 8, characterized in that: The feedforward neural network layer, encoder and decoder have the same structure, consisting of a fully connected layer and an activation function.

10. A power battery attenuation trajectory prediction system for multiple fast charging modes, characterized by: The system comprises: The conversion module is used to count the duration of the battery under each voltage bin and realize the conversion of time-voltage series to voltage-duration series; A reconstruction module is used to reconstruct the voltage curve and determine the voltage segmentation points based on time equalization; A construction module is used to construct an aging feature set in each voltage interval based on a voltage interval division method with equal time distribution; The selection module performs feature screening and selects key features based on the constructed aging feature library; The prediction module is used to build a capacity decay curve prediction model based on the transformer model based on the aging feature set.

Images