Search

CN-122017654-A - Battery capacity prediction method

CN122017654ACN 122017654 ACN122017654 ACN 122017654ACN-122017654-A

Abstract

The application provides a battery capacity prediction method, which comprises the steps of filtering and removing first battery capacity data to obtain first data, mapping the first data into second data, wherein the first battery capacity data comprise battery capacities of a first battery when the first battery is charged or discharged for the first N times respectively, splicing and fusing dimension remodelling capacity characteristics, difference trend characteristics and moving average characteristics generated based on the second data to obtain first time sequence data characteristics, obtaining second time sequence data characteristics after reconstruction, further adding position codes to obtain third time sequence data characteristics, adding the second time sequence data characteristics to the position codes to obtain third time sequence data characteristics, inputting the third time sequence data characteristics to a trained prediction model to obtain first battery capacity prediction data of the first battery, and the first battery capacity prediction data comprise the predicted battery capacities of the first battery when the first battery is charged or discharged for the (n+1) th time. By adopting the method, the accurate prediction of the battery capacity can be realized.

Inventors

  • Request for anonymity
  • Request for anonymity
  • Request for anonymity
  • Request for anonymity
  • Request for anonymity
  • Request for anonymity

Assignees

  • 深圳市瑞能实业股份有限公司
  • 华南理工大学

Dates

Publication Date
20260512
Application Date
20260413

Claims (10)

  1. 1. A battery capacity prediction method, comprising: Filtering and denoising the first battery capacity data of the first battery through a Savitzky-Golay filter, removing extreme abnormal data to obtain first data, and mapping the obtained first data to a 0-1 interval through extremum standardization operation to obtain second data, wherein the first battery capacity data comprises that the first battery is respectively before Battery capacity at sub-charge or discharge, wherein, Is a positive integer; Generating dimension remodeled capacity features, difference trend features and moving average features based on the second data, and splicing and fusing the dimension remodeled capacity features, the difference trend features and the moving average features to obtain first time sequence data features; performing feature screening and reconstruction on the first time sequence data features based on a random forest algorithm to obtain second time sequence data features; adding position codes to the second time sequence data characteristics to obtain third time sequence data characteristics, and inputting the third time sequence data characteristics into a trained prediction model to obtain first battery capacity prediction data of a first battery, wherein the first battery capacity prediction data comprises the predicted first battery in the first battery Battery capacity at secondary charge or discharge.
  2. 2. The battery capacity prediction method according to claim 1, wherein, Adding position codes to the second time sequence data features to obtain third time sequence data features, and inputting the third time sequence data features into a trained prediction model to obtain first battery capacity prediction data of a first battery, wherein the method further comprises the following steps: Obtaining training data of a prediction model, wherein the training data comprises the following steps of according to a preset batch size Divided into A batch feature, wherein, The said The batch characteristics are determined by the second battery capacity data Is generated after filtering and noise reduction, extreme abnormal data rejection, extremum standardization operation, splicing and fusion, feature screening and reconstruction, Positive integer, second battery capacity data Comprises a second battery respectively arranged at Battery capacity at sub-charge or discharge, wherein, Is a positive integer; Will be The first of the batch characteristics Individual lot characterization Input to the prediction model to obtain the first Individual lot characterization Corresponding batch forecast values And according to the batch forecast value And the first Individual lot characterization Corresponding batch true value Updating parameters of the prediction model to obtain a trained prediction model, wherein the batch true value According to each second battery capacity data Preset batch size The obtained product.
  3. 3. The battery capacity prediction method according to claim 1, wherein, Generating dimensional remodelling capacity features, differential trend features, and moving average features based on the second data, specifically including: Obtaining a battery capacity characteristic according to the formula (1), wherein, (1), A battery capacity characteristic representing dimensional remodeling, The second data is represented by a representation of the second data, The number of samples is represented and the number of samples, The time step is represented by a time step, Obtaining a differential trend characteristic of dimensional remodeling according to the formulas (2) (3) (4), wherein, (2), (3), (4), Wherein, the Represent the first Samples of (a) Capacity data for a single time step, Representing the differential battery capacity characteristics of adjacent time steps, A battery capacity differential characteristic representing adjacent time steps of all samples, Representing the result after the end of the battery capacity difference signature at adjacent time steps of all samples has been supplemented with a zero value, A differential trend feature representing dimensional remodelling and used to capture the amount of battery capacity change between adjacent time steps; A moving average characteristic is obtained according to (5) (6) (7), (5), Wherein, the (6), (7), Wherein, the Represent the first Samples of (a) Capacity data for a single time step, Capacity data representing all samples and all time steps, Representing a moving average characteristic of dimensional remodeling, wherein, 、 、 、 Respectively belong to positive integers.
  4. 4. The battery capacity prediction method according to claim 3, wherein, And splicing and fusing the dimension remolded capacity feature, the difference trend feature and the moving average feature to obtain a first time sequence data feature, which specifically comprises the following steps: obtaining a first time-ordered data characteristic according to equation (8), wherein, (8), Wherein, the Representing a first time-ordered data characteristic.
  5. 5. The battery capacity prediction method according to claim 4, wherein, The method for obtaining the second time sequence data features comprises the following steps of: obtaining a second time series data characteristic according to (9) (10) (11), (9), Wherein, the (10), (11), In the form of a two-dimensional feature matrix, Representing characteristics of a first time sequence The two-dimensional feature matrix after screening has the dimensions of , In order to select the number of features, The time step after the reconstruction is performed, Representing a second time sequence data characteristic after characteristic screening and reconstruction, wherein, , wherein, Represent the first The importance weights of the individual flattened features, Represent the first Reject the first in the decision tree After each feature, the variance of the mean square error of the prediction model, Representing the number of decision trees in the random forest, The threshold value of importance is represented as such, Is a positive integer.
  6. 6. The battery capacity prediction method according to claim 5, wherein, Adding position codes to the second time sequence data characteristic to obtain a third time sequence data characteristic, wherein the method comprises the following steps: A third temporal data characteristic is obtained by (12), (12), Wherein, the Representing the matrix of the position-coding that can be learned, Representing the third time series data characteristic after adding the position code.
  7. 7. The battery capacity prediction method according to claim 6, wherein, Adding position codes to the second time sequence data characteristics to obtain third time sequence data characteristics, and inputting the third time sequence data characteristics into a trained prediction model to obtain first battery capacity prediction data of a first battery, wherein the method specifically comprises the following steps of: obtaining first battery capacity prediction data through the steps of (13) - (20); (13) , wherein, (14), (15), (16), (17), (18), (19), (20), Wherein, the Representing the first battery capacity prediction data, Representing the time-series characteristics of the weighted attention, The regularization function is represented as a function of the regularization, The probability is represented by a probability that, Representing the full-connection layer weight matrix, Representing the full connection layer offset vector, Representation of Model No The fusion of the hidden states of the individual time steps, Representation of Last time step of model Is used for the fusion of the hidden states, Indicating that after adding position codes, the first The feature vectors of the individual time steps, By a third time series data feature The product of the process is obtained by the method, The feature stitching is represented and is performed, Representation of The gate control unit operates the function, 、 、 A query matrix, a key matrix, a value matrix respectively representing the attention mechanism, composed of Obtained by linear transformation, wherein, , 、 、 Respectively the first The sub-matrices corresponding to the individual attention headers, Representing the dimensions of the query matrix and the key matrix, The weight normalization function is represented as a function of the weights, Is that The output results of the individual attention heads, The characteristic stitching function is represented as a function of the feature, The linear transformation matrix is output for attention.
  8. 8. The battery capacity prediction method according to claim 7, wherein, After obtaining the first battery capacity prediction data of the first battery, further comprising: Performing shape calibration and outlier filtering on the first battery capacity data by means of formulas (21) and (22) to obtain first battery capacity suppression data, performing shape calibration and outlier filtering on the first battery capacity prediction data to obtain first battery suppression prediction data, The first battery capacity suppression data includes that the first battery is respectively before The first battery inhibition prediction data comprises the filtering data of the first battery capacity prediction data after shape calibration and outlier filtering; (21), (22), representing the first battery capacity prediction data, The data representing the capacity of the first battery, First battery rejection prediction data representing a first battery, The first battery capacity suppression data is represented, Indicating the rated capacity of the first battery, Representing an outlier rejection function.
  9. 9. The battery capacity prediction method according to claim 8, wherein, Performing shape calibration and outlier filtering on the first battery capacity data by the formulas (21) and (22) to obtain first battery capacity suppression data, performing shape calibration and outlier filtering on the first battery capacity prediction data to obtain first battery suppression prediction data, and then further comprising: Determining the predictive model according to equations (23) (24) (25) The loss, wherein, (23), (24), (25), Wherein, the Representing an error between the first battery capacity suppression data of the first battery and the first battery suppression prediction data, An absolute value representing an error between the first battery capacity suppression data of the first battery and the first battery suppression prediction data, Represent the first The error of the individual samples is determined by, Representation of The loss threshold value is set to be equal to the loss threshold value, Representing the predictive model Loss.
  10. 10. The battery capacity prediction method according to claim 9, wherein, Adding position codes to the second time sequence data features to obtain third time sequence data features, and inputting the third time sequence data features into a trained prediction model to obtain first battery capacity prediction data of a first battery, wherein the method further comprises the following steps: updating parameters of the predictive model with AdamW optimizers to obtain a trained predictive model, wherein, The method for updating parameters of the prediction model by adopting AdamW optimizers to obtain a trained prediction model specifically comprises the following steps: updating parameters of the predictive model using equations (26) (27) (28) (29) (30) (31) to obtain a trained predictive model, wherein, (26), (27), (28), (29), (30), (31), Wherein, the Represent the first The parameters of the predictive model at the time of the iteration, Represented as a gradient of a parameter, 、 Respectively representing the initialized smoothing coefficients, Representation of Is used for the first moment estimation of (a), Representation of Is used for the second moment estimation of (a), The learning rate is indicated as being indicative of the learning rate, Representing the minimum value of the initialization and, Representing the initialized weight decay coefficients, Is a positive integer which is used for the preparation of the high-voltage power supply, , 。

Description

Battery capacity prediction method Technical Field The application relates to the technical field of battery life prediction, in particular to a battery capacity prediction method. Background The existing battery capacity prediction method is mainly divided into a traditional model-based method and a data-driven deep learning method. The traditional model-based method takes a battery aging mechanism as a core, realizes capacity prediction by establishing a capacity attenuation model or an equivalent circuit model, and has the outstanding problems of being required to rely on accurate battery aging parameter calibration and complex manual feature extraction, being obviously influenced by factors such as battery type, charge and discharge working conditions, environmental temperature fluctuation and the like, being weak in anti-interference capability, limited in prediction precision and the like. The existing deep learning method still has the defects of insufficient data feature mining, redundant model structure, higher calculation cost, insufficient robustness of training strategies, low feature utilization rate and the like, and is difficult to realize accurate prediction of battery capacity and also incapable of providing reliable data support for battery life assessment. Disclosure of Invention In order to solve the technical problems, the application provides a battery capacity prediction method, which can realize high-precision prediction of battery capacity through feature enhancement modes such as multi-dimensional feature remodeling, splicing and fusion, and modes such as learning position coding addition. In a first aspect, the present application provides a battery capacity prediction method, the method comprising: Filtering and denoising the first battery capacity data of the first battery through a Savitzky-Golay filter, removing extreme abnormal data to obtain first data, and mapping the obtained first data to a 0-1 interval through extremum standardization operation to obtain second data, wherein the first battery capacity data comprises that the first battery is respectively before Battery capacity at sub-charge or discharge, wherein,Is a positive integer; Generating dimension remodeled capacity features, difference trend features and moving average features based on the second data, and splicing and fusing the dimension remodeled capacity features, the difference trend features and the moving average features to obtain first time sequence data features; performing feature screening and reconstruction on the first time sequence data features based on a random forest algorithm to obtain second time sequence data features; adding position codes to the second time sequence data characteristics to obtain third time sequence data characteristics, and inputting the third time sequence data characteristics into a trained prediction model to obtain first battery capacity prediction data of a first battery, wherein the first battery capacity prediction data comprises the predicted first battery in the first battery Battery capacity at secondary charge or discharge. With reference to the first aspect, in an alternative embodiment, Adding position codes to the second time sequence data features to obtain third time sequence data features, and inputting the third time sequence data features into a trained prediction model to obtain first battery capacity prediction data of a first battery, wherein the method further comprises the following steps: obtaining training data of a prediction model, wherein the training data comprises the following steps of Divided intoA batch feature, wherein, The saidThe batch characteristics are determined by the second battery capacity dataIs generated after filtering and noise reduction, extreme abnormal data rejection, extremum standardization operation, splicing and fusion, feature screening and reconstruction,Positive integer, second battery capacity dataComprises a second battery respectively arranged atBattery capacity at sub-charge or discharge, wherein,Is a positive integer; Will be The first of the batch characteristicsIndividual lot characterizationInput to the prediction model to obtain the firstIndividual lot characterizationCorresponding batch forecast valuesAnd according to the batch forecast valueAnd the firstIndividual lot characterizationCorresponding batch true valueUpdating parameters of the prediction model to obtain a trained prediction model, wherein the batch true valueAccording to each second battery capacity dataPreset batch sizeThe obtained product. With reference to the first aspect, in an alternative embodiment, Generating dimensional remodelling capacity features, differential trend features, and moving average features based on the second data, specifically including: Obtaining a battery capacity characteristic according to the formula (1), wherein, (1), A battery capacity characteristic representing dimensional remodeling,The second data is r