Lithium battery state data generation method based on fusion time sequence generative adversarial network

Abstract

The invention discloses a lithium battery state data generation method based on a fusion time sequence generative adversarial network. According to the method, a charging period is extracted from lithium battery state data to serve as real data for training the generative adversarial network. An environment temperature label and random noise corresponding to real data are used as condition information to be input into a generator, modeling is conducted on the condition information through a gating circulation unit, the spatial resolution of the hidden state output by a GRU module is gradually recovered through a convolution layer by means of a transposition convolution module, and lithium battery state data are generated. In the training process, a weight constraint strategy is introduced to control parameter distribution of a discriminator, a dynamic gradient penalty mechanism is introduced into a loss function of the discriminator, and the gradient is adaptively adjusted to improve the training stability. A voltage differential constraint is added in a loss function of the generator, and it is ensured that generated data meet a specific physical constraint condition. The confrontation process is optimized through an asymmetric gradient penalty mechanism, the model convergence time is shortened, and the authenticity of generated data is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of electric digital data processing, and relates to data generation using deep learning technology, specifically to a method for generating lithium battery state data based on a fusion time-series generative adversarial network. Background Art

[0002] Battery state data has significant temporal and non-linear characteristics, and there are complex coupling relationships among various factors, resulting in certain difficulties in effective feature extraction. At the same time, the problem of uneven distribution of battery state data is prominent, leading to the absence of key feature data, which directly affects the accuracy of lithium battery state model modeling and recognition.

[0003] As the core technology in the field of data generation, the generative adversarial network (GAN) shows significant advantages in fitting data distributions by virtue of the adversarial game mechanism between the generator and the discriminator, and can learn complex distributions from limited real data and generate approximate samples. However, due to the instability of its generation process, it faces significant technical challenges in the application of generating lithium battery state data: on the one hand, the generated lithium battery state data lacks temporality and is difficult to capture the long-term and short-term correlations of non-linear characteristics such as the change in remaining battery capacity and the fluctuation of internal resistance in multi-source time-series data such as temperature, current, and charge-discharge cycles; on the other hand, due to the ineffective embedding of physical characteristics such as voltage, current, and remaining battery capacity range in the model, it is difficult to adapt to the working condition changes of temperature and random noise, resulting in deviations between the generated data and the real operating state in terms of physical rationality and temporal consistency, and the quality and authenticity of the generated data are insufficient. Summary of the Invention

[0004] Aiming at the deficiencies of the prior art, the present invention proposes a method for generating lithium battery state data based on a fusion time-series generative adversarial network. Firstly, by integrating the time-series modeling of the gated recurrent unit (GRU) and the spatial reconstruction ability of transposed convolution, the generation accuracy of battery data is improved. Secondly, by integrating weight constraint strategies, dynamic gradient penalty mechanisms, and asymmetric gradient penalty mechanisms, the stability of the data generation process and the physical consistency between the output data and the real physical state of the battery are improved.

[0005] The method for generating lithium battery state data based on a fusion time-series generative adversarial network includes the following steps:

[0006] Step 1: Preprocessing of lithium battery state data

[0007] Define the time period when the current value is positive during the charging process of the battery as the charging moment, that is , where , respectively represent the control current and control voltage at the ), current ( ), and remaining battery charge ( ) from the original battery state dataset to extract the charging cycles , where is the number of extracted charging cycles.

[0008] To achieve comparability between different charging cycles, the number of time points for each charging cycle is made the same by linear interpolation:

[0009]

[0010] Among them, is the k-th sampling time point in the charging cycle, is the battery state data sampled at time is the target time point generated by interpolation, is the battery state data obtained by interpolation.

[0011] Normalize the interpolated charging cycle data to avoid deviations caused by data with different dimensions and ensure the accuracy of subsequent processing:

[0012]

[0013] Among them, and are the minimum and maximum values of the current state data respectively, is the original value of the state data, is the normalized state data value. Through the normalization operation, the values of battery state data with different dimensions are converted to the same range [-1, 1].

[0014] Through the above preprocessing steps, the dataset has consistent input features when performing model training.

[0015] Step 2: Design of the conditional generator

[0016] Design a conditional generator based on GRU - transposed convolution, using the temperature label and Gaussian noise as inputs. Consider the long - term dependence of time - series data during the data generation process, and combine the GRU time - series modeling and the spatial reconstruction ability of transposed convolution to generate physically consistent lithium - battery state data:

[0017] S2.1. Label embedding layer

[0018] To more conveniently use the ambient temperature as an input to the model later, discretely map the actual ambient temperature value to a label , and put these three labels into a set Then, the environmental temperature label is transformed into a learnable embedding vector through the embedding layer , and together with the random noise, it is passed as conditional information of the generator to the subsequent module. l is the dimension of the environmental temperature label .

[0019] S2.2. GRU-Transposed Convolution Module

[0020] Combining the GRU (Gated Recurrent Unit) module with the transposed convolution module forms an end-to-end generation architecture of "temporal learning + spatial restoration", enabling the model to have good temporal modeling and spatial restoration capabilities, and being able to generate more realistic, continuous, and controllable samples in the lithium battery state data generation task.

[0021] The GRU module includes an update gate ( ), a reset gate ( ), a candidate hidden state ( ), and a hidden state ( ). Through the gating mechanism, it can effectively learn and remember the important information in the time series data, capture the long-term dependencies in the lithium battery charging process, contribute to improving the accuracy of temporal data modeling, and ensure the consistency of the generated data in terms of time series.

[0022] The transposed convolution module gradually restores the spatial resolution of the last hidden state output by the GRU module through multiple convolutional layers, maps it from the low-dimensional temporal representation to the high-dimensional spatial output, performs spatial reconstruction, and generates lithium battery state data equal in length to the charging cycle :

[0023]

[0024] Among them, is the th element of the output data , representing the spatial output obtained through the convolution operation. is the position of the current predicted time point, , represents the number of time points in one charging cycle after interpolation. is the number of convolutional layers. is the th convolutional layer in the transposed convolution module, used to extract specific features from the input data. is the th element of the input hidden state , representing the position of the input hidden state in the spatial dimension, is the convolution stride.

[0025] Step 3: Discriminator Design

[0026] S3.1. Construct a multi-layer convolutional neural network as the discriminator, and design a spectral normalization layer based on a weight constraint strategy to normalize the weight matrix of the discriminator :

[0027]

[0028] Among them, is the weight matrix after normalization processing, represents calculating the standard deviation, is a learnable scaling factor, represents the total number of elements of the weight matrix W. Through spectral normalization, the scale of the weights in each dimension can be better adjusted during the optimization process, avoiding the influence of too large or too small weight values on the stability of the training process.

[0029] S3.2. Design dynamic gradient penalty , by monitoring the change of the gradient during the model training process in real time and dynamically adjusting the penalty coefficient according to the preset rules, while ensuring the convergence of the model, avoiding training failure or performance degradation caused by too large or too small gradients:

[0030]

[0031] Among them, represents the L2 norm, represents the true data corresponding to the ambient temperature label , represents the generated data output by the conditional generator for the ambient temperature label , represents the interpolation sample generated by linearly interpolating the real sample and the generated sample , represents the gradient output by the discriminator, , E[ ] represents taking the mathematical expectation.

[0032] Step 4: Generative Adversarial Network Training

[0033] Through the asymmetric gradient penalty strategy, update the discriminator designed in Step 3 and the conditional generator designed in Step 2 to improve the stability of the training process and the accuracy of data generation:

[0034] S4.1. Discriminator Update

[0035] By minimizing the loss function of the discriminator D , enabling the discriminator to better identify which samples are real and which are generated by the generator, so as to improve the overall performance of the model and the stability of the model training process:

[0036]

[0037] Among them, and respectively represent the outputs of the discriminator for the generated data and the generated data . and are constant coefficients.

[0038] S4.2. Conditional generator update

[0039] By minimizing the loss function of the conditional generator G, the conditional generator generates highly accurate data that satisfies the physical constraint conditions, and makes it difficult for the discriminator to distinguish between the generated data and the real data:

[0040]

[0041] Among them, the first term is the score of the discriminator D for the generated sample , and the second term is the voltage differential constraint, which is used to ensure that the generated sample satisfies specific physical constraint conditions; and and respectively represent the voltage, current, and remaining power data generated by the generator. is a constant representing the nominal capacitance of the lithium battery, is the constraint coefficient.

[0042] Step Five. Generation of lithium battery state data

[0043] The environmental temperature label and the Gaussian noise are concatenated and input into the trained conditional generator to obtain the lithium battery state data that meets the physical constraint conditions, and is restored to the original dimension through inverse normalization.

[0044] The present invention has the following beneficial effects:

[0045] 1. The present invention proposes a hybrid deep learning framework composed of a GRU - transposed convolution generation architecture and a dynamic gradient penalty mechanism. It generates lithium - battery state data based on environmental temperature labels and Gaussian noise, enhancing the physical consistency and temporal rationality of the generated data. This framework effectively improves the adaptability of the generated data to working conditions, and can stably generate highly accurate simulation data with high consistency with the battery's own state within a wide temperature range, thus improving the stability of the data generation process.

[0046] 2. In the proposed hybrid deep learning framework of the present invention, the GRU layer is responsible for capturing the dynamic change rules in time - series data and can accurately model the temporal characteristics during the battery charging and discharging process; the transposed convolution layer is used to reconstruct spatial features that conform to physical laws, further enhancing the authenticity of the generated data. The combination of the two enhances the model's ability to jointly model time information and spatial structure, especially suitable for the highly accurate synthesis of lithium - battery state data under complex working conditions.

[0047] 3. By introducing a dynamic gradient penalty mechanism to optimize the model training process, the stability of the training process is significantly improved, the convergence time is greatly shortened, and the consumption of computing resources is effectively reduced at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 is a flowchart of the generation of lithium - battery state data;

[0049] Figure 2 is a framework diagram of the data generation network model constructed in the embodiment;

[0050] Figure 3 is a schematic diagram of the conditional generator structure constructed in the embodiment;

[0051] Figure 4 is a schematic diagram of the discriminator structure constructed in the embodiment;

[0052] Figure 5 is a flowchart of the training process of the fused temporal adversarial generation network model. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0053] As Figure 1 shown, the present invention provides a method for generating lithium - battery state data based on a fused temporal adversarial generation network, aiming to generate high - precision lithium - battery state data and apply it to the evaluation of lithium - battery states. The following elaborates the implementation process of the present invention in detail with reference to the accompanying drawings:

[0054] Step 1. Pre - processing of lithium - battery state data

[0055] Collect the voltages of a lithium battery with a nominal capacity at 25 °C, 35 °C, and 45 °C, and currents and remaining battery power , and the corresponding ambient temperature label are respectively denoted as . The acquisition time span is 1000 seconds, and the sampling frequency is 1 Hz, forming an original battery state dataset.

[0056] Define the time period when the current value of the battery is positive during charging as the charging moment, and traverse each moment in the dataset , if it satisfies , then record the starting moment , and continue to traverse until the condition is not satisfied and then record the ending moment , forming a charging cycle . Extract the charging cycle from the original battery state dataset , where , represents the number of extracted charging cycles.

[0057] To ensure data integrity, select charging cycles with a length greater than 500 seconds. And through linear interpolation, make the number of data points in each charging cycle equal to that of the longest charging cycle. Normalize the interpolated charging cycle data.

[0058] Step 2: Design of the conditional generator

[0059] Design a conditional generator based on GRU - transposed convolution, as Figure 2 shown, using the temperature label and Gaussian noise as inputs to generate lithium - battery state data with the same length as the charging cycle :[[]]

[0060] S2.1 Label embedding layer

[0061] Denote the ambient temperature labels corresponding to different ambient temperature values as a set . Convert the ambient temperature label corresponding to the charging cycle into a one - hot encoding , and transform it into a learnable embedding vector , and together with the random noise z, pass it as the conditional information of the generator to the subsequent module. In this embodiment, the random noise is Gaussian noise.

[0062] S2.2, GRU - transposed convolution module

[0063] As Figure 3 shown, construct a GRU module, set the hidden - layer dimension to 128, and use as the activation function of the update gate and the reset gate, and use for the candidate hidden state. For the input embedding vector and Gaussian noise z, output a hidden state with a dimension of 128 .

[0064] The transposed convolution module includes 3 consecutive transposed convolutions, with a size of 4, a stride of 2, and a padding of 1. The input dimension of the first layer is 128, and the output dimension is 64. The input dimension of the second layer is 64, and the output dimension is 32. The input dimension of the third layer is 32, and the output dimension is 3. The spatial resolution of the last hidden state output by the GRU module is gradually restored through multiple convolutional layers to generate lithium battery state data equal in length to the charging cycle :

[0065]

[0066] Among them, is the th element of the output data is the position of the current predicted time point , represents the number of time points in one charging cycle after interpolation is the th convolutional layer in the transposed convolution module is the input hidden state th element of is the convolutional stride is the index of the convolutional kernel

[0067] Step 3: Discriminator design

[0068] S3.1. As Figure 4 shown, a 3-layer convolutional neural network (CNN) is used as the discriminator, and the dimension of each weight matrix is determined by the convolutional layer parameters. Design a spectral normalization layer based on the weight constraint strategy to perform a normalization operation on the weight matrix :

[0069]

[0070] Among them, is the weight matrix after normalization processing represents the calculation of the standard deviation, which is calculated by the power iteration method for 5 iterations is a learnable scaling factor with an initial value of 1

[0071] S3.2. Design dynamic gradient penalty :

[0072]

[0073] Among them, represents the L2 norm, represents the environmental temperature label corresponding true data, represents the interpolation sample generated by linearly interpolating the true sample and the generated sample and the generated sample, represents the gradient output by the discriminator. , E[ ] represents taking the mathematical expectation.

[0074] represents the penalty coefficient, with an initial value of 8. After every 100 iterations, it is adjusted according to the mean value of the gradient norm :

[0075]

[0076] Step 4. Generative adversarial network training

[0077] As Figure 5 shown, through the asymmetric gradient penalty strategy, the discriminator designed in Step 3 and the conditional generator designed in Step 2 are updated:

[0078] S4.1. Discriminator update

[0079] Minimize the loss function of the discriminator D:

[0080]

[0081] Among them, , respectively represent the outputs of the discriminator for the generated data and the generated data output. , are constant coefficients. In this embodiment, , are set.

[0082] S4.2. Conditional generator update

[0083] For the voltage, current, and remaining power data of the lithium battery, the following voltage differential constraint and remaining power constraint should be satisfied:

[0084]

[0085]

[0086] Therefore, for the voltage, current, and remaining power data generated by the conditional generator , , , set physical constraint conditions: when , it is considered that , , conforms to the battery charge-discharge law and the capacitance-current relationship. Introduce the physical constraint conditions into the loss function of the conditional generator G, and perform minimization training:

[0087]

[0088] where is the constraint coefficient, which is set to 0.1 in this embodiment.

[0089] Set the learning rate , select the Adam optimizer to update the parameters of the discriminator and the conditional generator. And it is set that after the discriminator is updated 5 times, the conditional generator is updated 1 time.

[0090] Step Five, Generate Lithium Battery State Data

[0091] Ambient temperature label and Gaussian noise are concatenated and input into the trained conditional generator to obtain the lithium battery state data that conforms to the physical constraint conditions, and is restored to the original dimension through inverse normalization.

Claims

1. A method for generating lithium battery state data based on a fused temporal adversarial generation network, which inputs the temperature label corresponding to the real battery state data and the random noise z into the generator to obtain the generated battery state data . The discriminator is used to compare the real battery state data with the generated data to train the adversarial generation network. The trained generator is used to generate lithium battery state data that meets the requirements. It is characterized in that: The generator includes a gated recurrent unit (GRU) module and a transposed convolution module. Among them, the bidirectional gated recurrent unit (GRU) module models the input data through a gating mechanism and outputs a hidden state ; The transposed convolution module performs spatial reconstruction on the hidden state through multiple consecutive convolutional layers and outputs generated data , , , respectively represent the generated lithium battery voltage, current, and remaining power; During the training process of the generative adversarial network, set the loss function of the discriminator D for the generated data and the real state data to compare the mathematical expectations between them, and introduce a dynamic penalty term for the output gradient of the discriminator ; set the loss function of the generator G as the mathematical expectation of the generated data and introduce a voltage differential constraint.

2. The method for generating lithium battery state data based on the fusion time-series adversarial generation network according to claim 1, wherein: Define the time period when the current value of the lithium battery is positive during charging as the charging time, and divide the continuous charging time data into a charging cycle as the true battery state data .

3. The method for generating lithium battery state data based on the fusion time series adversarial generation network according to claim 2, wherein: Equalize the number of time points for each charging cycle through linear interpolation and perform normalization to bring the voltage of the battery , current and remaining charge values into the same range.

4. The method for generating lithium battery state data based on the fusion time-series adversarial generation network according to claim 1, wherein: Convert the real battery state data corresponding temperature label into a learnable embedding vector , which is concatenated with the random noise z and then input into the generator; where l is the dimension of the ambient temperature tag of 5. The method for generating lithium battery state data based on the fusion time-series adversarial generation network according to claim 1, wherein: Set the hidden layer dimension of the gated recurrent unit (GRU) module to 128, and use as the activation functions for the update gate and the reset gate, and use .

6. The method for generating lithium battery state data based on the fusion time-series adversarial generation network according to claim 1 or 5, wherein: ； Among them, is the th element of the generated data, is the position of the currently predicted time point, , represents the number of time points in the true battery state data ; is the number of convolutional layers in the transposed convolution module; is the th convolutional layer in the transposed convolution module; is the th element of the hidden state, is the convolution stride.

7. The method for generating lithium battery state data based on the fusion time-series adversarial generation network according to claim 1, wherein: The discriminator is a 3-layer convolutional neural network, and the weight matrix of each layer The dimension is determined by the parameters of the convolutional layer, and through the spectral normalization layer based on the weight constraint strategy, the weight matrix is normalized.

8. The method for generating lithium battery state data based on the fusion time-series adversarial generation network according to claim 1 or 7, characterized in that: The dynamic gradient penalty : ； Among them, represents the L2 norm, represents the environmental temperature label corresponding true data, represents the interpolation sample generated by linearly interpolating the true sample and the generated sample ; represents the gradient output by the discriminator; , E[ ] represents taking the mathematical expectation; represents the penalty coefficient.

9. The method for generating lithium battery state data based on the fusion time series adversarial generation network according to claim 8, wherein: Set the penalty factor The initial value is 8, and after every 100 iterations, according to the gradient norm The mean Make adjustments when When Expand 1.2 times; when When Scale down by a factor of 0.8, otherwise keep the same.

10. The method for generating lithium battery state data based on the fusion time-series adversarial generation network according to claim 1, wherein: The voltage differential constraint is , where represents the nominal capacitance of the lithium battery, is the constraint coefficient.

Images