Lithium ion battery state-of-charge prediction method based on sparse adversarial self-coding, medium and equipment

Through sparse adversarial self-coding and self-attention mechanisms, combined with feature entropy and optimization algorithms, the problem of coupling relationship and initial value uncertainty in SOC and SOH estimation of lithium-ion batteries is solved, and high-precision and stable state of charge prediction are achieved. , improves the real-time monitoring capabilities of the battery management system.

CN118938024APending Publication Date: 2024-11-12NANJING INST OF TECH
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202411298092.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-09-14
Publication Date
2024-11-12

AI Technical Summary

Technical Problem

The prior art ignores the coupling relationship between the two in the estimation of SOC and SOH of lithium-ion batteries, resulting in a decrease in estimation accuracy. The SOC estimation method mainly focuses on the change process of 0% to 100%, and ignores the uncertainty of the initial value. The impact of lithium-ion batteries has led to challenges in high-precision prediction.

Method used

The method based on sparse adversarial autocoding is adopted to capture the global correlation between current, voltage and state of charge through the self-attention mechanism, dynamically optimize the sliding window size, and extract the deep features of the battery measurable variables using sparse adversarial autocoding. Combining feature entropy and improved loss function and optimization algorithm, joint estimation of SOC and SOH is achieved.

Benefits of technology

It significantly improves the accuracy and stability of state-of-charge prediction of lithium-ion batteries, reduces the computational complexity and overfitting risks, and provides efficient real-time monitoring support for battery management systems.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118938024A_ABST
    Figure CN118938024A_ABST
Patent Text Reader

Abstract

The invention provides a lithium ion battery charge state prediction method based on sparse adversarial self-coding, a medium and equipment, and the method comprises the steps: collecting measurable variables of a lithium ion battery under different operation conditions, and calculating the charge state through employing an ampere-hour integral method; a self-attention mechanism is adopted to capture global correlation between measurable variables of the lithium ion battery and the SOC, and the size of a sliding window is dynamically optimized according to correlation analysis; battery measurable variables in the sliding window are converted into a two-dimensional matrix, sparse adversarial self-coding is established, and deep features of the sparse adversarial self-coding are extracted; and constructing a feature entropy to estimate the complexity of the deep features, and sending the deep features to different feature analysis models such as an extreme learning machine, a least square support vector machine and a bidirectional long-short-term memory network according to a complexity value, so as to realize the high-precision prediction of the state of charge of the lithium ion battery. According to the method, the training efficiency of the model is remarkably improved, the calculation complexity is reduced, and the accuracy and stability of state-of-charge prediction are effectively improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of battery status diagnosis, and in particular relates to a method, medium and device for predicting the state of charge of a lithium-ion battery based on sparse adversarial autoencoding. Background Art

[0002] Lithium-ion batteries are widely used in energy storage systems, new energy vehicles and other fields due to their high energy density, long cycle life and low self-discharge rate. In the process of battery operation management, the state of charge (SOC) and the state of health (SOH) are important indicators for measuring battery performance, which can reflect the current charge level and overall health status of the battery in real time. However, these two state quantities are indirectly defined and are difficult to measure directly. They usually need to be estimated with the help of intelligent algorithms.

[0003] At present, a lot of research work is devoted to the independent estimation of battery SOC and SOH, that is, estimating another state quantity when one state quantity is determined. However, this type of method ignores the coupling relationship between the two, which is inconsistent with the actual operating conditions of lithium-ion batteries. Studies have shown that as the aging of lithium-ion batteries increases, the accuracy of independent estimation of SOC and SOH will seriously decrease. In comparison, the joint estimation method fully considers the coupling relationship between SOC and SOH, and can better adapt to the nonlinear changes of battery state. The model-based joint estimation method first establishes an equivalent circuit model, then uses dual observers to estimate SOC and SOH respectively, and finally realizes the joint estimation of the two through mutual iteration. However, this method has high requirements on the accuracy of battery modeling, and inaccurate battery parameters are bound to affect the accuracy of estimation. In addition, most SOC estimation methods only focus on the change process of SOC from 100% to 0%, while ignoring the impact of initial value uncertainty on the accuracy of SOC estimation. In view of this, high-precision prediction of lithium-ion batteries faces challenges, so it is urgent to develop a SOC and SOH estimation method with high accuracy and strong real-time performance. Summary of the invention

[0004] In view of the deficiencies in the prior art, the present invention provides a method, medium and device for predicting the state of charge of a lithium-ion battery based on sparse adversarial autoencoding.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] A method for predicting the state of charge of a lithium-ion battery based on sparse adversarial autoencoding is characterized by comprising:

[0007] S1: Collect measurable variables of lithium-ion batteries and calculate the state of charge using the ampere-hour integration method;

[0008] S2: Use self-attention mechanism to capture the global correlation between current, voltage and state of charge, and dynamically optimize the sliding window size according to the correlation;

[0009] S3: Convert the measurable variables of lithium-ion batteries in the sliding window into input data x of dimension O×P, and establish a sparse adversarial autoencoder to extract the deep features h of the input data x. deep , and adopt the loss function L SAAE to be trained;

[0010] S4: Construct feature entropy H s Estimate the deep feature h deep complexity;

[0011] S5: According to the characteristic entropy H s The value of the deep feature h deep The data is sent to different feature analysis models for analysis to predict the state of charge of the lithium-ion battery.

[0012] To optimize the above technical solutions, the specific measures taken also include:

[0013] Furthermore, in step S1, the measurable variables of the lithium-ion battery include voltage, current and temperature.

[0014] Furthermore, step S2 specifically includes the following sub-steps:

[0015] S2.1: Set an initial sliding window size;

[0016] S2.2: The correlation between the measurable variables and the state of charge of the lithium-ion battery in the sliding window is calculated using the self-attention mechanism. The self-attention mechanism is defined as Where Q, K, and V are embedded representations of measurable variables of lithium-ion batteries, d k is the embedding dimension;

[0017] S2.3: Adjust different sliding window sizes and calculate the correlation between the corresponding lithium-ion battery measurable variables and the state of charge;

[0018] S2.4: Repeat step S2.3 until the correlation value is maximized, thereby determining the optimal sliding window size.

[0019] Further, in step S3, the sparse adversarial autoencoder includes a generator and a discriminator, the generator includes a convolutional encoder and a convolutional decoder, and the discriminator includes a convolutional encoder and a fully connected layer; the sparse adversarial autoencoder processes the input data x as follows:

[0020] S3.1.1: Input data x is sent to the convolutional encoder e1 of the generator for processing: first, feature extraction is performed by Q1 convolutional filters C1 with a window size of c×c and a step size of q, then feature activation is performed by the ReLU activation function, and finally feature dimension reduction is performed by the maximum pooling layer P1 with a window size of o×o and a step size of l, and the output dimension is a deep feature h1 with a dimension of a1×b1×c1;

[0021] S3.1.2: The deep feature h1 is processed by the convolutional encoders e2, e3, and e4 of the generator in turn. The data processing process of the convolutional encoders e2, e3, and e4 is the same as that of the convolutional encoder e1, and the final output dimension is a4×b4×c4 deep feature h deep ;

[0022] S3.1.3: Deep features h deep The convolution decoder d1 of the generator is sent for processing: first, the upsampling layer U1 with a window size of g×g performs feature dimension upscaling, and then the convolution filter C5 and the ReLU activation function are used to process the deep feature h4;

[0023] S3.1.4: The deep feature h4 is processed by the generator's convolutional decoders d2, d3, and d4 in turn. The data processing process of the convolutional decoders d2, d3, and d4 is the same as that of the convolutional decoder d1, and the final output is the reconstructed data z with a dimension of O×P;

[0024] S3.1.5: The input data x and the reconstructed data z are input into the discriminator at the same time. They are first processed by the convolutional encoders e5, e6, and e7 of the discriminator, and then processed by flattening F and full connection f to determine whether they are "true" or "false", where the input data x is "true" and the reconstructed data z is "false".

[0025] Furthermore, in step S3, the loss function L SAAE for:

[0026] L SAAE =aL MSE +βL sparse +γL cc +δL GAN ;

[0027] In the formula, α, β, γ and δ all represent weight coefficients; L MSE is the mean square error loss function, O and P are the dimensions of the input data, i is the index of O, j is the index of P, and x is the ij is the input data at position (i, j), z ij is the reconstructed input data at position (i, j); KL represents divergence, which is used to measure the difference between the sparsity target and the actual activation rate, φ represents the sparsity target, Represents the actual activation rate of the i-th hidden layer unit; Δi and Δj represent the horizontal and vertical displacements of the input data x and the reconstructed data z, respectively; G(-) represents the generator output, D(-) represents the discriminator output, P(x) and P(z) are the probability distributions of the input data x and the reconstructed data z, respectively, and E P(x) is the expectation of P(x), E P(z) is the expectation of P(z).

[0028] Furthermore, in step S3, during the training process of the sparse adversarial autoencoder, the rime optimization algorithm is used to select the convolution kernel size and the number of hidden layer neurons, as follows:

[0029] S3.2.1: Set the population size N, generate the initial population, and each individual Represents a structural solution of sparse adversarial autoencoder, each individual The position of represents the combination of the convolution kernel size and the number of hidden layer neurons; the search range is set to Candidate solutions for the convolution kernel size and the number of hidden layer neurons are randomly generated by initialization;

[0030] S3.2.2: Set the maximum number of iterations MaxIter, cooling sparseness T, and freezing factor

[0031] S3.2.3: Iterative Updates In the t+1th generation, The update process is in represents the combination of convolution kernel size and number of hidden layer neurons of the i-th individual in the t+1th generation, represents the current global optimal solution, represents the local optimal solution, η represents the adaptive random variable, and the chaotic variable The update formula is represents the chaotic variable of the tth generation;

[0032] S3.2.4: Freeze Factor As the number of iterations decreases, it is used to control the search step size during the iteration, the freezing factor in the tth generation represents the initial freezing factor;

[0033] S3.2.5: For each new solution Perform fitness evaluation and obtain the fitness function value

[0034] S3.2.5: Update the global optimal solution based on the fitness evaluation results and local optimal solution

[0035] S3.2.6: Store the global optimal solution into the memory pool M, and randomly select a memory optimal solution M from the memory pool M best As the current global optimal solution, return to step S3.2.3 to perform Updates;

[0036] S3.2.7: When the algorithm iteration reaches the maximum number of MaxIter or the global optimal solution When the change in is less than a certain threshold, the algorithm stops and outputs the optimal convolution kernel size and number of hidden layer neurons.

[0037] Furthermore, in step S4, the characteristic entropy H s for:

[0038]

[0039] In the formula, p(h deep,i ) represents the i-th deep feature h deep,i The probability distribution of θ and are two control parameters, and n represents the total number of deep features.

[0040] Further, in step S5, according to the characteristic entropy H s The value of the deep feature h deep Sent to different feature analysis models for analysis, specifically:

[0041] like The deep features are sent to the extreme learning machine;

[0042] like The deep features are sent to the least squares support vector machine;

[0043] like The deep features are sent to the bidirectional long short-term memory network, where Both represent threshold values.

[0044] Accordingly, the present invention proposes a computer-readable storage medium storing a computer program, characterized in that the computer program enables a computer to execute the lithium-ion battery state of charge prediction method based on sparse adversarial autoencoding as described above.

[0045] Accordingly, the present invention proposes an electronic device, characterized in that it includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the method for predicting the state of charge of a lithium-ion battery based on sparse adversarial autoencoding as described above is implemented.

[0046] The beneficial effects of the present invention are:

[0047] (1) The present invention designs a self-attention mechanism to capture the global correlation between current, voltage and SOC, dynamically selects the optimal sliding window size, and can effectively smooth the fluctuations of data such as voltage and current, reduce the impact of data noise, and thus improve the accuracy and stability of SOC estimation;

[0048] (2) The present invention establishes a sparse adversarial autoencoder for extracting deep features of battery measurable variables, adopts a new loss function to improve training efficiency, and introduces an improved rime optimization algorithm to optimize the convolution kernel size and the number of hidden layer neurons, thereby significantly enhancing the learning ability of representative features of batteries;

[0049] (3) The present invention constructs a feature entropy to estimate the complexity of the deep features of the battery, and assigns the features to different prediction models according to the complexity values, thereby effectively improving the computational efficiency, reducing the model complexity and overfitting risk, and further improving the accuracy and stability of the prediction.

[0050] In summary, the present invention, based on sparse adversarial autoencoding, realizes the deep feature extraction of measurable variables of lithium-ion batteries, and effectively improves the accuracy and stability of SOC prediction. At the same time, with the help of improved loss function and optimization algorithm, the training efficiency of the model is significantly improved, the computational complexity is reduced, and reliable technical support is provided for efficient prediction and real-time monitoring of battery management systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 The present invention is a flow chart of a method for predicting the state of charge of a lithium-ion battery based on sparse adversarial autoencoding.

[0052] Figure 2 This is a sliding window size optimization method based on the attention mechanism of the present invention.

[0053] Figure 3 This is a schematic diagram of the network structure for deep feature extraction of battery measurable variables based on sparse adversarial autoencoding according to the present invention.

[0054] Figure 4 Schematic diagram of the feature analysis model operation strategy based on feature entropy of the present invention.

[0055] Figure 5 Schematic diagram of the structure of the extreme learning machine of the present invention.

[0056] Figure 6 It is a schematic diagram of the least squares support vector machine structure of the present invention.

[0057] Figure 7 It is a schematic diagram of the bidirectional long short-term memory network structure of the present invention.

[0058] Figure 8a and Figure 8b The schematic diagram of the state of charge prediction result of a lithium-ion battery using the method of the present invention is shown in FIG. Figure 8a This is a schematic diagram of the comparison between the predicted value and the actual value of the state of charge. Figure 8b Schematic diagram of state of charge prediction error. DETAILED DESCRIPTION

[0059] The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0060] In one embodiment, the present invention proposes a method for predicting the state of charge of a lithium-ion battery based on sparse adversarial autoencoding, such as Figure 1 As shown, the specific steps include the following steps.

[0061] Step 1: Collect the measurable variables of the lithium-ion battery (such as voltage, current and temperature) and calculate the state of charge (SOC) using the ampere-hour integration method.

[0062] Step 2: Use the self-attention mechanism to capture the global correlation between current, voltage and SOC, and dynamically optimize the sliding window size based on the correlation analysis. The specific process is as follows: Figure 2 As shown, it includes the following sub-steps:

[0063] S2.1: Set an initial sliding window size;

[0064] S2.2: The correlation between measurable variables such as current, voltage, temperature and SOC is calculated using the self-attention mechanism. The self-attention mechanism is defined as Among them, Q (Query), K (Key), and V (Value) are embedded representations of the measurable variables of the battery, d k is the embedding dimension;

[0065] S2.3: Adjust different sliding window sizes and calculate the global correlation between current, voltage and SOC;

[0066] S2.4: Repeat step S2.3 until the correlation value is maximized, thereby determining the optimal sliding window size.

[0067] Step 3: Convert the measurable variables of the battery in the sliding window into input data x of dimension O×P, and establish a sparse adversarial autoencoder (SAAE) to extract its deep features h deep , and adopt a new loss function L SAAE Be trained.

[0068] Sparse adversarial autoencoder structure such as Figure 3 As shown, it includes a generator and a discriminator. The generator includes a convolutional encoder and a convolutional decoder, and the discriminator includes a convolutional encoder and a fully connected layer. The specific implementation steps of sparse adversarial autoencoding are:

[0069] S3.1.1: Input data x is sent to the convolution encoder e1 for processing, which includes: first, feature extraction is performed by Q1 convolution filters C1 with a window size of c×c and a step size of q, then feature activation is performed by the ReLU activation function, and finally feature dimension reduction is performed by the maximum pooling layer P1 with a window size of o×o and a step size of l, and the output dimension is a deep feature h1 with a dimension of a1×b1×c1;

[0070] S3.1.2: The deep feature h1 is processed by convolutional encoders e2, e3, and e4 in sequence. The data processing process is similar to step S3.1.1, and the final output is a deep feature h with a dimension of a4×b4×c4. deep ;

[0071] S3.1.3: Deep features h deep It is sent to the convolution decoder d1 for processing. First, the upsampling layer U1(g×g) with a g×g window size performs feature dimension upscaling, and then it is processed by the convolution filter C5(Q5@c×c+q) and the ReLU activation function to obtain the deep feature h4.

[0072] S3.1.4: The deep feature h4 is processed by the convolution decoders d2, d3, and d4 in sequence. The data processing process is similar to step S3.1.3, and the final output is the reconstructed data z with a dimension of O×P;

[0073] S3.1.5: Input data x and reconstructed data z are input into the discriminator at the same time, first processed by convolution encoders e5, e6, and e7, and then processed by flattening F and full connection f to determine whether it is "true" or "false", where input data x is "true" and reconstructed data z is "false".

[0074] New loss function L used in sparse adversarial autoencoder training SAAE Specifically:

[0075] L SAAE =αL MSE +βL sparse +γLcc +δL GAN

[0076] Among them, α, β, γ and δ all represent weight coefficients; L M SE is the mean square error loss function, O and P are the dimensions of the input data, i is the index of O, j is the index of P, and x is the ij is the input data at position (i, j), z ij is the reconstructed input data at position (i, j); KL represents divergence, which is used to measure the difference between the sparsity target and the actual activation rate, φ represents the sparsity target (i.e., the expected activation rate), Represents the actual activation rate of the i-th hidden layer unit; Where Δi and Δj represent the horizontal and vertical displacements of the input data x and the reconstructed data z; G(-) represents the generator output, D(-) represents the discriminator output, P(x) and P(z) are the probability distribution of input data x and the probability distribution of reconstructed input data z, respectively, E P(x) is the expectation of P(x), E P(z) is the expectation of P(z).

[0077] During the training of the sparse adversarial autoencoder, the rime optimization algorithm is used to select the convolution kernel size and the number of hidden layer neurons, as follows:

[0078] S3.2.1: Initialize the population. Set the population size N and generate the initial population Each individual represents a structural solution of SAAE. It includes two dimensions, namely the convolution kernel size and the number of hidden layer neurons, and the search range is Candidate solutions for the convolution kernel size and the number of hidden layer neurons are randomly generated by initialization;

[0079] S3.2.2: Initialize parameters. Set the maximum number of iterations MaxIter, cooling sparseness T, and freezing factor By controlling the optimization process through these parameters, the algorithm can gradually converge to the optimal network structure;

[0080] S3.2.3: Location update. Each individual The position of represents the combination of the convolution kernel size and the number of hidden layer neurons. In the t+1th generation, The update process is in represents the combination of convolution kernel size and number of hidden layer neurons of the i-th individual in the t+1th generation, represents the current global optimal solution, represents the local optimal solution, η represents the adaptive random variable, and the chaotic variable The update formula is represents the chaotic variable of the tth generation, ensuring that the algorithm balances the local and global optimal solutions during the search process;

[0081] S3.2.4: Adaptive freezing mechanism. Set the freezing factor As the number of iterations decreases, the freezing factor in the tth generation is used to prevent the algorithm from falling into the local optimum. represents the initial freezing factor;

[0082] S3.2.5: Fitness evaluation. For each new solution The fitness is evaluated through SAAE performance, and the objective function is

[0083] S3.2.5: Optimal solution update. In each generation, the global optimal solution is updated according to the fitness evaluation results. and local optimal solution Ensure that the algorithm gradually converges to the optimal combination of convolution kernel size and number of hidden layer neurons;

[0084] S3.2.6: Memory mechanism. At each iteration, the global optimal solution is stored in the memory pool M. When a new solution is generated, a memory optimal solution M is randomly selected from the memory pool M. best As the current global optimal solution, return to step S3.2.3 to perform Updates;

[0085] S3.2.7: Termination condition. When the algorithm iteration reaches the maximum number of MaxIter or the global optimal solution When the change of is less than a certain threshold, the algorithm stops. At this time, the optimal convolution kernel size and number of hidden layer neurons are output.

[0086] The present invention innovatively introduces a freezing mechanism and a chaotic search mechanism, which are used to regulate the operation of the algorithm in a specific way to avoid local optimal traps and improve global search efficiency. The specific operation mechanism is as follows:

[0087] (1) Freezing mechanism

[0088] The freezing mechanism works by gradually reducing the freezing factor To limit the search range of the algorithm. This mechanism controls how the algorithm adjusts its search step size during the iteration process to achieve the transition from global exploration to local convergence. The specific implementation is:

[0089]

[0090] in, is the initial freezing factor, t is the current number of iterations, and MaxIter is the maximum number of iterations. As t increases, Gradually decreasing means that the algorithm has a larger search step size at the beginning of the iteration, which can perform global exploration, but when it is close to termination, the step size is reduced to ensure that the algorithm can search more accurately in the local area. This mechanism gradually guides the algorithm from global search to local convergence by limiting the size of the search step size.

[0091] (2) Chaotic Search Mechanism

[0092] Chaos search introduces the chaotic variable ξ i To limit the algorithm's search behavior near the local optimum. Chaotic variables are generated according to the Logistic map, and the update formula is:

[0093]

[0094] This chaotic variable ξ i Substituted into the individual position update formula to control the search range of the individual relative to the local optimal solution:

[0095]

[0096] in, Represents the local optimal solution. By introducing chaotic variables, the algorithm's search in the local area is no longer linear, but is affected by random disturbances, making the search path more diverse and avoiding premature convergence to the local optimum.

[0097] In summary, the freezing mechanism limits the algorithm's search step size, allowing for large-scale exploration in the early stages of the search and gradual refinement in the later stages. Chaotic search, on the other hand, limits the search mode of individuals near the local optimal solution by introducing chaotic variables, enhancing randomness and avoiding falling into the local optimum. Both mechanisms restrict and regulate the algorithm in different aspects, ensuring that the algorithm can perform global search and local convergence more effectively.

[0098] Step 4: Construct feature entropy H s Estimate the deep feature h deep Complexity, Where p(h deep,i ) represents the i-th deep feature h deep,i The probability distribution of θ and are two control parameters, and n represents the total number of deep features.

[0099] Step 5: According to the characteristic entropy H s The numerical value of hdeep is sent to different feature analysis models. The specific feature entropy operation strategy is as follows: Figure 4 If The deep features are sent to the extreme learning machine (ELM); if The deep features are sent to the least squares support vector machine (LSSVM); if The deep features are sent to the Bidirectional Long Short Term Memory (BiLSTM) network, where Both represent thresholds. The structure of the extreme learning machine ELM is as follows Figure 5 As shown, the structure of the least squares support vector machine LSSVM is as follows Figure 6 As shown, the structure of the bidirectional long short-term memory network BiLSTM is as follows Figure 7 As shown in the figure, BiLSTM connects two LSTM units with opposite time sequences to obtain bidirectional data information of the input sequence.

[0100] Figure 8a 3 is a schematic diagram of the comparison between the predicted state of charge and the true value. It can be seen that the lithium ion state of charge predicted by the method of the present invention is basically the same as the true value, indicating that the method of the present invention has a very high accuracy in predicting the state of charge of lithium ion batteries. Figure 8b Schematic diagram of the state of charge prediction error. It can be seen that the errors of the lithium ion state of charge predicted by the present invention are all less than 4%, indicating that the present invention has high prediction accuracy for the lithium ion battery state of charge.

[0101] In another embodiment, the present invention provides a computer-readable storage medium storing a computer program, wherein the computer program enables a computer to execute the method for predicting the state of charge of a lithium-ion battery based on sparse adversarial autoencoding as described in the first embodiment.

[0102] In another embodiment, the present invention proposes an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the method for predicting the state of charge of a lithium-ion battery based on sparse adversarial autoencoding as described in Example 1 is implemented.

[0103] In the embodiments disclosed in the present application, the computer storage medium can be a tangible medium that can contain or store a program for use by or in conjunction with an instruction execution system, device or equipment. The computer storage medium can include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices or equipment, or any suitable combination of the above. More specific examples of computer storage media can include electrical connections based on one or more lines, portable computer disks, hard disks, random access memories (RAM), read-only memories (ROM), erasable programmable read-only memories (EPROM or flash memory), optical fibers, portable compact disk read-only memories (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the above.

[0104] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed in this application can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0105] The above are only preferred embodiments of the present invention, and the protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should be regarded as the protection scope of the present invention.

Claims

1. A lithium-ion battery state of charge prediction method based on sparse adversarial autoencoding, characterized in that: include: S1: Collect measurable variables of lithium-ion batteries and calculate the state of charge using the ampere-hour integration method; S2: Use self-attention mechanism to capture the global correlation between current, voltage and state of charge, and dynamically optimize the sliding window size according to the correlation; S3: Convert the measurable variables of lithium-ion batteries in the sliding window into input data x of dimension O×P, and establish a sparse adversarial autoencoder to extract the deep features h of the input data x. deep , and adopt the loss function L SAAE to be trained; S4: Construct feature entropy H s Estimate the deep feature h deep complexity; S5: According to the characteristic entropy H s The value of the deep feature h deep The data is sent to different feature analysis models for analysis to predict the state of charge of the lithium-ion battery.

2. The method for predicting the state of charge of a lithium-ion battery based on sparse adversarial autoencoding according to claim 1, characterized in that: In step S1, the measurable variables of the lithium-ion battery include voltage, current and temperature.

3. The method for predicting state of charge of a lithium-ion battery based on sparse adversarial autoencoding according to claim 1, characterized in that: Step S2 specifically includes the following sub-steps: S2.1: Set an initial sliding window size; S2.2: The correlation between the measurable variables and the state of charge of the lithium-ion battery in the sliding window is calculated using the self-attention mechanism. The self-attention mechanism is defined as Where Q, K, and V are embedded representations of measurable variables of lithium-ion batteries, d k is the embedding dimension; S2.3: Adjust different sliding window sizes and calculate the correlation between the corresponding lithium-ion battery measurable variables and the state of charge; S2.4: Repeat step S2.3 until the correlation value is maximized, thereby determining the optimal sliding window size.

4. The method for predicting the state of charge of a lithium-ion battery based on sparse adversarial autoencoding according to claim 1, characterized in that: In step S3, the sparse adversarial autoencoder includes a generator and a discriminator, the generator includes a convolutional encoder and a convolutional decoder, and the discriminator includes a convolutional encoder and a fully connected layer; the sparse adversarial autoencoder processes the input data x as follows: S3.1.1: Input data x is sent to the convolutional encoder e1 of the generator for processing: first, feature extraction is performed by Q1 convolutional filters C1 with a window size of c×c and a step size of q, then feature activation is performed by the ReLU activation function, and finally feature dimension reduction is performed by the maximum pooling layer P1 with a window size of o×o and a step size of l, and the output dimension is a deep feature h1 with a dimension of a1×b1×c1; S3.1.2: The deep feature h1 is processed by the convolutional encoders e2, e3, and e4 of the generator in turn. The data processing process of the convolutional encoders e2, e3, and e4 is the same as that of the convolutional encoder e1, and the final output dimension is a4×b4×c4 deep feature h deep ; S3.1.3: Deep features h deep The convolution decoder d1 of the generator is sent for processing: first, the upsampling layer U1 with a window size of g×g performs feature dimension upscaling, and then the convolution filter C5 and the ReLU activation function are used to process the deep feature h4; S3.1.4: The deep feature h4 is processed by the generator's convolutional decoders d2, d3, and d4 in turn. The data processing process of the convolutional decoders d2, d3, and d4 is the same as that of the convolutional decoder d1, and the final output is the reconstructed data z with a dimension of O×P; S3.1.5: Input data x and reconstructed data z are input into the discriminator at the same time. They are first processed by the convolutional encoders e5, e6, and e7 of the discriminator, and then processed by flattening F and full connection f to determine whether they are "true" or "false", where input data x is "true" and reconstructed data z is "false".

5. The method for predicting the state of charge of a lithium-ion battery based on sparse adversarial autoencoding according to claim 4, characterized in that: In step S3, the loss function L SAAE for: L SAAE =αL MSE +βL sparse +γL cc +δL GAN ; In the formula, α, β, γ and δ all represent weight coefficients; L MS E is the mean square error loss function, O and P are the dimensions of the input data, i is the index of O, j is the index of P, and x is the ij is the input data at position (i, j), z ij is the reconstructed input data at position (i, j); KL represents divergence, which is used to measure the difference between the sparsity target and the actual activation rate, φ represents the sparsity target, Represents the actual activation rate of the i-th hidden layer unit; Δi and Δj represent the horizontal and vertical displacements of the input data x and the reconstructed data z, respectively; G(·) represents the generator output, D(·) represents the discriminator output, P(x) and P(z) are the probability distributions of input data x and reconstructed data z, respectively, and E P(x) is the expectation of P(x), E P(z) is the expectation of P(z).

6. The method for predicting state of charge of a lithium-ion battery based on sparse adversarial autoencoding according to claim 1, characterized in that: In step S3, during the training process of the sparse adversarial autoencoder, the rime optimization algorithm is used to select the convolution kernel size and the number of hidden layer neurons, as follows: S3.2.1: Set the population size N, generate the initial population, and each individual g i Represents a structural solution of sparse adversarial autoencoder, each individual The position of represents the combination of the convolution kernel size and the number of hidden layer neurons; the search range is set to Candidate solutions for the convolution kernel size and the number of hidden layer neurons are randomly generated by initialization; S3.2.2: Set the maximum number of iterations MaxIter, cooling sparseness T, and freezing factor S3.2.3: Iterative Updates In the t+1th generation, The update process is in represents the combination of convolution kernel size and number of hidden layer neurons of the i-th individual in the t+1th generation, represents the current global optimal solution, represents the local optimal solution, η represents the adaptive random variable, and the chaotic variable The update formula is represents the chaotic variable of the tth generation; S3.2.4: Freeze Factor As the number of iterations decreases, it is used to control the search step size during the iteration, the freezing factor in the tth generation represents the initial freezing factor; S3.2.5: For each new solution Perform fitness evaluation and obtain the fitness function value S3.2.5: Update the global optimal solution based on the fitness evaluation results and local optimal solution S3.2.6: Store the global optimal solution into the memory pool M, and randomly select a memory optimal solution M from the memory pool M best As the current global optimal solution, return to step S3.2.3 to perform Updates; S3.2.7: When the algorithm iteration reaches the maximum number of MaxIter or the global optimal solution When the change in is less than a certain threshold, the algorithm stops and outputs the optimal convolution kernel size and number of hidden layer neurons.

7. The method for predicting state of charge of a lithium-ion battery based on sparse adversarial autoencoding according to claim 1, characterized in that: In step S4, the characteristic entropy H s for: In the formula, p(h deep,i ) represents the i-th deep feature h deep,i The probability distribution of θ and are two control parameters, and n represents the total number of deep features.

8. The method for predicting the state of charge of a lithium-ion battery based on sparse adversarial autoencoding according to claim 7, characterized in that: In step S5, the characteristic entropy H s The value of the deep feature h deep Sent to different feature analysis models for analysis, specifically: like The deep features are sent to the extreme learning machine; like The deep features are sent to the least squares support vector machine; like The deep features are sent to the bidirectional long short-term memory network, where Both represent threshold values.

9. A computer-readable storage medium storing a computer program, characterized in that: The computer program enables a computer to execute the method for predicting the state of charge of a lithium-ion battery based on sparse adversarial autoencoding as described in any one of claims 1 to 8.

10. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the method for predicting the state of charge of a lithium-ion battery based on sparse adversarial autoencoding as described in any one of claims 1 to 8 is implemented.

Citation Information

Cited By

  • Storage battery operation state defect identification method and device

    CN119438907A