Early warning method and system for energy storage battery performance degradation

Abstract

The application discloses an early warning method for performance degradation of an energy storage battery, which comprises the following steps: obtaining a capacity degradation data set of the energy storage battery and preprocessing the capacity degradation data set to obtain a training data set; performing wavelet analysis to obtain multi-dimensional intrinsic components at different frequencies; performing principal component analysis to obtain a core frequency signal; constructing an early warning initial model for performance degradation of the energy storage battery and training the early warning initial model by using the core frequency signal to obtain an early warning model for performance degradation of the energy storage battery; and performing actual early warning for performance degradation of the energy storage battery by using the early warning model for performance degradation of the energy storage battery. The application further discloses a system for implementing the early warning method for performance degradation of the energy storage battery. The early warning method for performance degradation of the energy storage battery has higher reliability, better accuracy and better universality.

Description

Technical Field

[0001] This invention belongs to the field of electrical automation, and specifically relates to an early warning method and system for the performance degradation of energy storage batteries. Background Technology

[0002] With economic and technological development and the improvement of people's living standards, electricity has become an indispensable secondary energy source in people's production and daily life, bringing endless convenience. Therefore, ensuring a stable and reliable supply of electricity has become one of the most important tasks of the power system.

[0003] Currently, a large number of new energy power generation systems are being integrated into the power grid. However, the randomness of the output from these systems puts immense pressure on the power system's operation. To better absorb the power from new energy sources, and to better smooth peak and valley loads and ensure stable and reliable operation, the power system has begun to gradually equip itself with energy storage battery systems.

[0004] The integration of energy storage battery systems into the power system greatly ensures the stable and reliable operation of the power system. Therefore, predicting and providing early warnings of energy storage battery performance is extremely important. Currently, traditional early warning schemes for energy storage battery performance degradation often employ various types of neural network schemes for prediction, such as AlexNet, LSTM, or WaveNet-related schemes. However, these schemes still suffer from low reliability and poor accuracy. Summary of the Invention

[0005] One of the objectives of this invention is to provide an early warning method for the performance degradation of energy storage batteries that is highly reliable, accurate, and versatile.

[0006] The second objective of this invention is to provide a system for an early warning method of the performance degradation of the energy storage battery.

[0007] The early warning method for performance degradation of energy storage batteries provided by this invention includes the following steps:

[0008] S1. Obtain the capacity degradation dataset of the energy storage battery;

[0009] S2. Perform data preprocessing on the data obtained in step S1 to obtain the training dataset;

[0010] S3. Perform wavelet analysis on the training dataset obtained in step S2 to obtain multi-dimensional eigencomponents at different frequencies;

[0011] S4. Perform principal component analysis on the intrinsic component data information obtained in step S3 to obtain the core frequency signal;

[0012] S5. Based on the Informer network, construct an initial model for early warning of energy storage battery performance degradation;

[0013] S6. Using the core frequency signal obtained in step S4, train the initial model constructed in step S5 to obtain an early warning model for the performance degradation of the energy storage battery.

[0014] S7. Using the early warning model for energy storage battery performance degradation obtained in step S6, perform early warning for actual energy storage battery performance degradation.

[0015] Step S1, which involves obtaining the capacity degradation dataset of the energy storage battery, specifically includes the following steps:

[0016] The capacity degradation dataset of energy storage batteries is obtained using the following method;

[0017] Obtain a set of energy storage batteries;

[0018] The energy storage battery is fully charged in a set ambient temperature environment using the prescribed method. After the energy storage battery is placed in the test environment for a set time, it is discharged to the cutoff voltage at a 1C rate. At the same time, the capacity data of the energy storage battery during the discharge process is recorded.

[0019] Repeat the above process several times, record the capacity data each time, and take the average value to obtain the final energy storage battery capacity degradation dataset.

[0020] Step S3, which involves performing wavelet analysis on the training dataset obtained in step S2 to obtain multi-dimensional eigencomponents at different frequencies, specifically includes the following steps:

[0021] A. Select wavelet basis functions suitable for the capacity degradation dataset of energy storage batteries; the wavelet basis functions are expressed as:

[0022]

[0023] In the formula ψ a,b ψ(t) is the wavelet basis function; ψ(t) is the wavelet function before transformation; t is the time-domain variable of the function; a is the scale of the wavelet analysis; b is the translation amount of the wavelet analysis.

[0024] B. Select the wavelet basis function ψ a,b The inner product of f(t) and the capacity degradation data f(t) of the energy storage battery to be analyzed is calculated, and the coefficient C is used to represent the relationship between f(t) and ψ. a,b The degree of similarity between (t);

[0025] C. Shifting the wavelet basis function k units to the right yields the second wavelet basis function ψ. a,b(tk), and repeat step B until f(t) ends;

[0026] D. Extending the wavelet basis functions yields the third wavelet basis function. And repeat steps B and C;

[0027] E. Repeat step D to expand the wavelet basis functions until the set conditions are met;

[0028] Finally, the general formula for continuous wavelet transform is obtained as follows:

[0029]

[0030] In the formula, CWT(a,b) is the total function of the continuous wavelet transform;

[0031] Based on the general formula of continuous wavelet transform, the multidimensional eigencomponents at different frequencies are obtained.

[0032] Step S4, which involves performing principal component analysis on the intrinsic component data obtained in step S3 to obtain the core frequency signal, specifically includes the following steps:

[0033] Obtain the intrinsic component data information obtained in step S3, which includes n frequency signals of wavelet components. Each frequency signal sequence includes m index variables, where the j-th index of the i-th frequency signal is represented by x. ij ;

[0034] The indicators are standardized using the following formula:

[0035]

[0036] In the formula This is the j-th index of the standardized i-th frequency signal; s is the sample mean of the j-th indicator; j Let be the standard deviation of the j-th indicator;

[0037] The correlation coefficient matrix R is calculated using the following formula:

[0038]

[0039] In the formula r ij Let be the correlation coefficient between the i-th indicator and the j-th indicator;

[0040] Calculate the m eigenvalues ​​λ1 to λ2 of the correlation coefficient matrix R. m and the corresponding eigenvectors u1~u m ; where the j-th eigenvector u j Represented as u j =(u 1j ,u2j ,...,u nj ) T , λ1≥λ2≥...≥λ m ;

[0041] Based on the feature vectors u1~u m The following formula is used to calculate m index variables:

[0042]

[0043] In the formula y m It is the m-th principal component; Let n be the standardized indicator variable;

[0044] Calculate the information contribution rate and cumulative contribution rate of m eigenvalues:

[0045]

[0046]

[0047] In the formula b j For the j-th principal component y j Information contribution rate; α j For the j-th principal component y j The cumulative contribution rate;

[0048] Based on the cumulative contribution rate of the m eigenvalues, the following judgment is made:

[0049] If |α p -1|≤ε T Then, the principal components corresponding to the first p eigenvalues ​​are selected as the core frequency signal; ε T The threshold value is set.

[0050] Step S5, which involves constructing an initial early warning model for energy storage battery performance degradation based on the Informer network, specifically includes the following steps:

[0051] Based on the Informer network, an initial model for early warning of energy storage battery performance degradation is constructed.

[0052] The model includes an encoder, a decoder, and a fully connected layer;

[0053] The encoder uses a probabilistic self-attention mechanism to encode the input sequence and then inputs the encoded data into the decoder.

[0054] The decoder uses a one-time generative prediction method to decode the input encoded data and uploads the decoded data to the fully connected layer;

[0055] Fully connected layers are used to obtain the final output result based on the input decoded data.

[0056] The input sequence specifically includes the following:

[0057] The core frequency signal obtained in step S4 is used as the input multivariate sequence and expressed as:

[0058]

[0059] In the formula X t For input multivariate sequences; L x The length of the current input sequence; Let be the i-th point in the input sequence at time t; Let the vector dimension be the number of points in the sequence.

[0060] The model employs a dot product self-attention mechanism and uses timestamps to encode location information. The timestamp calculation includes local timestamps (PE) and global timestamps (SE).

[0061] The local timestamp (PE) is calculated using the following formula:

[0062]

[0063] In the formula, PE(pos,2j) is the local timestamp; pos is the location information; [] represents the rounding operation; d model The feature dimensions after input;

[0064] After encoding the timestamps and aligning them with the input dimensions, we obtain the model's input representation vector:

[0065]

[0066] In the formula The final input data is the data fed into the encoder; α is a factor that balances the size between the scalar mapping and the local / global embedding. Let i be a characteristic scalar, i∈{1,...,L} x}; Local timestamp; Global timestamp; This indicates that there are p types of global timestamps.

[0067] The encoder specifically includes the following:

[0068] Self-attention is defined when receiving a tuple input (Q,K,V) and performing a scaled dot product; where Q is the query value, K is the key value, and V is the value value.

[0069] (Q,K,V) is obtained through a linear transformation of the input matrix X, and is expressed as:

[0070] Q = XW Q

[0071] K = XW K

[0072] V = XW V

[0073] In the formula W Q W is the query parameter matrix to be trained. K W is the key parameter matrix to be trained; V The parameter matrix is ​​the value to be trained;

[0074] The self-attention mechanism is obtained:

[0075]

[0076] In the formula, Attention(Q,K,V) is the calculation formula for the attention mechanism; Softmax() is the normalized exponential function; d is the dimension of the input vector;

[0077] Use q i Represents the i-th row of Q, k i Represents the i-th row of K, v i Let V represent the i-th row. The attention for obtaining the i-th query is defined as the probabilistic form of kernel smoothing:

[0078]

[0079] In the formula A(q) i K,V) represents the attention of the i-th query; k(q) i ,k j ) is an asymmetric exponential kernel function, and v j The j-th row of V; This is the kernel-smoothed probability form for attention;

[0080] According to the above formula, the attention of the i-th query on all keys is defined as a probability distribution p(k j |q i ), Here we introduce a uniform distribution

[0081] The KL divergence is used to measure the similarity between probability distribution p and uniform distribution q:

[0082]

[0083] In the formula, KL(q||p) is the KL-divergence between distributions p and q; Lk The length of the vector;

[0084] After removing constants, the sparsity metric for the i-th query is defined as:

[0085]

[0086] In the formula M(q) i K) is the sparsity measure of the i-th query;

[0087] Using KL-divergence values, probabilistic sparse self-attention is achieved so that each key only focuses on u queries, represented as:

[0088]

[0089] In the formula It is a sparse matrix of the same size as q and contains only the most important u queries under the sparse metric M(q,K);

[0090] Therefore, the approximate metric for query sparsity is:

[0091]

[0092] In the formula To query the approximate calculation results of the sparse metric;

[0093] As a natural consequence of the probabilistic self-attention mechanism, the feature map obtained by the encoder has redundant combinations of values. A distillation operation is used to preserve the core features. The process of advancing the distillation operation from layer j to layer j+1 is represented as follows:

[0094]

[0095] In the formula The feature map is the (j+1)th layer; MaxPool() is the max pooling downsampling function; ELU() is the activation function; Conv1d() is the one-dimensional convolution function; Let [equation] be the feature map of the j-th layer; AB These are the basic operations of probabilistic self-attention;

[0096] To enhance the robustness of the distillation operation, the encoder architecture establishes a stack of several encoders, each of which is an independent sub-encoder. As the stack increases, one layer is discarded at a time, thereby reducing the number of distillation operation layers and ultimately aligning the output dimensions. Finally, the outputs of all stacks are concatenated to obtain the final representation of the encoder.

[0097] The decoder specifically includes the following:

[0098] The decoder includes a probabilistic sparse self-attention module and a multi-head self-attention module;

[0099] The decoder produces a long sequence of outputs through a forward process:

[0100]

[0101] In the formula The data is processed and then input into the Decoder; Concat() concatenates the two vectors. This is the Start token; Placeholder for the target sequence to be predicted; L token L is the length of the sampled start token sequence. y d is the length of the decoder's predicted sequence. model For data feature dimensions;

[0102] By selecting a mask size, autoregression is avoided by preventing each element in the X sequence from focusing on its future position.

[0103] The training described in step S6 yields an early warning model for the performance degradation of energy storage batteries, specifically including the following steps:

[0104] During training, the LogCosh loss function was used to train the model.

[0105] This invention also provides a system for implementing an early warning method for the performance degradation of the energy storage battery, comprising a data acquisition module, a data processing module, a wavelet analysis module, a principal component analysis module, a model building module, a model training module, and an early warning module; the data acquisition module, data processing module, wavelet analysis module, principal component analysis module, model building module, model training module, and early warning module are connected in series; the data acquisition module is used to acquire the capacity degradation dataset of the energy storage battery and upload the data to the data processing module; the data processing module is used to preprocess the received data to obtain a training dataset and upload the data to the wavelet analysis module; the wavelet analysis module is used to perform wavelet analysis on the training dataset based on the received data to obtain multi-dimensional data at different frequencies. The intrinsic components are analyzed and the data is uploaded to the principal component analysis module. The principal component analysis module performs principal component analysis on the received intrinsic component data to obtain the core frequency signal, and then uploads this data to the model building module. The model building module constructs an initial early warning model for energy storage battery performance degradation based on the Informer network, and uploads this data to the model training module. The model training module trains the initial model using the core frequency signal to obtain an early warning model for energy storage battery performance degradation, and then uploads this model to the early warning module. The early warning module uses the obtained early warning model to provide early warnings of actual energy storage battery performance degradation based on the received data.

[0106] The early warning method and system for energy storage battery performance degradation provided by this invention uses wavelet analysis algorithm to obtain intrinsic components at different frequencies from the original dataset, and uses principal component analysis to refine the obtained intrinsic components to obtain a recognizable core frequency signal. Furthermore, it employs an Informer network based on an attention mechanism to construct an early warning model. Therefore, this invention can not only achieve early warning of energy storage battery performance degradation, but also has higher reliability, better accuracy, and better versatility. Attached Figure Description

[0107] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0108] Figure 2 This is a schematic diagram of the functional modules of the system of the present invention. Detailed Implementation

[0109] like Figure 1 The diagram shown is a flowchart of the method of the present invention: The early warning method for the performance degradation of energy storage batteries disclosed in this invention includes the following steps:

[0110] S1. Obtain the capacity degradation dataset of the energy storage battery; specifically including the following steps:

[0111] The capacity degradation dataset of energy storage batteries is obtained using the BMS device in the following manner;

[0112] Obtain a set of energy storage batteries;

[0113] The energy storage battery is fully charged in a set ambient temperature environment using the prescribed method. After the energy storage battery is placed in the test environment for a set time, it is discharged to the cutoff voltage at a 1C rate. At the same time, the capacity data of the energy storage battery during the discharge process is recorded.

[0114] Repeat the above process several times, record the capacity data each time, and take the average value to obtain the final energy storage battery capacity degradation dataset.

[0115] S2. Perform data preprocessing on the data obtained in step S1 to obtain the training dataset;

[0116] S3. Perform wavelet analysis on the training dataset obtained in step S2 to obtain multi-dimensional eigencomponents at different frequencies; specifically, this includes the following steps:

[0117] A. Select wavelet basis functions suitable for the capacity degradation dataset of energy storage batteries; the wavelet basis functions are expressed as:

[0118]

[0119] In the formula ψ a,b ψ(t) is the wavelet basis function; ψ(t) is the wavelet function before transformation; t is the time-domain variable of the function; a is the scale of wavelet analysis, used to control the scaling of the wavelet function, which corresponds to frequency (inverse ratio); b is the translation amount of wavelet analysis, used to control the translation of the wavelet function, which corresponds to time.

[0120] B. Select the wavelet basis function ψ a,b The inner product of f(t) and the capacity degradation data f(t) of the energy storage battery to be analyzed is calculated, and the coefficient C is used to represent the relationship between f(t) and ψ. a,b The degree of similarity between (t);

[0121] C. Shifting the wavelet basis function k units to the right yields the second wavelet basis function ψ. a,b (tk), and repeat step B until f(t) ends;

[0122] D. Extending the wavelet basis functions yields the third wavelet basis function. And repeat steps B and C;

[0123] E. Repeat step D to expand the wavelet basis functions until the set conditions are met;

[0124] Finally, the general formula for continuous wavelet transform is obtained as follows:

[0125]

[0126] In the formula, CWT(a,b) is the total function of the continuous wavelet transform;

[0127] Based on the general formula of continuous wavelet transform, the multidimensional eigencomponents at different frequencies are obtained;

[0128] S4. Perform principal component analysis on the intrinsic component data obtained in step S3 to obtain the core frequency signal; specifically including the following steps:

[0129] Obtain the intrinsic component data information obtained in step S3, which includes n frequency signals of wavelet components. Each frequency signal sequence includes m index variables, where the j-th index of the i-th frequency signal is represented by x. ij ;

[0130] The indicators are standardized using the following formula:

[0131]

[0132]

[0133]

[0134] In the formula This is the j-th index of the standardized i-th frequency signal; s is the sample mean of the j-th indicator; j Let j be the standard deviation of the j-th indicator;

[0135] The correlation coefficient matrix R is calculated using the following formula:

[0136]

[0137] In the formula r ij Let be the correlation coefficient between the i-th indicator and the j-th indicator;

[0138] Calculate the m eigenvalues ​​λ1 to λ2 of the correlation coefficient matrix R. m and the corresponding eigenvectors u1~u m ; where the j-th eigenvector u j Represented as u j =(u 1j ,u 2j ,...,u nj ) T, λ1≥λ2≥...≥λ m ;

[0139] Based on the feature vectors u1~u m The following formula is used to calculate m index variables:

[0140]

[0141] In the formula y m It is the m-th principal component; Let n be the standardized indicator variable;

[0142] Calculate the information contribution rate and cumulative contribution rate of m eigenvalues:

[0143]

[0144]

[0145] In the formula b j For the j-th principal component y j Information contribution rate; α j For the j-th principal component y j The cumulative contribution rate;

[0146] Based on the cumulative contribution rate of the m eigenvalues, the following judgment is made:

[0147] If |α p -1|≤ε T Then, the principal components corresponding to the first p eigenvalues ​​are selected as the core frequency signal; ε T The set threshold;

[0148] S5. Based on the Informer network, construct an initial model for early warning of energy storage battery performance degradation; specifically including the following steps:

[0149] Based on the Informer network, an initial model for early warning of energy storage battery performance degradation is constructed.

[0150] The model includes an encoder, a decoder, and a fully connected layer;

[0151] The encoder uses a probabilistic self-attention mechanism to encode the input sequence and then inputs the encoded data into the decoder. Using a probabilistic self-attention mechanism reduces the computational cost and enables the processing of extremely long input sequences.

[0152] The decoder uses a one-step generative prediction method to decode the input encoded data and then uploads the decoded data to the fully connected layer. The one-step generative prediction method is faster and more effective than single-step inference.

[0153] Fully connected layers are used to obtain the final output result based on the input decoded data;

[0154] In specific implementation, the input sequence includes the following:

[0155] The core frequency signal obtained in step S4 is used as the input multivariate sequence and expressed as:

[0156]

[0157] In the formula X t For input multivariate sequences; L x The length of the current input sequence; Let be the i-th point in the input sequence at time t; Let the vector dimension be the number of points in the sequence.

[0158] The model employs a dot product self-attention mechanism and uses timestamps to encode location information. The timestamp calculation includes local timestamps (PE) and global timestamps (SE).

[0159] The local timestamp (PE) is calculated using the following formula:

[0160]

[0161] In the formula, PE(pos,2j) is the local timestamp; pos is the location information; [] represents the rounding operation; d model The feature dimensions after input;

[0162] After encoding the timestamps and aligning them with the input dimensions, we obtain the model's input representation vector:

[0163]

[0164] In the formula The final input data is the encoder; α is a factor balancing the size between the scalar mapping and the local / global embedding. Let i be a characteristic scalar, i∈{1,...,L} x}; Local timestamp; Global timestamp; This indicates that there are p types of global timestamps.

[0165] The encoder specifically includes the following:

[0166] Self-attention is defined when receiving a tuple input (Q,K,V) and performing a scaled dot product; where Q is the query value, K is the key value, and V is the value value.

[0167] (Q,K,V) is obtained through a linear transformation of the input matrix X, and is expressed as:

[0168] Q = XW Q

[0169] K = XW K

[0170] V = XW V

[0171] In the formula W Q W is the query parameter matrix to be trained. K W is the key parameter matrix to be trained; V The parameter matrix is ​​the value to be trained;

[0172] The self-attention mechanism is obtained:

[0173]

[0174] In the formula, Attention(Q,K,V) is the calculation formula for the attention mechanism; Softmax() is the normalized exponential function; d is the dimension of the input vector;

[0175] Use q i Represents the i-th row of Q, k i Represents the i-th row of K, v i Let V represent the i-th row. The attention for obtaining the i-th query is defined as the probabilistic form of kernel smoothing:

[0176]

[0177] In the formula A(q) i K,V) represents the attention of the i-th query; k(q) i ,k j ) is an asymmetric exponential kernel function, and v j The j-th row of V; This is the kernel-smoothed probability form for attention;

[0178] According to the above formula, the attention of the i-th query on all keys is defined as a probability distribution p(k j |q i ), Here we introduce a uniform distribution

[0179] To distinguish important queries, the KL divergence is used to measure the similarity between distributions p and q:

[0180]

[0181] In the formula, KL(q||p) is the KL-divergence between distributions p and q; L k The length of the vector;

[0182] After removing constants, the sparsity metric for the i-th query is defined as:

[0183]

[0184] In the formula M(q) i K) is the sparsity measure of the i-th query;

[0185] Using KL-divergence values, probabilistic sparse self-attention is achieved so that each key only focuses on u queries, represented as:

[0186]

[0187] In the formula It is a sparse matrix of the same size as q and contains only the most important u queries under the sparse metric M(q,K);

[0188] Therefore, the approximate metric for query sparsity is:

[0189]

[0190] In the formula To query the approximate calculation results of the sparse metric;

[0191] The above calculation process only involves taking the maximum value and summing the results, i.e., the dot product. Compared with the traditional self-attention mechanism, the computational cost has been greatly reduced, and the model performs better in actual engineering applications.

[0192] As a natural consequence of the probabilistic self-attention mechanism, the feature map obtained by the encoder has redundant combinations of values. Distillation is used to preserve the core features, generating a focused self-attention feature map in the next layer, significantly reducing the feature dimension of the input. The process of distillation from layer j to layer j+1 is represented as follows:

[0193]

[0194] In the formula The feature map is the (j+1)th layer; MaxPool() is the max pooling downsampling function; ELU() is the activation function; Conv1d() is the one-dimensional convolution function; Let [equation] be the feature map of the j-th layer; AB These are the basic operations of probabilistic self-attention;

[0195] To enhance the robustness of the distillation operation, the encoder architecture establishes a stack of several encoders, each of which is an independent sub-encoder. As the stack increases, one layer is discarded at a time, thereby reducing the number of distillation operation layers and ultimately aligning the output dimensions. Finally, the outputs of all stacks are concatenated to obtain the final representation of the encoder.

[0196] The decoder specifically includes the following:

[0197] The decoder includes a probabilistic sparse self-attention module and a multi-head self-attention module;

[0198] The decoder produces a long sequence of outputs through a forward process:

[0199]

[0200] In the formula The data is processed and then input into the Decoder; Concat() concatenates the two vectors. This is the Start token; Placeholder for the target sequence to be predicted; L token L is the length of the sampled start token sequence. y d is the length of the decoder's predicted sequence. model For data feature dimensions;

[0201] By selecting the mask size, autoregression is avoided by preventing each element in the X sequence from focusing on its future position.

[0202] S6. Using the core frequency signal obtained in step S4, train the initial model constructed in step S5 to obtain an early warning model for the performance degradation of the energy storage battery; the training to obtain the early warning model for the performance degradation of the energy storage battery specifically includes the following steps:

[0203] During training, the LogCosh loss function is used to train the model;

[0204] S7. Using the early warning model for energy storage battery performance degradation obtained in step S6, perform early warning for actual energy storage battery performance degradation.

[0205] To address the issue of suboptimal preprocessing of raw data in existing methods, this invention employs wavelet analysis to obtain intrinsic components at different frequencies from the original dataset. The advantages of this method are: ① Multi-scale analysis: Wavelet analysis possesses multi-scale analysis capabilities, simultaneously capturing features of signals at different frequencies and time scales. This makes it highly useful for analyzing non-stationary signals and signals containing multiple frequency components; ② Compressed representation: Wavelet transform typically generates sparse representations, meaning it can provide effective signal compression, reducing data storage and transmission costs; ③ Fast algorithm: Compared to some other signal analysis methods, wavelet analysis has a faster computation speed, especially when using Discrete Wavelet Transform (DWT). This makes it suitable for real-time processing and large-scale datasets; ④ Noise filtering: Wavelet analysis can be used to remove noise from signals because noise typically has different frequency characteristics, which can be separated and filtered in the wavelet domain; ⑤ Adaptability: Wavelet analysis can be adaptively adjusted according to application requirements, including selecting different wavelet basis functions and scales. This makes it highly flexible and suitable for various types of signal processing tasks.

[0206] To address the problem in existing solutions where excessive feature information after initial data preprocessing leads to slow convergence in subsequent network model training, this invention applies Principal Component Analysis (PCA) to refine the intrinsic components obtained in the previous step, yielding recognizable core frequency signals. The advantages are: ① Dimensionality reduction: PCA can map high-dimensional data to a low-dimensional space, preserving most of the variance information. This helps reduce data dimensionality, decrease dataset complexity, and make the data easier to visualize and analyze; ② Improved model performance: In some cases, applying PCA can improve the performance of machine learning models because dimensionality reduction removes noise and reduces the risk of overfitting; ③ Reduced computational complexity: Through dimensionality reduction, PCA can reduce the computational complexity of data analysis and modeling, saving time and resources; ④ Removal of redundant information: PCA removes redundant information in the data by identifying the main directions of change (principal components). This helps reduce noise and improve data clarity.

[0207] To address the issues of excessive complexity, slow training convergence, high resource consumption, and short prediction strides in existing neural network models, this invention applies the Informer network, which consists of an Encoder and Decoder based on an attention mechanism. Its advantages are as follows: ① Long sequence processing: The Informer model is designed to process long-term sequence data. It introduces a self-attention mechanism, which can better capture long-term dependencies in the sequence, thus performing better on long sequences than some traditional models; ② Global information capture: Informer utilizes a global self-attention mechanism, enabling it to focus on all parts of the sequence simultaneously, rather than just a local window. This helps to better capture global patterns and trends in the sequence; ③ End-to-end learning: Informer has end-to-end learning capabilities, which can optimize model parameters through backpropagation without the need for manually designing complex feature engineering and preprocessing processes; ④ Multi-timescale: Informer supports prediction at multiple time scales, from hours to months, making it suitable for prediction tasks across different time ranges; ⑤ Model interpretability: Informer's self-attention mechanism allows for the analysis of the model's attention distribution in the time series, thus providing a degree of model interpretability and helping to understand the basis of the prediction; ⑥ Performance superiority: Informer performs exceptionally well in some time series prediction tasks, especially in long-series and multi-scale prediction tasks. It often outperforms some traditional time series prediction methods.

[0208] like Figure 2The diagram shows the functional modules of the system of this invention: The system for implementing an early warning method for the performance degradation of the energy storage battery disclosed in this invention includes a data acquisition module, a data processing module, a wavelet analysis module, a principal component analysis module, a model building module, a model training module, and an early warning module; these modules are connected in series. The data acquisition module acquires the capacity degradation dataset of the energy storage battery and uploads the data to the data processing module. The data processing module preprocesses the received data to obtain a training dataset and uploads the data to the wavelet analysis module. The wavelet analysis module performs wavelet analysis on the training dataset based on the received data to obtain a training dataset. The system receives multi-dimensional intrinsic components at the same frequency and uploads the data to the principal component analysis module. The principal component analysis module performs principal component analysis on the intrinsic component data based on the received data to obtain the core frequency signal, and uploads the data to the model building module. The model building module constructs an initial early warning model for energy storage battery performance degradation based on the received data and an Informer network, and uploads the data to the model training module. The model training module trains the initial model using the core frequency signal based on the received data to obtain an early warning model for energy storage battery performance degradation, and uploads the data to the early warning module. The early warning module uses the obtained early warning model for energy storage battery performance degradation to provide early warnings of actual energy storage battery performance degradation based on the received data.

Claims

1. An early warning method for performance degradation of energy storage batteries, comprising the following steps: S1. Obtain the capacity degradation dataset of energy storage batteries; S2. Perform data preprocessing on the data obtained in step S1 to obtain the training dataset; S3. Perform wavelet analysis on the training dataset obtained in step S2 to obtain multi-dimensional eigencomponents at different frequencies; S4. Perform principal component analysis on the intrinsic component data information obtained in step S3 to obtain the core frequency signal; S5. Based on the Informer network, construct an initial model for early warning of energy storage battery performance degradation; specifically including the following steps: Based on the Informer network, an initial model for early warning of energy storage battery performance degradation is constructed. The model includes an encoder, a decoder, and a fully connected layer; The encoder uses a probabilistic self-attention mechanism to encode the input sequence and then inputs the encoded data into the decoder. The decoder uses a one-time generative prediction method to decode the input encoded data and uploads the decoded data to the fully connected layer; Fully connected layers are used to obtain the final output result based on the input decoded data; S6. Using the core frequency signal obtained in step S4, train the initial model constructed in step S5 to obtain an early warning model for the performance degradation of the energy storage battery. S7. Using the early warning model for energy storage battery performance degradation obtained in step S6, perform early warning for actual energy storage battery performance degradation.

2. The early warning method for energy storage battery performance degradation according to claim 1, characterized in that... Step S1, which involves obtaining the capacity degradation dataset of the energy storage battery, specifically includes the following steps: The capacity degradation dataset of energy storage batteries is obtained using the following method; Obtain a set of energy storage batteries; The energy storage battery is fully charged in a set ambient temperature environment using the prescribed method. After the energy storage battery is placed in the test environment for a set time, it is discharged to the cutoff voltage at a 1C rate. At the same time, the capacity data of the energy storage battery during the discharge process is recorded. Repeat the above process several times, record the capacity data each time, and take the average value to obtain the final energy storage battery capacity degradation dataset.

3. The early warning method for energy storage battery performance degradation according to claim 2, characterized in that... Step S3, which involves performing wavelet analysis on the training dataset obtained in step S2 to obtain multi-dimensional eigencomponents at different frequencies, specifically includes the following steps: A. Select wavelet basis functions suitable for the capacity degradation dataset of energy storage batteries; the wavelet basis functions are expressed as: In the formula These are wavelet basis functions; is the wavelet function before transformation; t is the time-domain variable of the function; The scale for wavelet analysis; This represents the shift amount in wavelet analysis; B. Select wavelet basis functions With the capacity degradation data of the energy storage battery to be analyzed Perform the inner product and calculate the coefficient C; the coefficient C is used to represent and The degree of similarity; C. Shift the wavelet basis function k units to the right to obtain the second wavelet basis function. And repeat step B until Finish; D. Extending the wavelet basis functions yields the third wavelet basis function. And repeat steps B and C; E. Repeat step D to expand the wavelet basis functions until the set conditions are met; Finally, the general formula for continuous wavelet transform is obtained as follows: In the formula It is the total function of the continuous wavelet transform; Based on the general formula of continuous wavelet transform, the multidimensional eigencomponents at different frequencies are obtained.

4. The early warning method for energy storage battery performance degradation according to claim 3, characterized in that... Step S4, which involves performing principal component analysis on the intrinsic component data obtained in step S3 to obtain the core frequency signal, specifically includes the following steps: Obtain the intrinsic component data information obtained in step S3, where there are n frequency signals of wavelet components, and each frequency signal sequence includes m index variables, where the j-th index of the i-th frequency signal is represented as... ; The indicators are standardized using the following formula: In the formula This is the j-th index of the standardized i-th frequency signal; Let be the sample mean of the j-th indicator; Let j be the standard deviation of the j-th indicator; The correlation coefficient matrix R is calculated using the following formula: In the formula Let be the correlation coefficient between the i-th indicator and the j-th indicator; Calculate the m eigenvalues ​​of the correlation coefficient matrix R. and the corresponding feature vector ; where the j-th eigenvector Represented as , ; Based on the feature vector The following formula is used to calculate m index variables: In the formula It is the m-th principal component; Let n be the standardized indicator variable; Calculate the information contribution rate and cumulative contribution rate of m eigenvalues: In the formula For the j-th principal component Information contribution rate; For the j-th principal component The cumulative contribution rate; Based on the cumulative contribution rate of the m eigenvalues, the following judgment is made: like Then, the principal components corresponding to the first p eigenvalues ​​are selected as the core frequency signal; The threshold value is set.

5. The early warning method for performance degradation of energy storage batteries according to claim 4, characterized in that... The input sequence specifically includes the following: The core frequency signal obtained in step S4 is used as the input multivariate sequence and expressed as: In the formula For input multivariate sequences; The length of the current input sequence; for The first time in the input sequence One point; Let the vector dimension be the number of points in the sequence. The model employs a dot product self-attention mechanism and uses timestamps to encode location information. The timestamp calculation includes local timestamps (PE) and global timestamps (SE). The local timestamp (PE) is calculated using the following formula: In the formula Local timestamp; Location information; , This is a rounding operation; The feature dimensions after input; After encoding the timestamps and aligning them with the input dimensions, we obtain the model's input representation vector: In the formula This is the data that is ultimately input into the encoder; A factor to balance the size between scalar mapping and local / global embedding; For characteristic scalars, ; Local timestamp; Global timestamp; This indicates that there are p types of global timestamps.

6. The early warning method for performance degradation of energy storage batteries according to claim 5, characterized in that... The encoder specifically includes the following: Self-attention is the process of receiving tuple input. Defined at time, and performs the scaled dot product; where Q is the query value, K is the key value, and V is the value value; It is obtained by linear transformation of the input matrix X, and is expressed as: In the formula The query parameter matrix to be trained; The key parameter matrix is ​​to be trained; The parameter matrix is ​​the value to be trained; The self-attention mechanism is obtained: In the formula The formula for calculating the attention mechanism; It is a normalized exponential function; The dimension of the input vector; use Represents the i-th row of Q, This represents the i-th row of K. Let V represent the i-th row. The attention for obtaining the i-th query is defined as the probabilistic form of kernel smoothing: In the formula Attention to the i-th query; It is an asymmetric exponential kernel function, and ; The j-th row of V; This is the kernel-smoothed probability form for attention; According to the above formula, the attention of the i-th query on all keys is defined as a probability distribution. , ; The KL divergence is used to measure the similarity between distributions p and q: In the formula Let p be the KL divergence between distributions p and q. The length of the vector; After removing constants, the sparsity metric for the i-th query is defined as: In the formula Let i be the sparsity measure of the i-th query; Using KL-divergence values, probabilistic sparse self-attention is achieved so that each key only focuses on u queries, represented as: In the formula A sparse matrix of the same size as q, containing only sparse metrics. The most important query below; Therefore, the approximate metric for query sparsity is: In the formula To query the calculation results of the sparse metric; As a natural consequence of the probabilistic self-attention mechanism, the feature map obtained by the encoder has redundant combinations of values. A distillation operation is used to preserve the core features. The process of advancing the distillation operation from layer j to layer j+1 is represented as follows: In the formula This is the feature map of the (j+1)th layer; This is the max-pooling downsampling function; For activation functions; It is a one-dimensional convolution function; This is the feature map of the j-th layer; These are the basic operations of probabilistic self-attention; To enhance the robustness of the distillation operation, the encoder architecture establishes a stack of several encoders, each of which is an independent sub-encoder. As the stack increases, one layer is discarded at a time, thereby reducing the number of distillation operation layers and ultimately aligning the output dimensions. Finally, the outputs of all stacks are concatenated to obtain the final representation of the encoder.

7. The early warning method for performance degradation of energy storage batteries according to claim 6, characterized in that... The decoder specifically includes the following: The decoder includes a probabilistic sparse self-attention module and a multi-head self-attention module; The decoder produces a long sequence of outputs through a forward process: In the formula This refers to the processed data input into the Decoder. Concatenate the two sets of vectors; This is the Start token; Placeholders for the target sequence to be predicted; The length of the sampled start token sequence; The length of the predicted sequence for the decoder; For data feature dimensions; By selecting a mask size, autoregression is avoided by preventing each element in the X sequence from focusing on its future position.

8. The early warning method for performance degradation of energy storage batteries according to claim 7, characterized in that... The training described in step S6 yields an early warning model for the performance degradation of energy storage batteries, specifically including the following steps: During training, the LogCosh loss function was used to train the model.

9. A system for implementing the early warning method for performance degradation of energy storage batteries according to any one of claims 1 to 8, characterized in that... It includes a data acquisition module, a data processing module, a wavelet analysis module, a principal component analysis module, a model building module, a model training module, and an early warning module; the data acquisition module, data processing module, wavelet analysis module, principal component analysis module, model building module, model training module, and early warning module are connected in series; the data acquisition module is used to acquire the capacity degradation dataset of energy storage batteries and upload the data to the data processing module; The data processing module is used to preprocess the received data to obtain the training dataset and then upload the data to the wavelet analysis module. The wavelet analysis module is used to perform wavelet analysis on the training dataset based on the received data, obtain multi-dimensional eigencomponents at different frequencies, and upload the data to the principal component analysis module. The principal component analysis module is used to perform principal component analysis on the intrinsic component data information based on the received data, obtain the core frequency signal, and upload the data to the model building module; the model building module is used to build an early warning initial model for the performance degradation of energy storage batteries based on the Informer network based on the received data, and upload the data to the model training module. The model training module is used to train the initial model based on the received data and the core frequency signal to obtain an early warning model for the performance degradation of the energy storage battery, and then upload the data to the early warning module. The early warning module is used to provide early warnings of actual energy storage battery performance degradation based on the received data and the obtained early warning model for energy storage battery performance degradation.

Images