Method and system for predicting service life degradation of energy storage lithium battery

Abstract

The invention discloses an energy storage lithium battery life degradation prediction method and system, and the method comprises the steps: obtaining and preprocessing lithium battery operation data, constructing a time sequence, carrying out the normalization processing, and dividing a training set and a test set; based on the normalized data, constructing a multi-head self-attention mechanism MSA module with masks, capturing multiple features in the input data through different time windows, and extracting an important feature set through a feature fusion method after screening; based on the important feature set, key hyper-parameters are preliminarily set, and an MSA-LSTM comprehensive prediction model is constructed and used for processing time sequence data and predicting the service life of the battery; an artificial bee colony algorithm is adopted to optimize hyper-parameters of the MSA-LSTM comprehensive prediction model; and training the optimized MSA-LSTM comprehensive prediction model by using the training set, verifying and evaluating the prediction performance by using the test set, predicting unknown data by using the MSA-LSTM comprehensive prediction model, and calculating the remaining service life RUL of the energy storage lithium battery. And safety accidents caused by battery faults are prevented.

Description

Technical Field

[0001] The present invention belongs to the technical field of remaining service life prediction of lithium-ion batteries, and more specifically, relates to a method and system for predicting service life degradation of energy storage lithium batteries. Background Art

[0002] Lithium-ion batteries are increasingly used in distributed energy storage due to their high energy density, low self-discharge rate, large capacity, and strong adaptability. However, as the use time increases, the internal chemical reaction of the battery slows down and the electrode material degrades, resulting in a decrease in battery capacity and an increase in internal resistance. If the battery management system (BMS) cannot effectively monitor and adjust the battery status, or the battery is subjected to adverse conditions such as overcharge, overdischarge, high temperature, or physical damage, it may cause battery failure or serious safety accidents such as overheating, combustion, or even explosion. Therefore, it is necessary to accurately predict the remaining useful life (RUL) of lithium-ion batteries and obtain information related to the battery life in advance so that batteries that have reached the threshold can be replaced in time to ensure the safe and reliable operation of the system.

[0003] At present, there are three main methods for RUL prediction of lithium-ion batteries: model-based methods, data-driven methods, and fusion algorithm-based methods. With the rapid development of cloud computing and artificial intelligence, data-driven methods are more powerful in dealing with complex nonlinear problems. They do not need to master the electrochemical mechanism inside the battery, and can avoid the influence of insufficient experience and knowledge on the accuracy of estimation and prediction results, and gradually become the mainstream of RUL prediction. However, the current data-driven RUL prediction methods still have some shortcomings: RUL prediction models with good performance require multiple training of neural networks to seek the optimal parameters of the model; the workload of manually adjusting the hyperparameters of the neural network is large, and the prediction model uses the gradient descent algorithm to update the parameters, which is very easy to fall into the local minimum, causing the model to overfit; the simple model architecture is difficult to capture the nonlinearity and uncertainty capacity recovery problems in the process of battery capacity degradation; the mining ability of battery capacity decay data needs to be improved, and it is impossible to effectively capture the long-term and short-term time dependencies in the battery performance data, making it impossible to guarantee the high accuracy and strong robustness of the RUL prediction results of lithium-ion batteries. Summary of the invention

[0004] In order to solve the problem of low accuracy and weak robustness in the current prediction of the remaining service life of lithium-ion batteries, the present invention proposes a method for predicting the remaining service life of lithium-ion batteries based on the ABC-MSA-LSTM model. First, a multi-head self-attention mechanism with mask is added to enhance the expressiveness of sequence features. Secondly, the dropout regularization method is used to effectively reduce the advantage of overfitting and improve the generalization ability of the prediction model. Then, an activation layer network structure is introduced to address capacity recovery and data noise issues, significantly improving the model's ability to handle complex nonlinear data. Finally, the hyperparameters of the MSA-LSTM comprehensive prediction model are optimized in combination with the ABC algorithm to avoid the model falling into a local optimal solution, thereby improving the RUL prediction accuracy and robustness.

[0005] The present invention adopts the following technical solution.

[0006] The first aspect of the present invention provides a method for predicting the life degradation of a lithium energy storage battery, which specifically comprises the following steps:

[0007] Step 1: Obtain and preprocess lithium battery operation data, construct time series and perform normalization, and roll-divide training and test sets;

[0008] Step 2: Based on the normalized data, a masked multi-head self-attention mechanism MSA module is constructed to capture multiple features in the input data through different time windows, and dynamic weights are assigned to different attention heads to extract important features. After screening, the important feature set is extracted through feature fusion method;

[0009] Step 3: Based on the important feature set, preliminarily set the key hyperparameters, build multiple multi-head self-attention-long short-term memory network MSA-LSTM comprehensive prediction models with dropout and activation function layers, and dynamically adjust the hyperparameters;

[0010] Step 4, optimize the hyperparameters of the multi-head self-attention-long short-term memory network MSA-LSTM comprehensive prediction model through the artificial bee colony algorithm to find the optimal solution set;

[0011] Step 5: Use the training set divided in step 1 to train the optimized MSA-LSTM comprehensive prediction model, use the test set to perform multiple rounds of cross-validation, evaluate the prediction performance, apply the MSA-LSTM comprehensive prediction model to predict unknown data, and calculate the remaining service life RUL of the energy storage lithium battery.

[0012] Preferably, step 1 comprises:

[0013] Step 1.1, obtain multi-dimensional data of lithium battery capacity, voltage, temperature, discharge rate and external environment, and perform preprocessing, including removing missing values ​​and outliers;

[0014] Step 1.2, use the isolation forest algorithm to detect and remove abnormal data, and use linear interpolation to repair missing values;

[0015] Step 1.3, construct the time series data of sliding windows and extended windows according to different time granularities;

[0016] Step 1.4, according to the constructed time series features, select the optimal normalization method, including min-max normalization and z-score normalization;

[0017] In step 1.5, the generative adversarial network (GAN) is used to generate expanded training data, and the training set and test set are dynamically divided through the rolling partition strategy.

[0018] Preferably, in step 1.4, the min-max normalization is calculated as follows:

[0019]

[0020] Where:

[0021] x i Represents the preprocessed data; y i represents the normalized data;

[0022] max(x) and min(x) represent the corresponding maximum and minimum values ​​in the original capacity data respectively;

[0023] The z-score standardization is calculated as follows:

[0024]

[0025] Where:

[0026] x i represents the preprocessed data; μ represents the average value of the preprocessed data features; σ represents the standard deviation of the preprocessed data features, which is used to measure the volatility of the data.

[0027] Preferably, step 2 comprises:

[0028] Step 2.1, input the normalized data into the MSA module, and capture the important features of the time series by combining the sliding window and the jumping window of the multi-head self-attention mechanism;

[0029] Step 2.2, map the input sequence θ into query q, key k, and value v vectors through three weight matrices, as follows:

[0030]

[0031] Where:

[0032] θ i represents the input feature vector at time step i;

[0033] W q represents the query matrix, W k represents the bond matrix, W v represents the value matrix, W q , W k ,

[0034] (W q ) h represents the query weight matrix of the h-th attention head, (W k ) h represents the key weight matrix of the h-th attention head, (W v ) h represents the value weight matrix of the h-th attention head, h∈{1,…,P};

[0035] represents the query vector generated by the h-th attention head at time step i; represents the key vector generated by the h-th attention head at time step i; represents the value vector generated by the h-th attention head at time step i;

[0036] Step 2.3, by scaling the query vector with the key vector, the attention weight is calculated, and the mask mechanism is applied to make the current time step focus only on the past time steps. The formula is as follows:

[0037]

[0038] Where:

[0039] represents the attention weight of the h-th attention head at time step i to time step τ;

[0040] Represents the scaling factor, which is used to control the dot product result;

[0041] mask i,τ Represents the mask value, which is used to process the information of time step τ and to block the information of future time steps; if τ>i, the mask is negative infinity, otherwise it is 0;

[0042] Represents the attention weight Apply the softmax function, and the resulting weights are between 0 and 1, representing the attention distribution at each time step, which is used to convert the weights into probabilities;

[0043] represents the attention output after weighted summation; λ i-τ represents the time attenuation coefficient;

[0044] Step 2.4, by defining multiple sliding windows, sliding on the time series with different window lengths, extracting different time segment data, representing the performance of the time series at different time scales, and aggregating the features of different time windows into the feature set;

[0045] Step 2.5, evaluate the importance of features through decision trees, check the correlation between features and the remaining useful life RUL of lithium-ion batteries, fuse features through principal component analysis PCA method, and construct an important feature set;

[0046] Step 2.6, output the masked MSA result, the formula is as follows:

[0047]

[0048] Where:

[0049] It means concatenating the output vectors of all P attention heads at time step i;

[0050] W o Represents the weight matrix of linear projection, with a size of (P·d k )×d em , d k is the dimension of each head, d em is the dimension of the final output;

[0051] linear{·} represents a linear transformation operation.

[0052] Preferably, in step 3, constructing the MSA-LSTM comprehensive prediction model specifically includes the following steps:

[0053] Step 3.1: Take the extracted important features as the current input x t The forget gate is introduced to retain important information after filtering through the Sigmoid function. The formula is as follows:

[0054] f t =σ(w f ·[h t-1 ,x t ]+b f )

[0055] Where:

[0056] x t 、f t Indicates the gate input capacity and output capacity;

[0057] w frepresents the connection weight parameter;

[0058] h t-1 Indicates the output status at the previous moment;

[0059] b f is the offset parameter;

[0060] σ means the output range of the Sigmoid function is [0,1];

[0061] Step 3.2, the input gate controls the transmission of new information to the cell state, the Sigmoid function is used to determine the value to be updated, and the tanh function generates the candidate state. The formula is as follows:

[0062] i t =σ(w i ·[h t-1 ,x t ]+b i )

[0063]

[0064] Where:

[0065] i t represents the output of the input gate;

[0066] w i represents the weight parameter of the input gate; w s The weight parameter representing the candidate cell state;

[0067] b i represents the offset parameter of the input gate; b s The offset parameter representing the candidate cell state;

[0068] represents the candidate cell state;

[0069] Step 3.3, combine the output of the forget gate and the input gate to update the current cell state. The formula is as follows:

[0070] O t =σ(w o ·[h t-1 , x t ]+b o )

[0071]

[0072] h t =O t ×tanh(C t )

[0073] Where:

[0074] O t represents the output of the output gate;

[0075] w o Represents the weight matrix of the output gate;

[0076] b o Represents the offset parameter of the output gate;

[0077] h t Indicates the hidden state at the current moment;

[0078] C t Indicates the cell state at the current moment; C t-1 Indicates the cell state at the previous moment;

[0079] tanh(C t ) represents the hyperbolic tangent function, which is used to convert the cell state C t Map to the interval [-1,1];

[0080] Step 3.4, taking into account the nonlinearity and noise in the battery capacity data, the Dropout layer is used in the LSTM deep neural network to train the generalization ability of the model; the formula of Dropout in the training stage is as follows:

[0081] y=F(Wx)*m

[0082] Where:

[0083] y represents the output of the layer;

[0084] F represents the activation function of the layer;

[0085] x represents the input of the layer;

[0086] m represents the Dropout rate;

[0087] W represents the weight matrix;

[0088] In step 3.5, add a LeakyReLU activation function layer after the LSTM layer to build the MSA-LSTM comprehensive prediction model, form a complete model structure, output the battery capacity prediction value, solve the problem that the standard ReLU has a zero gradient in the negative input area, and optimize the nonlinear expression ability of the model.

[0089] Preferably, in step 3.5, the LeakyReLU activation function formula is as follows:

[0090]

[0091] Where:

[0092] α = 0.01;

[0093] When the input x>0, the output of LeakyReLU is directly the input value, which is consistent with the behavior of the ReLU function and remains positive;

[0094] When the input x≤0, the output of LeakyReLU is α, which is different from the behavior of the ReLU function. The standard ReLU outputs 0 when the input is negative, while the LeakyReLU outputs a negative value when the input is negative.

[0095] Preferably, step 4 comprises:

[0096] Step 4.1: Initialize the control parameters in the ABC algorithm, including the number of honey sources S N , the maximum number of cycles S for finding nectar sources and the number of termination cycles S mc ;

[0097] Step 4.2, use the root mean square error RMSE to set the fitness function in the ABC algorithm to evaluate the quality of each nectar source. The formula is as follows:

[0098]

[0099] Where:

[0100] represents the predicted value;

[0101] y i represents the true value;

[0102] n represents the sample size;

[0103] Step 4.3, define the initial search range of the hyperparameters, set the initial honey source position of the ABC algorithm, and the honey source includes the hyperparameter combination;

[0104] Step 4.4, based on the determined search range, generate new nectar sources by using hired bee local search to explore the solution space; the formula is as follows:

[0105]

[0106] Where:

[0107] χ′ id Indicates the d-th dimension value of the current nectar source;

[0108] Represents a random number between [-1,1]; i = 1, 2, ..., S N ; d = 1, 2, ..., D; k is any nectar source except i, k ≠ i;

[0109] χ kd represents the d-th dimension value of another nectar source;

[0110] Step 4.5, based on the information transmitted by the employed bees, select the optimal nectar source through the fitness probability; the fitness probability formula is as follows:

[0111]

[0112] Where:

[0113] P i represents the probability of the i-th nectar source being selected;

[0114] fit i represents the fitness value of the i-th nectar source;

[0115] S N Indicates the total number of nectar sources;

[0116] Represents the sum of fitness values ​​of all nectar sources;

[0117] Step 4.6, based on the selected optimal nectar source, train and test the MSA-LSTM comprehensive prediction model, use scout bees to search and replace the nectar source whose fitness has not been improved, and continue to explore new solution space; the formula is as follows:

[0118]

[0119] Where:

[0120] χ id represents the value of the i-th nectar source in the d-th dimension, indicating the position of the nectar source in the multidimensional search space;

[0121] represents the lower bound of the d-th dimension;

[0122] represents the upper bound of the d-th dimension;

[0123] rand(0,1) represents a random number in the interval [0,1], which is used to generate a random position of the nectar source in the dth dimension;

[0124] Represents the range of the dth dimension, which is used to scale the random numbers so that the random positions are within the range of the search space.

[0125] Preferably, in step 4.3, the definition of the initial search range includes setting the number of attention heads P between 1 and 8, the batch size B between 16 and 256, the number of hidden layer units between 32 and 512, and the Dropout rate between 0.1 and 0.5.

[0126] Preferably, step 5 comprises:

[0127] Step 5.1: Based on the optimal solution set found in step 4, use K-fold cross validation to split the data set into different training sets and validation sets multiple times, and train the MSA-LSTM model on the training data set.

[0128] Step 5.2: Use the test data set to verify the trained model and calculate the mean absolute error (MAE), root mean square error (RMSE) and R-squared (R 2 ) performance indicator to evaluate the predictive ability on the test data set; the formula is as follows:

[0129]

[0130] Where:

[0131] represents the predicted value;

[0132] y i represents the true value;

[0133] n represents the total number of samples.

[0134] The second aspect of the present invention provides a system for predicting the life degradation of a lithium energy storage battery, including: a data acquisition and preprocessing module, a feature extraction and selection module, a model training and optimization module, a prediction and diagnosis module, and a result display and feedback module;

[0135] The data acquisition and preprocessing module is used to collect battery operating capacity data and preprocess the data, including cleaning, denoising, and normalization;

[0136] The feature extraction and selection module is used to extract important features from the preprocessed data and perform feature selection to capture key information about battery degradation;

[0137] The model training and optimization module is used to train the battery life prediction model, using machine learning and deep learning algorithms to improve the model's prediction performance through evaluation and optimization;

[0138] The prediction and diagnosis module is used to use the trained model to predict the remaining service life (RUL) of the battery and diagnose the battery health status;

[0139] The result display and feedback module is used to display the prediction results and diagnostic information to users in a visual manner to provide decision support; at the same time, the model is updated and maintained based on actual operating conditions and new data.

[0140] Compared with the prior art, the beneficial effects of the present invention include at least:

[0141] (1) By acquiring and preprocessing lithium battery operation data, constructing time series and normalizing them, and dividing them into training sets and test sets, the quality and consistency of the data are guaranteed, the efficiency and accuracy of model training are improved, and the problem of model overfitting is effectively prevented.

[0142] (2) By constructing a masked multi-head self-attention mechanism (MSA) module, the long-term and short-term time dependencies in the input data can be accurately captured, significantly improving the expressiveness of sequence features, thereby avoiding the problem that a single model architecture cannot capture the nonlinearity and uncertainty in the battery capacity degradation process, and improving the accuracy of feature extraction.

[0143] (3) By constructing an MSA-LSTM comprehensive prediction model, the advantages of MSA and LSTM are combined to better process time series data and capture the complex nonlinear relationship in the process of battery capacity attenuation, thereby improving the prediction performance and robustness of the model.

[0144] (4) By using the artificial bee colony algorithm (ABC) to optimize the hyperparameters of the MSA-LSTM comprehensive prediction model, the problems of heavy workload and low efficiency caused by manually adjusting the hyperparameters are effectively avoided. At the same time, the global search capability of the ABC algorithm is utilized to avoid the problem that the gradient descent algorithm is prone to falling into local minima, thereby improving the generalization ability and prediction accuracy of the model.

[0145] (5) By training and verifying the optimized MSA-LSTM comprehensive prediction model, the optimized model is applied to predict the remaining useful life (RUL) of the battery, providing high-precision prediction results, obtaining battery life information in advance, and ensuring the safe and reliable operation of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0146] Figure 1 It is an overall flow chart of the ABC-MSA-LSTM prediction model provided according to an embodiment of the present invention;

[0147] Figure 2 is a structural diagram of a masked MSA module provided in accordance with an embodiment of the present invention;

[0148] Figure 3 (a) is a structural diagram of an LSTM unit provided according to an embodiment of the present invention;

[0149] Figure 3 (b) is a structural diagram of an LSTM deep neural network model provided according to an embodiment of the present invention;

[0150] Figure 4 It is a structural diagram of the MSA-LSTM comprehensive prediction model provided according to an embodiment of the present invention;

[0151] Figure 5It is a specific optimization flow chart of the ABC-MSA-LSTM method provided in accordance with an embodiment of the present invention;

[0152] Figure 6 is a capacity decay curve diagram provided according to an embodiment of the present invention;

[0153] Figure 7 (a) is a diagram showing the prediction results of the RUL of a B05 battery based on 40% training data according to an embodiment of the present invention;

[0154] Figure 7 (b) is a diagram showing the prediction results of the RUL of a B06 battery based on 40% training data according to an embodiment of the present invention;

[0155] Figure 7 (c) is a diagram showing the prediction results of the RUL of a B07 battery based on 40% training data according to an embodiment of the present invention;

[0156] Figure 7 (d) is a diagram showing the prediction results of the RUL of a B18 battery based on 40% training data according to an embodiment of the present invention;

[0157] Figure 8 (a) is a diagram showing the prediction results of the RUL of a B05 battery based on 60% training data according to an embodiment of the present invention;

[0158] Figure 8 (b) is a diagram showing the prediction results of the RUL of a B06 battery based on 60% training data according to an embodiment of the present invention;

[0159] Figure 8 (c) is a diagram showing the prediction results of the RUL of a B07 battery based on 60% training data according to an embodiment of the present invention;

[0160] Figure 8 (d) is a diagram showing the prediction results of the RUL of a B18 battery based on 60% training data according to an embodiment of the present invention. DETAILED DESCRIPTION

[0161] In order to make the purpose, technical scheme and advantages of the present invention clearer, the technical scheme of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. The embodiments described in this application are only part of the embodiments of the present invention, not all of them. Based on the spirit of the present invention, other embodiments obtained by ordinary technicians in this field without creative work are all within the scope of protection of the present invention.

[0162] like Figure 1 As shown, embodiment 1 of the present invention provides a method for predicting the life degradation of a lithium energy storage battery, comprising the following steps:

[0163] Step 1: Acquire and preprocess lithium battery operation data, construct time series and perform normalization, and divide the data into training set and test set;

[0164] Preferably, step 1 specifically includes the following steps:

[0165] Step 1.1, obtaining the operating data of the lithium battery capacity, temperature, voltage and current;

[0166] Specifically, the capacity data decay curve of the battery aging data set is analyzed. This time, the first batch of lithium-ion battery cycle aging data sets are used, namely B0005 (B05), B006 (B06), B007 (B07) and B0018 (B18). At room temperature of 24°C, the battery is subjected to continuous charge and discharge tests. When the battery reaches the end of life (EOL), the test is stopped and the rated capacity is reduced by 30% (from 2Ah to 1.4Ah). The capacity change curves of batteries B05, B06, B07 and B18 with the number of cycles are shown in the figure below. Figure 6 As shown. The capacity data after data set preprocessing is used as input features for example verification. Among them, EOL is 70% of the battery capacity (1.4Ah), and the B07 battery will not drop to the lower limit, so its EOL is 72% of the rated capacity of the battery (1.44Ah);

[0167] Step 1.2, preprocessing the data, including removing missing values ​​and outliers;

[0168] Step 1.3, reconstructing the preprocessed data into a time series format to reflect the operating status and historical behavior of the battery;

[0169] Step 1.4, normalize the time series data;

[0170] Preferably, in step 1.4, the min-max normalization is calculated as follows:

[0171]

[0172] Where:

[0173] x i Represents the preprocessed data; y i represents the normalized data;

[0174] max(x) and min(x) represent the corresponding maximum and minimum values ​​in the original capacity data respectively;

[0175] Specifically, data such as capacity and voltage usually fluctuate within a fixed range (such as battery capacity between 0% and 100%, and voltage between 3.0V and 4.2V). They are mapped to the 0, 1 range through Min-Max normalization to ensure that the model will not produce deviations due to differences in the magnitude of feature values ​​when processing different features.

[0176] Adapting to input of time series features: Since the present invention adopts the MSA-LSTM model, different features of the time series data need to be in the same numerical range to ensure that the model can efficiently learn these features.

[0177] The z-score standardization is calculated as follows:

[0178]

[0179] Where:

[0180] x i represents the preprocessed data; μ represents the average value of the preprocessed data features; σ represents the standard deviation of the preprocessed data features, which is used to measure the volatility of the data.

[0181] Specifically, data such as temperature and discharge rate may fluctuate greatly over time and working conditions. Using Z-Score normalization to convert these data into a distribution with a mean of 0 and a standard deviation of 1 can reduce the impact of fluctuations on model training.

[0182] Handling extreme values: Extreme temperatures (such as high temperature) or abnormal currents may occur in certain battery usage conditions. Z-Score normalization can reduce the interference of these extreme values ​​on the model prediction results.

[0183] Step 1.5: Divide the processed data into training set and test set according to the set ratio.

[0184] Step 2: Based on the normalized data, a masked multi-head self-attention mechanism MSA module is constructed to capture important features in the input data.

[0185] Specifically, the multi-head self-attention mechanism, as a key module of the Transformer architecture, is used to capture important features in any historical sequence regardless of distance. The multi-head structure can significantly improve the learning performance of long-term dependencies by combining multiple self-attention modules with feature learning in different state subspaces. In essence, MSA (Multi-head Self-Attention) can achieve better prediction performance through the interaction of each head, while the single-head structure only emphasizes important features of certain stages and ignores other stages. In addition, in order to solve the problem of time series prediction, an additional masking mechanism is introduced in MSA to prevent future values ​​from being added to the calculation.

[0186] Preferably, step 2 specifically includes the following steps:

[0187] Step 2.1, input the normalized data into the MSA module to capture the important features of the time series through the multi-head mechanism;

[0188] Step 2.2, for the h-th attention head, h∈{1,…,P}, in cycle i (i=1,2,…,t), three vectors are obtained by multiplying the input sequence θ with the three weight matrices, named query (q), key (k) and value (v), respectively. This process can be expressed as:

[0189]

[0190] Where:

[0191] θ i Represents the input feature vector at time step i, with dimension d em ;

[0192] W q represents the query matrix, W k represents the bond matrix, W v represents the value matrix, W q , W k ,

[0193] (W q ) h represents the query weight matrix of the h-th attention head, (W k ) h represents the key weight matrix of the h-th attention head, (W v ) h The value weight matrix of the hth attention head has the dimension h∈{1,…,P};

[0194] represents the query vector generated by the h-th attention head at time step i; represents the key vector generated by the h-th attention head at time step i; represents the value vector generated by the h-th attention head at time step i;

[0195] Step 2.3, by scaling the query vector with the key vector, the attention weight is calculated, and the mask mechanism is applied to make the current time step focus only on the past time steps. The formula is as follows:

[0196]

[0197] Where:

[0198] represents the attention weight of the h-th attention head at time step i to time step τ;

[0199] Represents the scaling factor, which is used to control the dot product result;

[0200] mask i,τ Represents the mask value, which is used to process the information of time step τ and to block the information of future time steps; if τ>i, the mask is negative infinity, otherwise it is 0;

[0201] Represents the attention weight Apply the softmax function, and the resulting weights are between 0 and 1, representing the attention distribution at each time step, which is used to convert the weights into probabilities;

[0202] represents the attention output after weighted summation; λ i-τ represents the time attenuation coefficient;

[0203] Specifically, in lithium battery life prediction, time series data contains a lot of historical information. However, as time goes by, the impact of older data on the prediction of the current state gradually weakens. If the model assigns equal weights to all time steps, it may cause noise or outdated information in historical data to interfere with the prediction. Therefore, introducing time decay weights can ensure that the model pays more attention to recent data and reduce the negative impact of historical information.

[0204] Step 2.4, by defining multiple sliding windows, sliding on the time series with different window lengths, extracting different time segment data, representing the performance of the time series at different time scales, and aggregating the features of different time windows into the feature set;

[0205] Step 2.5, evaluate the importance of features through decision trees, check the correlation between features and the remaining service life RUL of lithium-ion batteries, fuse features through PCA (Principal Component Analysis) method, and construct an important feature set;

[0206] Step 2.6, output the masked MSA result, the formula is as follows:

[0207]

[0208] Where:

[0209] It means concatenating the output vectors of all P attention heads at time step i;

[0210] Wo Represents the weight matrix of linear projection, with a size of (P·d k )×d em , d k is the dimension of each head, d em is the dimension of the final output;

[0211] linear{·} represents a linear transformation operation.

[0212] Step 3: Based on the extracted important features, the key hyperparameters are preliminarily set to build the MSA-LSTM comprehensive prediction model to form a complete model structure for processing time series data and predicting battery life;

[0213] Specifically, the long short-term memory network is a variant of the recurrent neural network (RNN). Since three gating units, namely input gate, output gate and forget gate, are added to the RNN, LSTM has obvious advantages in processing long time series data, so it is often used to solve the gradient disappearance, gradient explosion and nonlinear mapping problems of traditional neural networks in the process of sequence training. The specific structure is as follows Figure 3 (a) shown.

[0214] Preferably, in step 3, constructing the MSA-LSTM comprehensive prediction model specifically includes the following steps:

[0215] Step 3.1: Take the extracted important features as the current input x t The forget gate is introduced. The forget gate is mainly used to selectively "forget" the initial output of the previous moment and the input capacity information of this moment, and retain important information after filtering through the Sigmoid function. The formula is as follows:

[0216] f t =σ(w f ·[h t-1 ,x t ]+b f )

[0217] Where:

[0218] x t 、f t Indicates the gate input capacity and output capacity;

[0219] w f represents the connection weight parameter;

[0220] h t-1 Indicates the output status at the previous moment;

[0221] b f is the offset parameter;

[0222] σ means the output range of the Sigmoid function is [0,1];

[0223] Step 3.2, the input gate controls the transmission of new information to the cell state, the Sigmoid function is used to determine the value to be updated, and the tanh function generates the candidate state. The formula is as follows:

[0224] i t =σ(w i ·[h t-1 ,x t ]+b i )

[0225]

[0226] Where:

[0227] i t represents the output of the input gate;

[0228] w i represents the weight parameter of the input gate; w s The weight parameter representing the candidate cell state;

[0229] b i represents the offset parameter of the input gate; b s The offset parameter representing the candidate cell state;

[0230] represents the candidate cell state;

[0231] Step 3.3, combine the output of the forget gate and the input gate to update the current cell state. The formula is as follows:

[0232] O t =σ(w o ·[h t-1 , x t ]+b o )

[0233]

[0234] h t =O t ×tanh(C t )

[0235] Where:

[0236] O t represents the output of the output gate;

[0237] w o Represents the weight matrix of the output gate;

[0238] b o Represents the offset parameter of the output gate;

[0239] h t Indicates the hidden state at the current moment;

[0240] C t Indicates the cell state at the current moment; C t-1 Indicates the cell state at the previous moment;

[0241] tanh(C t ) represents the hyperbolic tangent function, which is used to convert the cell state C t Map to the interval [-1,1];

[0242] Generally speaking, the hidden layer of an LSTM deep neural network is composed of many identical LSTM units stacked together. Figure 3 (b) is the model structure diagram of the deep LSTM network. The circles in the figure represent the nodes of the LSTM deep neural network, and each node is an LSTM unit. N is the number of hidden layers.

[0243] Step 3.4 Considering the nonlinearity and noise in the battery capacity data, using the Dropout layer in the LSTM deep neural network helps to improve the generalization ability of the overall model. For a given layer of the designed LSTM neural network, the mathematical representation of Dropout in the training phase is as follows:

[0244] y=F(Wx)*m

[0245] Where:

[0246] y represents the output of the layer;

[0247] F represents the activation function of the layer;

[0248] x represents the input of the layer;

[0249] m represents the Dropout rate;

[0250] W represents the weight matrix;

[0251] In step 3.5, considering the local fluctuation and recovery problems during the battery capacity degradation process, a LeakyReLU activation function layer is added after the LSTM layer to construct the MSA-LSTM comprehensive prediction model, form a complete model structure, and output the battery capacity prediction value; the activation function is the LeakyReLU (Leaky Rectified Linear Unit) activation function, which is an improved ReLU activation function. It solves the problem that the standard ReLU has a zero gradient in the negative input area and optimizes the nonlinear expression ability of the model.

[0252] Preferably, in step 3.5, the Leaky ReLU activation function formula is as follows:

[0253]

[0254] Where:

[0255] α = 0.01;

[0256] When the input x>0, the output of Leaky ReLU is directly the input value, which is consistent with the behavior of the ReLU function and remains positive;

[0257] When the input x≤0, the output of Leaky ReLU is α, which is different from the behavior of the ReLU function. The standard ReLU outputs 0 when the input is negative, while the LeakyReLU outputs a negative value when the input is negative.

[0258] It is worth noting that the MSA-LSTM comprehensive prediction model structure is constructed. In order to reduce the complexity of the algorithm iteration process, combined with the characteristics of the small capacity data volume and the existence of capacity recovery in the data set, a MSA layer, two LSTM layers, a Dropout layer, a Dense layer, and finally an activation layer are set to output the battery capacity prediction value. LeakyReLU is placed after the fully connected layer to ensure that the output of the LSTM layer is not affected by the negative slope of LeakyReLU. At the same time, nonlinearity is introduced after the features are linearly transformed by the fully connected layer, which helps to improve the ability to capture nonlinear features. The specific network structure is as follows: Figure 4 shown.

[0259] Step 4: Use ABC (Artificial Bee Colony) to optimize the hyperparameters of the MSA-LSTM comprehensive prediction model;

[0260] Specifically, the artificial bee colony algorithm is a swarm intelligence optimization algorithm inspired by the foraging behavior of bees. Compared with other algorithms, because it only needs to seek the optimal solution to the problem and does not need to understand the special information of the problem, it can effectively avoid falling into the problem of local optimal solution; the calculation process of the ABC algorithm is relatively simple, and it mainly searches and optimizes through the simulation of bee behavior. When processing large-scale data or complex models, the calculation cost of the ABC algorithm is low and the efficiency is higher; the ABC algorithm only needs to set a small number of control parameters, which reduces the difficulty of algorithm tuning and makes the algorithm more convenient in practical applications; the ABC algorithm shows strong robustness when processing data with large noise and uncertainty, and can better cope with the complexity and variability of data in practical applications, and provide stable and reliable optimization results. The ABC algorithm regards the solution process of the optimization problem as the process of searching for the optimal solution in D-dimensional space, compares the new solution with the original solution, and retains the better solution through the greedy algorithm selection strategy.

[0261] Preferably, if Figure 5 As shown, step 4 specifically includes the following steps:

[0262] Step 4.1: Initialize the control parameters in the ABC algorithm, including the number of honey sources S N , the maximum number of cycles S for finding nectar sources and the number of termination cycles S mc ;

[0263] Specifically, initialize the control parameters in the ABC algorithm and set the number of honey sources S N is 10, the maximum number of cycles S for finding nectar sources is 50, and the number of termination cycles is 200.

[0264] Step 4.2, use the root mean square error RMSE to set the fitness function in the ABC algorithm to evaluate the quality of each nectar source. The formula is as follows:

[0265]

[0266] Where:

[0267] represents the predicted value;

[0268] y i represents the true value;

[0269] n represents the sample size;

[0270] Step 4.3, define the initial search range of the hyperparameters, set the initial honey source position of the ABC algorithm, and the honey source includes the hyperparameter combination;

[0271] Preferably, in step 4.3, the definition of the initial search range includes setting the number of attention heads P between 1 and 8, the batch size B between 16 and 256, the number of hidden layer units between 32 and 512, and the Dropout rate between 0.1 and 0.5.

[0272] Step 4.4, based on the determined search range, generate new nectar sources by using hired bee local search to explore the solution space; the formula is as follows:

[0273]

[0274] Where:

[0275] χ′ id Indicates the d-th dimension value of the current nectar source;

[0276] Represents a random number between [-1,1]; i = 1, 2, ..., S N ; d = 1, 2, ..., D; k is any nectar source except i, k ≠ i;

[0277] x kd represents the d-th dimension value of another nectar source;

[0278] Step 4.5, based on the information transmitted by the employed bees, select the optimal nectar source through the fitness probability; the fitness probability formula is as follows:

[0279]

[0280] Where:

[0281] P i represents the probability of the i-th nectar source being selected;

[0282] fit i represents the fitness value of the i-th nectar source;

[0283] S N Indicates the total number of nectar sources;

[0284] Represents the sum of fitness values ​​of all nectar sources;

[0285] Step 4.6, based on the selected optimal nectar source, train and test the MSA-LSTM comprehensive prediction model, use scout bees to search and replace the nectar source whose fitness has not been improved, and continue to explore new solution space; the formula is as follows:

[0286]

[0287] Where:

[0288] χ idrepresents the value of the i-th nectar source in the d-th dimension, indicating the position of the nectar source in the multidimensional search space;

[0289] represents the lower bound of the D-th dimension;

[0290] represents the upper bound of the Dth dimension;

[0291] rand(0,1) represents a random number in the interval [0,1], which is used to generate a random position of the nectar source in the Dth dimension;

[0292] Represents the range of the Dth dimension, which is used to scale the random numbers so that the random positions are within the range of the search space.

[0293] Step 5: Use the training set divided in step 1 to train the optimized MSA-LSTM comprehensive prediction model, use the test set for verification, evaluate the prediction performance, apply the MSA-LSTM comprehensive prediction model to predict unknown data, and calculate the remaining service life RUL of the energy storage lithium battery.

[0294] Preferably, step 5 specifically includes the following steps:

[0295] Step 5.1: Based on the optimal solution set found in step 4, use K-fold cross validation to split the data set into different training sets and validation sets multiple times, and train the MSA-LSTM model on the training data set.

[0296] Specifically, by using multiple rounds of cross validation (K-fold Cross Validation) in the energy storage lithium battery life degradation prediction model, the generalization ability and prediction performance of the model can be significantly improved. Since lithium battery operation data is scarce and expensive, cross validation splits the data set into multiple subsets for training and validation in turn to make full use of the data and ensure that the model does not perform poorly due to insufficient data. At the same time, multiple rounds of validation can reduce the randomness of the model results and make the prediction more reliable by averaging multiple test results. This method can also help the model better adapt to complex time series data changes, ensure consistent performance in different time periods and environments, and improve the robustness of the model. In addition, cross validation can find the best hyperparameter configuration, avoid suboptimal selection caused by a single division, and enable the model to have stable prediction capabilities under a variety of conditions. Overall, multiple rounds of cross validation ensure the stability, accuracy and long-term adaptability of the model in diversified data, and provide more accurate remaining service life (RUL) prediction support for battery management systems.

[0297] Step 5.2: Use the test data set to verify the trained model and calculate the mean absolute error (MAE), root mean square error (RMSE), and R-squared error (R) of the model. 2Performance indicator, which evaluates the predictive ability on the test data set; the formula is as follows:

[0298]

[0299] Where:

[0300] represents the predicted value;

[0301] y i represents the true value;

[0302] n represents the total number of samples.

[0303] Specifically, performance evaluation indicators are set. Mean Absolute Error (MAE), Root Mean Square Error (RMSE) and R-square (R^2) are selected as evaluation criteria. The closer RMSE or MAE is to zero, the higher the accuracy of the proposed method. The closer R2 is to 1, the closer the predicted capacity value is to the actual capacity value, and the better the model fitting effect is.

[0304] The data is divided in a ratio of 4:6, and the model trained on the first 40% of the data is used to test the model performance of the remaining 60% of the data. The specific prediction results are as follows: Figure 7 As shown. Figure 7 It can be seen that the proposed algorithm has a high degree of fitting to the real capacity curve and shows a high accuracy of prediction performance. At the same time, for the case where the B06 and B18 training data fluctuate greatly, the proposed algorithm still shows a good prediction effect.

[0305] The data is divided in a ratio of 6:4. Similarly, the first 60% is used for model training, and the last 40% is used for model performance testing. The test results are as follows: Figure 8 As shown. Figure 8 , combined with the prediction results based on 40% of the training data, it can be concluded that: whether it is a small sample data or a large sample data, ABC-MSA-LSTM has shown excellent prediction performance, which can not only capture the global trend of capacity decay, but also cope with unknown disturbances caused by local capacity recovery and noise, and has high prediction accuracy and robustness. Among them, the MAE in the prediction results based on 60% of the training data is maintained at around 0.005, the RMSE is as low as 0.0084, and the fitting coefficient is as high as 97.5%.

[0306] Example 2 of the present invention provides a system for predicting the life degradation of a lithium energy storage battery, including: a data acquisition and preprocessing module, a feature extraction and selection module, a model training and optimization module, a prediction and diagnosis module, and a result display and feedback module;

[0307] The data acquisition and preprocessing module is used to collect battery operation data including voltage, current, and temperature, and preprocess the data, including cleaning, denoising, and normalization;

[0308] The feature extraction and selection module is used to extract important features from the preprocessed data and perform feature selection to capture key information about battery degradation;

[0309] The model training and optimization module is used to train the battery life prediction model, using machine learning and deep learning algorithms to improve the model's prediction performance through evaluation and optimization;

[0310] The prediction and diagnosis module is used to use the trained model to predict the remaining service life (RUL) of the battery and diagnose the battery health status;

[0311] The result display and feedback module is used to display the prediction results and diagnostic information to users in a visual manner to provide decision support; at the same time, the model is updated and maintained based on actual operating conditions and new data.

[0312] In summary, the practical effects of the present invention are summarized as follows:

[0313] The present invention significantly improves the accuracy and stability of the remaining useful life (RUL) prediction of lithium-ion batteries by adopting a masked multi-head self-attention mechanism combined with the LeakyReLU activation function and the ABC optimization algorithm. The ABC-MSA-LSTM method not only optimizes the computational efficiency and accelerates the convergence of the model, but also enhances the generalization ability of the model, so that it can adapt to different battery types and working conditions, and has broad application prospects. In addition, battery degradation monitoring is crucial to ensure the safety, reliability and economy of the battery system. The present invention can accurately capture the degradation characteristics of the battery under different conditions of use, realize real-time monitoring and prediction of the battery health status, and thus provide a battery management system with an efficient and intelligent decision support tool. The application of this technology can not only extend the service life of the battery and reduce maintenance costs, but also prevent safety accidents caused by battery failures, and has important social and economic value.

[0314] The present invention may be a system, a method and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for causing a processor to implement various aspects of the present invention.

[0315] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents, and any modifications or equivalent replacements that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for predicting the life degradation of energy storage lithium batteries. It is characterized in that The following steps are involved: Step 1: Obtain and preprocess lithium battery operation data, construct time series and perform normalization, and roll-divide training and test sets; Step 2: Based on the normalized data, a masked multi-head self-attention mechanism MSA module is constructed to capture multiple features in the input data through different time windows, and dynamic weights are assigned to different attention heads to extract important features. After screening, the important feature set is extracted through feature fusion method; Step 3: Based on the important feature set, preliminarily set the key hyperparameters, build multiple multi-head self-attention-long short-term memory network MSA-LSTM comprehensive prediction models with dropout and activation function layers, and dynamically adjust the hyperparameters; Step 4, optimize the hyperparameters of the multi-head self-attention-long short-term memory network MSA-LSTM comprehensive prediction model through the artificial bee colony algorithm to find the optimal solution set; Step 5: Use the training set divided in step 1 to train the optimized MSA-LSTM comprehensive prediction model, use the test set to perform multiple rounds of cross-validation, evaluate the prediction performance, apply the MSA-LSTM comprehensive prediction model to predict unknown data, and calculate the remaining service life RUL of the energy storage lithium battery.

2. A method for predicting the life degradation of an energy storage lithium battery according to claim 1, Features: Step 1 includes: Step 1.1, obtain multi-dimensional data of lithium battery capacity, voltage, temperature, discharge rate and external environment, and perform preprocessing, including removing missing values ​​and outliers; Step 1.2, use the isolation forest algorithm to detect and remove abnormal data, and use linear interpolation to repair missing values; Step 1.3, construct the time series data of sliding windows and extended windows according to different time granularities; Step 1.4, according to the constructed time series features, select the optimal normalization method, including min-max normalization and z-score normalization; In step 1.5, the generative adversarial network (GAN) is used to generate expanded training data, and the training set and test set are dynamically divided through the rolling partition strategy.

3. A method for predicting the life degradation of an energy storage lithium battery according to claim 2, Features: In step 1.4, the min-max normalization is calculated as follows: Where: x i Represents the preprocessed data; y i represents the normalized data; max(x) and min(x) represent the corresponding maximum and minimum values ​​in the original capacity data respectively; The z-score standardization is calculated as follows: Where: x i represents the preprocessed data; μ represents the average value of the preprocessed data features; σ represents the standard deviation of the preprocessed data features, which is used to measure the volatility of the data.

4. A method for predicting the life degradation of an energy storage lithium battery according to claim 1, Features: Step 2 includes: Step 2.1, input the normalized data into the MSA module, and capture the important features of the time series by combining the sliding window and the jumping window of the multi-head self-attention mechanism; Step 2.2, in the MSA module, dynamically assign weights to each attention head to map the input sequence θ into query q, key k, and value v vectors through three weight matrices, as follows: Where: θ i represents the input feature vector at time step i; W q represents the query matrix, W k represents the bond matrix, W v represents the value matrix, (W q ) h represents the query weight matrix of the h-th attention head, (W k ) h represents the key weight matrix of the h-th attention head, (W v ) h represents the value weight matrix of the h-th attention head, h∈{1,…,P}; represents the query vector generated by the h-th attention head at time step i; represents the key vector generated by the h-th attention head at time step i; represents the value vector generated by the h-th attention head at time step i; Step 2.3, by scaling the query vector with the key vector, the attention weight is calculated, and the mask mechanism is applied to make the current time step focus only on the past time steps. The formula is as follows: Where: represents the attention weight of the h-th attention head at time step i to time step τ; Represents the scaling factor, which is used to control the dot product result; mask i，τ Represents the mask value, which is used to process the information of time step τ and to block the information of future time steps; if τ>i, the mask is negative infinity, otherwise it is 0; Represents the attention weight Apply the softmax function, and the weight obtained is between 0 and l, which represents the attention distribution at each time step and is used to convert the weight into probability; represents the attention output after weighted summation; λ i-τ represents the time attenuation coefficient; Step 2.4, by defining multiple sliding windows, sliding on the time series with different window lengths, extracting different time segment data, representing the performance of the time series at different time scales, and aggregating the features of different time windows into the feature set; Step 2.5, evaluate the importance of features through decision trees, check the correlation between features and the remaining useful life RUL of lithium-ion batteries, fuse features through principal component analysis PCA method, and construct an important feature set; Step 2.6, output the masked MSA result, the formula is as follows: Where: It means concatenating the output vectors of all P attention heads at time step i; W o Represents the weight matrix of linear projection, with a size of (P·d k )×d em , d k is the dimension of each head, d em is the dimension of the final output; linear{·} represents a linear transformation operation.

5. A method for predicting the life degradation of an energy storage lithium battery according to claim 1, Features: In step 3, the MSA-LSTM comprehensive prediction model is constructed, including: Step 3.1: Take the extracted important features as the current input x t The forget gate is introduced to retain important information after filtering through the Sigmoid function. The formula is as follows: T t =σ(w f ·[h t-1 ,x t ]+b f ) Where: x t , T t Indicates the gate input capacity and output capacity; w f represents the connection weight parameter; h t-1 Indicates the output status at the previous moment; b f is the offset parameter; σ means the output range of the Sigmoid function is [0, 1]; Step 3.2, the input gate controls the transmission of new information to the cell state, the Sigmoid function is used to determine the value to be updated, and the tanh function generates the candidate state. The formula is as follows: i t =σ(w i ·[h t-1 ,x t ]+b i ) Where: i t represents the output of the input gate; w i represents the weight parameter of the input gate; w s The weight parameter representing the candidate cell state; b i represents the offset parameter of the input gate; b s The offset parameter representing the candidate cell state; represents the candidate cell state; Step 3.3, combine the output of the forget gate and the input gate to update the current cell state. The formula is as follows: The t =σ(w o ·[h t-1 ,x t ]+b o ) h t =O t ×tanh(C t ) Where: O t represents the output of the output gate; w o Represents the weight matrix of the output gate; b o Represents the offset parameter of the output gate; h t Indicates the hidden state at the current moment; C t Indicates the cell state at the current moment; C t-1 Indicates the cell state at the previous moment; tanh(C t ) represents the hyperbolic tangent function, which is used to convert the cell state C t Map to the interval [-1, 1]; Step 3.4, taking into account the nonlinearity and noise in the battery capacity data, the Dropout layer is used in the LSTM deep neural network to train the generalization ability of the model; the formula of Dropout in the training stage is as follows: y=F(Wx)*m Where: y represents the output of the layer; F represents the activation function of the layer; x represents the input of the layer; m represents the Dropout rate; W represents the weight matrix; In step 3.5, add a LeakyReLU activation function layer after the LSTM layer to build the MSA-LSTM comprehensive prediction model, form a complete model structure, output the battery capacity prediction value, solve the problem that the standard ReLU has a zero gradient in the negative input area, and optimize the nonlinear expression ability of the model.

6. A method for predicting the life degradation of an energy storage lithium battery according to claim 5, Features: In step 3.5, the LeakyReLU activation function formula is as follows: Where: a＝0.01； When the input x>0, the output of LeakyReLU is directly the input value, which is consistent with the behavior of the ReLU function and remains positive; When the input x≤0, the output of LeakyReLU is α, which is different from the behavior of the ReLU function. The standard ReLU outputs 0 when the input is negative, while the LeakyReLU outputs a negative value when the input is negative.

7. A method for predicting the life degradation of an energy storage lithium battery according to claim 1, Features: Step 4 includes: Step 4.1: Initialize the control parameters in the artificial bee colony algorithm, including the number of nectar sources S N , the maximum number of cycles S for finding nectar sources and the number of termination cycles S mc ; Step 4.2, use the root mean square error RMSE to set the fitness function in the ABC algorithm to evaluate the quality of each nectar source. The formula is as follows: Where: represents the predicted value; y i represents the true value; n represents the sample size; Step 4.3, define the initial search range of the hyperparameters, and set the initial honey source position of the artificial bee colony algorithm, where the honey source is a combination of hyperparameters; Step 4.4, based on the determined search range, generate new nectar sources by using hired bee local search to explore the solution space; the formula is as follows: Where: χ′ id Indicates the d-th dimension value of the current nectar source; represents a random number between [-1, 1]; i = 1, 2, ..., S N ; d = 1, 2, ..., D; k is any nectar source except i, k ≠ i; χ kd represents the d-th dimension value of another nectar source; Step 4.5, based on the information transmitted by the employed bees, select the optimal nectar source through the fitness probability; the fitness probability formula is as follows: Where: P i represents the probability of the i-th nectar source being selected; fit i represents the fitness value of the i-th nectar source; S N Indicates the total number of nectar sources; Represents the sum of fitness values ​​of all nectar sources; Step 4.6, based on the selected optimal nectar source, train and test the MSA-LSTM comprehensive prediction model, use scout bees to search and replace the nectar source whose fitness has not been improved, and continue to explore new solution space; the formula is as follows: Where: χ id represents the value of the i-th nectar source in the d-th dimension, indicating the position of the nectar source in the multidimensional search space; represents the lower bound of the d-th dimension; represents the upper bound of the d-th dimension; rand(0,1) represents a random number in the interval [0,1], which is used to generate a random position of the nectar source in the dth dimension; Represents the range of the dth dimension, which is used to scale the random numbers so that the random positions are within the range of the search space.

8. A method for predicting the life degradation of an energy storage lithium battery according to claim 7, Features: In step 4.3, the initial search range is defined including the number of attention heads P being set between 1 and 8, the batch size B being set between 16 and 256, the number of hidden layer units being set between 32 and 512, and the Dropout rate being set between 0.1 and 0.

5.

9. A method for predicting the life degradation of an energy storage lithium battery according to claim 1, Features: Step 5 includes: Step 5.1: Based on the optimal solution set found in step 4, use K-fold cross validation to split the data set into different training sets and validation sets multiple times, and train the MSA-LSTM model on the training data set. Step 5.2: Use the test set to verify the trained MSA-LSTM model and calculate the mean absolute error (MAE), root mean square error (RMSE), and R 2 Performance indicator, which evaluates the predictive ability on the test data set; the formula is as follows: Where: represents the predicted value; y i represents the true value; n represents the total number of samples.

10. A system for predicting the life degradation of energy storage lithium batteries. include: Data collection and preprocessing module, feature extraction and selection module, model training and optimization module, prediction and diagnosis module, and result display and feedback module; A method for predicting the life degradation of a lithium energy storage battery according to any one of claims 1 to 9 is implemented, characterized in that: The data acquisition and preprocessing module is used to collect battery operating capacity data and preprocess the data, including cleaning, denoising, and normalization; The feature extraction and selection module is used to extract important features from the preprocessed data and perform feature selection to capture key information about battery degradation; The model training and optimization module is used to train the battery life prediction model, using machine learning and deep learning algorithms to improve the model's prediction performance through evaluation and optimization; The prediction and diagnosis module is used to use the trained model to predict the remaining service life (RUL) of the battery and diagnose the battery health status; The result display and feedback module is used to display the prediction results and diagnostic information to users in a visual manner to provide decision support; at the same time, the model is updated and maintained based on actual operating conditions and new data.

Images