Lithium ion battery remaining service life prediction method based on fusion model

Abstract

The invention discloses a method for predicting the remaining service life of a lithium ion battery based on a fusion model, and belongs to the technical field of battery life prediction.The method comprises the steps of data preprocessing, RUL prediction model building and hyper-parameter optimization, feature parameters are efficiently mined, simplification and specification of the parameters are achieved, and the preprocessed data serve as credible input variables of the RUL prediction model. The RUL prediction model adopts a multi-channel parallel input mode, battery data are input into the fusion model MC-BiTCN-BiGRU-EAA, battery charging and discharging data are fully utilized, and the problem of insufficient data volume is solved; according to the multi-channel parallel input, the models are fused on the basis of maximizing the advantages of a single model, the gradient disappearance and gradient explosion phenomena are avoided, and the calculation cost is reduced. In addition, the precision and the calculation efficiency of model prediction are further improved through hyper-parameter optimization.

Description

Technical Field

[0001] This application relates to the technical field of battery life prediction, and particularly to a method for predicting the remaining useful life of a lithium-ion battery based on a fusion model. Background Art

[0002] Lithium-ion batteries have become the core energy carriers in clean energy application scenarios such as new energy vehicle power systems and energy storage power stations due to their high energy density, fast charge and discharge capabilities, and environmental friendliness. However, due to the complex internal aging mechanism of lithium batteries, their performance deteriorates during cyclic charge and discharge processes, leading to safety accidents such as power outages, fires, or explosions. The RUL (Remaining Useful Life) of a lithium battery represents the remaining number of charge and discharge cycles from the measurement start to the end of life (End of Life, EOL), which can be used as a key indicator to characterize the battery health state. Therefore, the accuracy of the lithium battery RUL evaluation method is directly related to the overall performance of the battery management system, and its accurate prediction has important engineering value for realizing fault warning, optimizing maintenance strategies, and ensuring the safe operation of the energy system.

[0003] Currently, the technical system for lithium battery RUL mainly includes three types of methods: mechanism model-based methods, data-driven methods, and fusion model-based methods. Among them, mechanism model-based methods include equivalent circuit models, electrochemical models, and empirical models. Due to the complex internal electrochemical characteristics of the battery, which are easily affected by external factors such as temperature and humidity, it is difficult to establish a general and accurate mechanism model. Data-driven methods directly study the relationship between model inputs and outputs by using historical data to predict RUL without the need to understand the internal mechanism of the battery, which enhances the feasibility of modeling. Fusion model-based methods can give full play to the advantages of multiple driving methods to improve the accuracy of machine learning prediction models and have become the main research direction of domestic and foreign scholars at present. Due to the limited data availability of current lithium batteries, it restricts the offline training of data-driven models, resulting in low prediction method accuracy and inaccurate prediction results. Moreover, existing algorithms are affected by battery experimental data noise and the defects of single algorithm models themselves, and phenomena such as gradient disappearance and gradient explosion will occur, and they usually have high computational requirements and weak generalization capabilities.

[0004] Therefore, there is an urgent need to research and develop a prediction method based on a fusion model to improve the accuracy of predicting the remaining useful life of lithium batteries. Summary of the Invention

[0005] In view of the above deficiencies in the prior art, a method for predicting the remaining useful life of a lithium-ion battery based on a fusion model provided by this application solves problems such as insufficient data volume, inaccurate prediction results, and low calculation efficiency.

[0006] To achieve the above-mentioned invention purpose, the technical solution adopted in this application is as follows: This application provides a method for predicting the remaining service life of a lithium-ion battery based on a fusion model, including: S1: Extract the charge and discharge data during all charge and discharge cycles in the lithium-ion battery detection dataset publicly available from NASA; S2: Perform feature extraction, feature dimensionality reduction, and feature scaling on the extracted charge and discharge data, and form the extracted charge and discharge features into an initial dataset; S3: Divide the initial dataset into a training set and a test set, and build an MC-BiTCN-BiGRU-EAA fusion model based on the training set as the RUL prediction model for the lithium battery; S4: Use the TPE algorithm to obtain the hyperparameters of the lithium battery RUL prediction model, and optimize based on the hyperparameters to obtain an optimized lithium battery RUL prediction model; S5: Predict the remaining service life of the lithium-ion battery according to the lithium battery RUL prediction model and the test set; S6: Evaluate the accuracy of the prediction according to the predicted remaining service life of the lithium-ion battery by using the mean absolute error, mean square error, mean absolute percentage error, and root mean square error.

[0007] Further, the specific steps of S2 include: S201: Perform feature extraction on the extracted charge and discharge data based on the Pearson correlation coefficient method to obtain the input features of the lithium battery RUL prediction model. The formula of the Pearson correlation coefficient method is:

[0008] Among them, is the value of the x variable in the sample, is the average value of the x variable, is the value of the y variable in the sample, is the average value of the y variable, is the total number of samples, is the th sample, is the correlation between variables; S202: Perform principal component analysis on the input features of the lithium battery RUL prediction model to obtain the input features after dimensionality reduction; S203: Use the maximum-minimum scaling method to normalize all features in the lithium-ion battery detection dataset to obtain the initial dataset. The formula of the maximum-minimum scaling method is:

[0009] Among them, is the value after the extracted feature is normalized, is the value of the extracted feature, is the minimum value of the extracted feature, is the maximum value of the extracted feature.

[0010] Furthermore, the principal component analysis in S202 specifically includes: A1: Set the D-dimensional training set X representing the original data matrix, and calculate the covariance matrix of X:

[0011] Among them, , is the number of battery charge and discharge cycles, is the th sample vector, with a dimension of , is the mean vector, with a dimension of ; A2: Use the non-linear iterative partial least squares or singular value decomposition method to solve the eigenvalues and eigenvectors of the covariance matrix; A3: Sort the eigenvalues in descending order, select the top d ones, project the original features onto the selected eigenvectors, and obtain the new d -dimensional input features after dimensionality reduction.

[0012] Furthermore, building the MC-BiTCN-BiGRU-EAA fusion model in S3 specifically includes: S301: Use a single-layer BiTCN model to extract the multi-channel charging data features in the training set, and use the extracted charging current, charging voltage, and charging temperature features as the inputs of a single layer of BiTCN respectively; S302: Connect the extracted multi-channel charging data features through a connection layer, input them into a multi-layer BiTCN model with a residual connection, and embed an additive attention mechanism layer into the BiTCN model to output charging features; S303: Input the discharge capacity data in the training set into the BiGRU model to output discharge features; S304: Connect the output charging features of the multi-layer BiTCN model and the output discharge features of the BiGRU model through a fully connected layer to obtain the MC-BiTCN-BiGRU-EAA fusion model.

[0013] The BiTCN model consists of two parts: dilated convolution and causal convolution, including an input layer, two hidden layers, and an output layer. Among them, the calculation formula of the dilated convolution is:

[0014] Among them, represents the output value of the dilated convolution at time step . is the convolution kernel size, represents the weight parameter at the th position in the product kernel, represents the input value at time step , represents the dilation factor.

[0015] Furthermore, the residual connection calculation formula of the multi-layer BiTCN model with the residual link is:

[0016] Among them, represents the convolution input, represents the convolution output, represents the output of the residual connection, is the activation function.

[0017] Furthermore, the output of the efficient additive attention in the additive attention mechanism layer is:

[0018] Among them, represents the normalized query matrix, represents the linear transformation, represents a single global query vector, represents the key matrix, is the efficient additive attention.

[0019] Furthermore, the S4 specifically includes: S401: Determine whether the lithium battery RUL prediction model has completed hyperparameter initialization. If so, update the hyperparameter combination based on the historical training records and enter S402; if not, randomly generate an initial hyperparameter combination within the preset search space and enter S402; S402: Based on the historical hyperparameter combinations and their corresponding performance metrics, divide the hyperparameters into two categories: the set of the top 20% hyperparameters with the optimal performance , and the remaining set of ordinary hyperparameters ; S403: Use Gaussian process regression to construct probability density models for the two types of hyperparameters respectively: the excellent distribution characterizing the statistical characteristics of , the ordinary distribution characterizing the statistical characteristics of , among which, Refers to the hyperparameter combination to be optimized; S404: Define the expected improvement function as , and select a better hyperparameter combination by maximizing the value of the expected improvement function ; S405: Based on the new hyperparameter combination Train the RUL prediction model of the lithium battery and calculate the prediction error of the training set; S406: Determine whether the prediction error meets the error minimization. If so, enter S409; if not, enter S407; S407: Determine whether the calculation reaches the maximum number of iterations or the calculation time limit. If so, enter S409; if not, enter S408; S408: Dynamically shrink the search space according to the high-probability region during the iteration I(x) , re-divide and after each iteration, and update I(x) and G(x) distribution parameters and enter S404; S409: Output the hyperparameter combination with the minimum training set error , and obtain the optimized RUL prediction model configuration based on the hyperparameter combination .

[0020] Furthermore, the calculation formulas for the mean absolute error, mean square error, mean absolute percentage error, and root mean square error in S6 are as follows:

[0021]

[0022] In the formula, is the original battery capacity value, is the predicted value, is the mean absolute error, is the mean square error, is the mean absolute percentage error, is the root mean square error, represents the number of samples, is the th sample.

[0023] Beneficial effects of the embodiments of the present application: A method for predicting the remaining useful life of a lithium-ion battery based on a fusion model provided by this application. This method includes data preprocessing, construction of an RUL prediction model, and hyperparameter optimization. Among them, data preprocessing adopts methods of feature selection, dimensionality reduction, and scaling. By analyzing the internal laws of the collected data, feature parameters are efficiently mined, and parameter reduction and regularization are achieved, so that the preprocessed data serves as reliable input variables for the RUL prediction model. The RUL prediction model adopts a multi-channel parallel input method, inputting battery data into the fusion model MC-BiTCN-BiGRU-EAA, making full use of battery charge and discharge data, and bidirectionally extracting battery information to make up for the problem of insufficient data volume; multi-channel parallel input maximizes the advantages of a single model and then fuses the models, avoiding the occurrence of gradient disappearance and gradient explosion phenomena, and reducing the computational cost. In addition, hyperparameter optimization further improves the accuracy and computational efficiency of battery life prediction. Description of the Drawings

[0024] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of this application, and those of ordinary skill in the art can also obtain other embodiments based on these drawings.

[0025] Figure 1 It is a schematic diagram of a method for predicting the remaining useful life of a lithium-ion battery based on a fusion model provided by an embodiment of this application.

[0026] Figure 2 It is a graph showing the capacity decay of NASA lithium batteries B0005, B0006, B0007, and B0018 provided by an embodiment of this application.

[0027] Figure 3 It is a graph showing the capacity decay of NASA lithium batteries B0025 - B0028 provided by an embodiment of this application.

[0028] Figure 4 It is a graph of the charge and discharge process at a specific cycle provided by an embodiment of this application.

[0029] Figure 5 It is a graph showing the analysis results of the Pearson correlation coefficient provided by an embodiment of this application.

[0030] Figure 6 It is a schematic diagram of the MC-BiTCN-BiGRU-EAA fusion model provided by an embodiment of this application.

[0031] Figure 7 It is a schematic diagram of the two-way causal convolution principle of BiTCN provided by an embodiment of this application.

[0032] Figure 8 A schematic diagram of BiGRU provided by an embodiment of this application.

[0033] Figure 9 A comparison chart of the single-channel model prediction results of lithium battery B0007 provided by an embodiment of this application.

[0034] Figure 10 A comparison chart of the single-channel model prediction results of lithium battery B0018 provided by an embodiment of this application.

[0035] Figure 11 A comparison chart of the multi-channel model prediction results of lithium battery B0007 provided by an embodiment of this application.

[0036] Figure 12 A comparison chart of the multi-channel model prediction results of lithium battery B0018 provided by an embodiment of this application. Detailed implementation manners

[0037] Next, the technical solutions in the embodiments of this application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of this application. Obviously, the described embodiments are only a part of the embodiments of this application, rather than all the embodiments. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art based on this application belong to the protection scope of this application.

[0038] An embodiment of this application provides a method for predicting the remaining useful life of a lithium-ion battery based on a fusion model. This method can be referred to Figure 1 , Figure 1 As shown in the following figure, a schematic diagram of a method for predicting the remaining useful life of a lithium-ion battery based on a fusion model provided by an embodiment of this application includes: S1: Extract the charge and discharge data during all charge and discharge cycles in the lithium-ion battery detection dataset publicly available from NASA.

[0039] In an embodiment of this application, a detection dataset of 18650-type lithium-ion batteries with a rated capacity of 2 Ah publicly available from NASA is used to extract effective information such as voltage, current, and temperature during all charge and discharge cycles of lithium batteries B0005, B0006, B0007, B0018, and B0025 - B0028 in the dataset. Among them, the capacity decay curves of the lithium batteries are as Figure 2 and Figure 3 shown, Figure 2 is a capacity decay curve diagram of NASA lithium batteries B0005, B0006, B0007, and B0018 provided by an embodiment of this application. Figure 3This is the capacity degradation curve of NASA lithium battery B0025 - B0028 provided by the embodiments of the present application, and the two show the capacity degradation of NASA lithium batteries B0005, B0006, B0007, B0018, B0025 - B0028.

[0040] As Figure 4 shown, Figure 4 This is a curve graph of the charge - discharge process for a specific cycle provided by the embodiments of the present application. During the charging process, the battery is charged at a constant current. When the voltage reaches the maximum value, it maintains a constant - voltage charging mode until the current drops to 20 mA, and the charging ends; during the discharging process, the battery discharges at a constant current until the battery voltage drops to a certain value and the discharging ends.

[0041] S2: Perform feature extraction, feature dimensionality reduction, and feature scaling on the extracted charge - discharge data, and form an initial data set with the extracted charge - discharge features.

[0042] Further, the S2 specifically includes: S201: Perform feature extraction on the extracted charge - discharge data based on the Pearson correlation coefficient method to obtain the input features of the lithium - battery RUL prediction model. Among them, the formula of the Pearson correlation coefficient method is:

[0043] Among them, is the value of the x variable in the sample, is the average value of the x variable, is the value of the y variable in the sample, is the average value of the y variable, is the total number of samples, is the th sample, is the correlation between variables.

[0044] In an embodiment of the present application, the Pearson correlation coefficient (PCC) method is introduced to extract features. As Figure 5 shown, Figure 5 This is a Pearson correlation coefficient analysis result graph provided by the embodiments of the present application. Using the distribution scales of the linear correlations among five different degrees of correlations of voltage, current, temperature, time, and capacity, the linear correlations among the selected five features are quantitatively evaluated. Finally, the charging current, voltage, temperature, and discharging capacity are selected as the four input features of the model.

[0045] S202: Perform principal - component analysis on the input features of the lithium - battery RUL prediction model to obtain the input features after dimensionality reduction.

[0046] Further, the principal - component analysis in S202 specifically includes: A1: Set the D - dimensional training set X representing the original data matrix, where d represents the new dimension (d ≤ D), and calculate the covariance matrix of X:

[0047] where, , is the number of battery charge - discharge cycles, is the th sample vector with dimension , is the mean vector with dimension ; A2: Use the non - linear iterative partial least squares or singular value decomposition method to solve the eigenvalues and eigenvectors of the covariance matrix; A3: Sort the eigenvalues in descending order, select the top d ones, project the original features onto the selected eigenvectors, and obtain the new d - dimensional input features after dimensionality reduction.

[0048] In an embodiment of the present application, principal component analysis (PCA) reduces the dimension of the dataset by removing irrelevant information while retaining most of the relevant information. The initial PCC selects four features, namely voltage, current, temperature, and capacity, as input to the model. After principal component analysis, the number of features becomes three. However, these three features are not three of the four original input features. Specifically, for each factor, 160 points are selected from a total of 432000 points. After PCC and PCA, the total number of data points is reduced from 3050800 to 1659300.

[0049] S203: Normalize all features in the lithium - ion battery detection dataset using the min - max scaling method to obtain the initial dataset, where the formula of the min - max scaling method is:

[0050] where, is the value of the extracted feature after normalization, is the value of the extracted feature, is the minimum value of the extracted feature, is the maximum value of the extracted feature.

[0051] In an embodiment of the present application, after feature selection and dimensionality reduction, the feature scaling method is used to normalize all features in the lithium - ion battery detection dataset to the interval [0, 1]. This will help the optimization mechanism of the model quickly reach the minimum cost function and improve the training performance of the model. Here, the min - max scaling method is used as the feature scaling method.

[0052] The beneficial effects of the above further solution are as follows: By performing correlation analysis on the data, features with higher correlation with RUL are obtained, and feature dimensionality reduction and scaling are performed on them to reduce the computational cost of the computer and improve the training performance of the model.

[0053] S3: Divide the initial data set into a training set and a test set, and build an MC-BiTCN-BiGRU-EAA fusion model based on the training set as the lithium battery RUL prediction model.

[0054] In an embodiment of the present application, the processed initial data set is divided into a training set and a test set. Among them, batteries B0005, B0006, B00025 - B00028 are used as the training set, and B0007 and B00018 are used as the test set. The present application selects the PyTorch framework of Python to build an MC-BiTCN-BiGRU-EAA fusion model as the lithium battery RUL prediction model. It can be seen from Figure 6 , Figure 6 which is the schematic diagram of the MC-BiTCN-BiGRU-EAA fusion model provided by the embodiment of the present application. The specific construction steps are as follows: S301: Use a single-layer BiTCN model to extract multi-channel charging data features in the training set, and use the extracted charging current, charging voltage, and charging temperature features as the inputs of one layer of BiTCN respectively.

[0055] Among them, the BiTCN layer has 32 convolutional kernels with a size of 3x3, and the dilation rate is , which is used to expand the receptive field and enhance the generalization ability of the model. A Droput layer is added in BiTCN to prevent overfitting, and the droput rate is 0.6. As shown in Figure 7 . Figure 7 which is the schematic diagram of the two-way causal convolution principle of BiTCN provided by the embodiment of the present application. BiTCN consists of two parts: dilated convolution and causal convolution. Each part includes an input layer, two hidden layers, and an output layer. The calculation formula of the dilated convolution is:

[0056] In the formula, represents the output value of the dilated convolution at time step , is the convolution kernel size represents the weight parameter at the th position in the product kernel, represents the input value at time step , means that the convolution operation can only be performed on past inputs, represents the dilation factor.

[0057] The BiTCN model adopts bidirectional dilated causal convolution. By introducing bidirectional dilation factors, the receptive field of the convolutional kernel is continuously expanded, thereby obtaining a larger receptive field. This structure enhances the information extraction ability and improves the quality of feature extraction.

[0058] S302: Connect the extracted multi-channel charging data features through a connection layer, input them into a multi-layer BiTCN model with residual connections, and embed an additive attention mechanism layer into the BiTCN model to output charging features.

[0059] Among them, the calculation formula for the residual connection is:

[0060] In the formula, represents the convolutional input, represents the convolutional output, represents the output of the residual connection, is the activation function.

[0061] It can be understood that the multi-layer BiTCN has the same parameters as the single-layer BiTCN, but a residual connection module is added. Using dilated causal convolution to increase the receptive field of BiTCN will bring problems such as gradient vanishing and gradient explosion. The introduction of the residual block can avoid these problems and at the same time achieve efficient sequence feature extraction.

[0062] Embed the additive attention mechanism (Efficient Additive Attention, EAA) layer into BiTCN, where the efficient additive attention in the additive attention mechanism layer The output of is:

[0063] In the formula, represents the normalized query matrix, represents the linear transformation, represents a single global query vector, represents the key matrix, is the efficient additive attention.

[0064] The embedding of the additive attention mechanism layer reduces the computational amount, provides a faster inference speed, and can still capture important global context information.

[0065] S303: Input the discharge capacity data in the training set into the BiGRU model to output discharge features.

[0066] In one embodiment of the present application, due to the different capacity data and charging characteristics, which have more obvious time correlation and cumulative effect, the capacity data during the discharge process is input into the BiGRU model. Here, a BiGRU layer containing 32 convolutional kernels is used. As Figure 8 shown, Figure 8 FIG. 3 is a schematic diagram of a BiGRU provided by an embodiment of the present application. Two standard GRU neural networks that make up the BiGRU neural network. One GRU neural network is forward-trained using historical time series data, and the other GRU neural network is backward-trained using future time series data. The output features obtained from the forward training and the output features obtained from the backward training are fused to obtain the final output of the BiGRU neural network at a certain time series.

[0067] S304: Connect the output charging characteristics of the multi-layer BiTCN model and the output discharge characteristics of the BiGRU model through a fully connected layer to obtain an MC-BiTCN-BiGRU-EAA fusion model.

[0068] In one embodiment of the present application, the feature vectors output from BiTCN and BiGRU are combined and flattened, and then output through a fully connected layer, which are 32 and 1 node respectively.

[0069] S4: Use the TPE algorithm to optimize the hyperparameters of the lithium battery RUL prediction model to obtain an optimized lithium battery RUL prediction model.

[0070] Further, it specifically includes: S401: Determine whether the hyperparameters of the lithium battery RUL prediction model have been initialized. If so, update the hyperparameter combination based on the historical training records and enter S402; if not, randomly generate an initial hyperparameter combination within the preset search space and enter S402; S402: Based on the historical hyperparameter combinations and their corresponding performance metrics, divide the hyperparameters into two categories: the top 20% hyperparameter set with the optimal performance , and the remaining ordinary hyperparameter set ; S403: Use Gaussian process regression to construct probability density models for the two types of hyperparameters respectively: the excellent distribution representing the statistical characteristics of , the ordinary distribution representing the statistical characteristics of , where refers to the hyperparameter combination to be optimized; S404: Define the expected improvement function as , and select a better hyperparameter combination by maximizing the value of the expected improvement function ; S405: Based on the new hyperparameter combination Train the RUL prediction model of the lithium battery, and calculate the prediction error of the training set; S406: Determine whether the prediction error satisfies error minimization. If so, enter S409; if not, enter S407; S407: Determine whether the calculation reaches the maximum number of iterations or the calculation time limit. If so, enter S409; if not, enter S408; S408: Dynamically shrink the search space according to the high-probability region during the iteration I(x) Each time after iteration, re-divide and and update I(x) and G(x) distribution parameters of, and enter S404; S409: Output the hyperparameter combination with the minimum error of the training set and based on the hyperparameter combination obtain the optimized RUL prediction model configuration.

[0071] In an embodiment of the present application, the Tree-structured Parzen Estimation (TPE) algorithm is applied to optimize the model parameters to find the correct parameter combination to significantly improve the performance of the model. The hyperparameters of the MC-BiTCN-BiGRU-EAA model include: learning rate (lr), the number of output channels of the BiTCN layer (output-Channel), convolutional kernels (kernels), the number of filters (filters), input step (seq), batch size (batch_size), the number of single training iterations (epochs), and dropout probability (Dropout). At the same time, the mean absolute error is used as the objective function value of the TPE algorithm. Taking B007 as an example, the total number of trials of the TPE algorithm is set to n_trials = 1500 times, and the training cycle of BiTCN-BiGRU for each trial = 150 times. Among a total of 1500 tests, 985 tests are pruned, and the optimal hyperparameter results are lr = 0.0015, out_channels = 35, kernels = 5, filters = 32, seq = 5, epochs = 100, batch_Size = 13, Dropout = 0.6. For the hyperparameter results of other batteries, please refer to Table 1.

[0072] Table 1 Lithium battery hyperparameter optimization results

[0073] S5: According to the lithium battery RUL prediction model and the test set, predict the remaining service life of the lithium-ion battery.

[0074] S6: According to the predicted remaining service life of the lithium-ion battery, the mean absolute error, mean square error, mean absolute percentage error, and root mean square error are used to evaluate the accuracy of the prediction.

[0075] Among them, the calculation formulas for the mean absolute error, mean square error, mean absolute percentage error, and root mean square error are as follows:

[0076]

[0077] In the formula, is the original battery capacity value, is the predicted value, is the mean absolute error, is the mean square error, is the mean absolute percentage error, is the root mean square error, represents the number of samples, is the th sample.

[0078] For , , and indicators, when they approach zero, the higher the RUL prediction accuracy.

[0079] In an embodiment of the present application, compared with the traditional single-channel model (only using the lithium battery discharge capacity as one-dimensional input), multi-channel feature extraction can consider more factors and theoretically can capture the capacity regeneration phenomenon more accurately. The MC-BiTCN-BiGRU model proposed in the present application was experimentally compared with single-channel models such as the SC-BiTCN-BiGRU model, the SC-TCN-GRU model, and the common SC-CNN-LSTM model. Among them, the SC-BiTCN-BiGRU model has the same structure as the MC-BiTCN-BiGRU model, only the input method is different. The SC-TCN-GRU model uses a unidirectional feature extraction method, where TCN is responsible for extracting local features of the time series, and GRU is used to process time dependence. Compared with BiTCN-BiGRU, SC-TCN-GRU only extracts forward features, while BiTCN can capture both forward and backward features, improving the ability to capture long-term dependence relationships in the sequence. The SC-CNN-LSTM model, as a common RUL prediction model, uses CNN to extract local spatial features, and LSTM processes the time series. Compared with GRU, LSTM introduces an additional gating mechanism when dealing with long-term dependence, but usually has a higher computational complexity. The prediction results of the above single-channel models are summarized in Table 2.

[0080] Table 2 Comparison Table of Single-Channel Model Prediction Results

[0081] The experimental results show that the MC-BiTCN-BiGRU-EAA model proposed in this application performs optimally among all the comparison models. In terms of MAE, it has improved by 55.87% compared to the SC-BiTCN-BiGRU, 61.36% compared to the SC-TCN-GRU, and 57.29% compared to the SC-CNN-LSTM. At the same time, we also observe that in the single-channel model, the prediction results of the BiTCN-BiGRU model are better than those of the TCN-GRU model, indicating that bidirectional feature extraction has significant advantages in improving prediction accuracy. The prediction results of each model for the remaining useful life (RUL) of lithium batteries are as Figure 9 and Figure 10 shown, where Figure 9 is the comparison chart of the single-channel model prediction results of lithium battery B0007 provided by the embodiment of this application, Figure 10 is the comparison chart of the single-channel model prediction results of lithium battery B0018 provided by the embodiment of this application. The results clearly show that the MC-BiTCN-BiGRU model performs best in predicting the remaining useful life of lithium batteries, especially on the dataset of B007 battery, where its prediction effect is particularly prominent. This phenomenon can be attributed to the fact that the B007 battery dataset contains 160 cycles, which is larger in scale than the B018 battery dataset, thus providing more abundant feature information and making the prediction results more accurate.

[0082] In an embodiment of this application, in order to explore the performance of the multi-channel model in predicting the remaining useful life (RUL), the MC-BiTCN-BiGRU-EAA model proposed in this application is compared with the MC-TCN-GRU-EAA and MC-CNN-LSTM-EAA models. Among them, MC-TCN-GRU extracts the time series features of the charging curve through TCN, and GRU processes the discharge capacity data; MC-CNN-LSTM extracts the local features in the charging curve through CNN, and LSTM is used to capture the long-term dependence of the discharge capacity. The experimental results are summarized in Table 3, showing the performance differences of each model on different battery datasets.

[0083] Table 3 Comparison Table of Multi-Channel Model Prediction Results

[0084] It can be seen that MC-BiTCN-BiGRU is significantly superior to MC-TCN-GRU and MC-CNN-LSTM models in terms of prediction accuracy. For example, in terms of MAE, the average error of MC-BiTCN-BiGRU is 1.03%. Among them, the MAE of B007 battery is 0.76%, which is 41.98% higher than that of the MC-TCN-GRU model and 50% higher than that of the MC-CNN-LSTM model. The MAE of B018 battery is 1.23%, which is 23.13% higher than that of MC-TCN-GRU and 38.5% higher than that of MC-CNN-LSTM respectively. In addition, Figure 11 and Figure 12 show the performance comparison of each multi-channel model in the prediction of lithium battery RUL, where Figure 11 is the comparison chart of the prediction results of the multi-channel model of lithium battery B0007 provided by the embodiment of the present application, Figure 12 is the comparison chart of the prediction results of the multi-channel model of lithium battery B0018 provided by the embodiment of the present application. It can be observed from the figure that the MC-CNN-LSTM model performs poorly in peak prediction, while the structure of the TCN network effectively avoids overfitting and underfitting phenomena through its causal convolution and skip connections, so it performs more accurately in the prediction of data turning points. The bidirectional feature extraction mechanism (BiTCN) can not only consider past and future information simultaneously, but also enhance the model's perception ability of dynamic changes in battery data. Combining these advantages, the prediction results of the MC-BiTCN-BiGRU model are closer to the true values, especially when dealing with the capacity regeneration phenomenon, showing stronger robustness and stability.

[0085] A method for predicting the remaining useful life of a lithium-ion battery based on a fusion model proposed in the present application. The method includes data preprocessing, RUL prediction model, hyperparameter optimization and prediction evaluation. Among them, data preprocessing adopts feature selection, dimensionality reduction, and scaling methods to efficiently mine feature parameters and achieve parameter reduction and regularization by analyzing the internal laws of the collected data, so that the preprocessed data can be used as reliable input variables for machine learning models; RUL prediction adopts a multi-channel parallel input method, and inputs battery data into the fusion model MC-BiTCN-BiGRU-EAA, making full use of battery charge and discharge data and bidirectionally extracting battery information to make up for the problem of insufficient data volume; multi-channel parallel input maximizes the advantages of a single model and then fuses the models, avoiding the occurrence of gradient disappearance and gradient explosion phenomena and reducing the computational cost; hyperparameter optimization further improves the accuracy and computational efficiency of battery life prediction.

[0086] 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 foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features, and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting the remaining useful life of a lithium-ion battery based on a fusion model, characterized in that, Including: S1: Extract the charge and discharge data during all charge and discharge cycles in the lithium-ion battery detection dataset publicly available from NASA; S2: Perform feature extraction, feature dimensionality reduction, and feature scaling on the extracted charge and discharge data, and form the initial dataset with the extracted charge and discharge features; S3: Divide the initial dataset into a training set and a test set, and build an MC-BiTCN-BiGRU-EAA fusion model based on the training set as the lithium battery RUL prediction model; S4: Use the TPE algorithm to obtain the hyperparameters of the lithium battery RUL prediction model, and optimize based on the hyperparameters to obtain the optimized lithium battery RUL prediction model; S5: Predict the remaining useful life of the lithium-ion battery according to the lithium battery RUL prediction model and the test set; S6: Evaluate the prediction accuracy according to the predicted remaining useful life of the lithium-ion battery using the mean absolute error, mean square error, mean absolute percentage error, and root mean square error.

2. The method for predicting the remaining service life of a lithium-ion battery based on a fusion model according to claim 1, wherein The specific content of S2 includes: S201: Perform feature extraction on the extracted charge and discharge data based on the Pearson correlation coefficient method to obtain the input features of the lithium battery RUL prediction model. The formula of the Pearson correlation coefficient method is: Among them, is the value of the x variable in the sample, is the average value of the values of the x variable, is the value of the y variable in the sample, is the average value of the values of the y variable, is the total number of samples, is the th sample, is the correlation between variables; S202: Perform principal component analysis on the input features of the lithium battery RUL prediction model to obtain the input features after dimensionality reduction; S203: Use the maximum-minimum scaling method to normalize all features in the lithium-ion battery detection dataset to obtain the initial dataset. The formula of the maximum-minimum scaling method is: Among them, is the value after the extracted feature is normalized, is the value of the extracted feature, is the minimum value of the extracted feature, is the maximum value of the extracted feature.

3. The method for predicting the remaining service life of a lithium-ion battery based on a fusion model according to claim 2, wherein The principal component analysis in S202 specifically includes: A1: Set the D-dimensional training set X representing the original data matrix, and calculate the covariance matrix of X: Among them, , is the number of battery charge and discharge cycles, is the th sample vector with a dimension of , is the mean vector with a dimension of ; A2: Use the non-linear iterative partial least squares or singular value decomposition method to solve the eigenvalues and eigenvectors of the covariance matrix; A3: Sort the eigenvalues in descending order, select the top d ones, project the original features onto the selected eigenvectors, and obtain the new d -dimensional input features after dimensionality reduction.

4. The method for predicting the remaining service life of a lithium-ion battery based on a fusion model according to claim 1, wherein The specific content of building the MC-BiTCN-BiGRU-EAA fusion model in S3 includes: S301: Use a single-layer BiTCN model to extract the multi-channel charging data features in the training set, and use the extracted charging current, charging voltage, and charging temperature features as the inputs of one layer of BiTCN respectively; S302: Connect the extracted multi-channel charging data features through a connection layer, input them into a multi-layer BiTCN model with a residual connection, and embed an additive attention mechanism layer into the BiTCN model to output charging features; S303: Input the discharge capacity data in the training set into the BiGRU model to output discharge features; S304: Connect the output charging features of the multi-layer BiTCN model and the output discharge features of the BiGRU model through a fully connected layer to obtain the MC-BiTCN-BiGRU-EAA fusion model.

5. The method for predicting the remaining service life of a lithium-ion battery based on a fusion model according to claim 4, wherein, The BiTCN model consists of two parts: dilated convolution and causal convolution, including an input layer, two hidden layers, and an output layer. The formula of the dilated convolution is: Among them, represents the output value of the dilated convolution at time step . is the convolution kernel size, represents the weight parameter at the th position in the product kernel, represents the input value at time step , represents the dilation factor.

6. The method for predicting the remaining service life of a lithium-ion battery based on a fusion model according to claim 4, characterized in that, The residual connection calculation formula of the multi-layer BiTCN model with a residual connection is: Among them, represents the convolution input, represents the convolution output, represents the output of the residual connection, is the activation function.

7. The method for predicting the remaining service life of a lithium-ion battery based on a fusion model according to claim 4, wherein The efficient additive attention in the additive attention mechanism layer has an output of: Among them, represents a normalized query matrix, represents a linear transformation, represents a single global query vector, represents a key matrix, is efficient additive attention.

8. The method for predicting the remaining service life of a lithium-ion battery based on a fusion model according to claim 1, wherein The specific content of S4 includes: S401: Determine whether the hyperparameter initialization of the lithium battery RUL prediction model has been completed. If so, update the hyperparameter combination based on the historical training records and proceed to S402; if not, randomly generate an initial hyperparameter combination within the preset search space and proceed to S402; S402: Based on the historical hyperparameter combinations and their corresponding performance metrics, the hyperparameters are divided into two categories: the set of the top 20% hyperparameters with the optimal performance , and the set of the remaining ordinary hyperparameters ; S403: Use Gaussian process regression to construct probability density models for two types of hyperparameters respectively: excellent distributions characterizing the statistical characteristics , and ordinary distributions characterizing the statistical characteristics , where refers to the hyperparameter combination to be optimized; S404: Define the expected improvement function as , and select a better hyperparameter combination by maximizing the value of the expected improvement function ; S405: Based on the new hyperparameter combination Train the RUL prediction model of the lithium battery and calculate the prediction error of the training set; S406: Determine whether the prediction error meets the error minimization. If so, proceed to S409; if not, proceed to S407; S407: Determine whether the calculation has reached the maximum number of iterations or the calculation time limit. If so, proceed to S409; if not, proceed to S408; S408: Dynamically shrink the search space according to the high-probability region during the iteration process, re-partition I(x) I(x) after each iteration, and and , and update I(x) the distribution parameters of G(x) and enter S404; S409: Output the hyperparameter combination with the smallest training set error , and based on the hyperparameter combination obtain the optimized RUL prediction model configuration.

9. The method for predicting the remaining service life of a lithium-ion battery based on a fusion model according to claim 1, wherein The calculation formulas for the mean absolute error, mean square error, mean absolute percentage error, and root mean square error in S6 are as follows: Wherein, is the original battery capacity value, is the predicted value, is the mean absolute error, is the mean square error, is the mean absolute percentage error, is the root mean square error, represents the number of samples, is the th sample.

Images