Lithium battery life prediction method based on modal decomposition and Informer-LSTM

Abstract

A lithium battery life prediction method based on modal decomposition and Informer-LSTM belongs to the field of battery life prediction, and adopts a CEEMDAN method to decompose a battery capacity data sequence into a plurality of intrinsic modal components serving as high-frequency components and a residual component serving as a low-frequency component, so as to reduce the influence caused by the irregular recovery phenomenon of the battery capacity; inputting the low-frequency component into the LSTM network model to obtain a low-frequency component prediction result; a high-frequency component is secondarily decomposed into a trend term and a residual term by adopting a Dlinear method, and the fluctuation degree of the high-frequency component is reduced while key information is reserved, so that the prediction stability is improved. Respectively inputting a trend term and a residual term decomposed by each high-frequency component into an Informer network model for prediction, then superposing results, extracting implicit features in the complex fluctuation data, and obtaining a prediction result of the high-frequency component; and carrying out weighted fusion on the low-frequency component prediction result and the prediction results of the plurality of high-frequency components to obtain a final prediction result.

Description

Technical Field

[0001] The present invention relates to a battery life prediction method, and specifically to a lithium battery life prediction method based on modal decomposition and Informer-LSTM. Background Art

[0002] Lithium batteries, with their high energy density, long cycle life, and environmental friendliness, are widely used in energy storage devices, electric vehicles, and aerospace. However, as the number of charge and discharge cycles increases, their performance gradually deteriorates. This is manifested as a decrease in the battery's usable capacity, an inability to deliver the rated power output, and ultimately, failure. Failed lithium batteries, if not promptly replaced, can cause power system failure or damage. Therefore, predicting the lifespan of lithium batteries is necessary to ensure safe equipment operation.

[0003] With the continuous development of machine learning and artificial intelligence, data-driven life prediction methods are becoming mainstream. This approach uses machine learning or deep learning models to extract data degradation characteristics from historical data, thereby establishing a mapping relationship between this data and remaining life indicators. However, battery data contains not only low-frequency features that reflect the long-term degradation trend of battery capacity, but also high-frequency features with local oscillation characteristics, such as brief capacity rebounds caused by factors such as battery material aging, SEI film thickening, and active material loss. A single model cannot account for these different frequency characteristics, resulting in large prediction errors. Currently used methods include decomposing battery data into high-frequency and low-frequency components through empirical mode decomposition and applying different models to each component, which has achieved some improvement in prediction accuracy. However, due to the strong and frequent oscillations of high-frequency components, directly predicting these high-frequency components often fails to accurately capture their dynamic characteristics, resulting in large prediction errors. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to overcome the defect of large error in lithium battery life prediction and provide a lithium battery life prediction method based on modal decomposition and Informer-LSTM.

[0005] The technical solution adopted by the present invention to solve the above technical problems is: a lithium battery life prediction method based on modal decomposition and Informer-LSTM, which uses the CEEMDAN method to decompose the battery capacity data sequence into multiple intrinsic modal components as high-frequency components and a residual component as a low-frequency component, uses different neural network models to extract the features of different components respectively and perform predictions, and then fuses the prediction results to obtain the final prediction result, and inputs the low-frequency component into the LSTM network model to obtain the low-frequency component prediction result; uses the Dlinear method to secondary decompose the high-frequency component into trend terms and residual terms, and inputs the trend terms and residual terms of each high-frequency component decomposition into the Informer network model for prediction, and then superimposes the results to obtain the prediction result of the high-frequency component; then, the low-frequency component prediction result and the prediction results of multiple high-frequency components are weighted and fused to obtain the final prediction result.

[0006] After decomposing the battery capacity data sequence into multiple intrinsic mode components and one residual component, the Pearson correlation coefficient between each component and the original battery capacity data sequence is calculated and used as the weight of different components. The weight is used to weightedly fuse the prediction results of the low-frequency component and the prediction results of multiple high-frequency components.

[0007] The formula for the Pearson correlation coefficient is: Where r is the correlation coefficient obtained, 、 For two sample variables to calculate the correlation, 、 They are 、 The sample mean of .

[0008] The weight of the residual component was set to 1, and the weights of the other components were calculated in proportion to the obtained Pearson correlation coefficients.

[0009] The Dlinear method is used to decompose the high-frequency component into trend terms and residual terms. The decomposition formula is: in, is the trend item, is the residual term, is the input sequence, k is the sequence length, and i is the time position.

[0010] By performing secondary decomposition on the high-frequency components, the sharp fluctuations in the high-frequency components can be reduced while retaining key information.

[0011] The battery capacity data sequence is a curve data sequence that reflects the change of battery capacity with the number of cycles. The collected battery capacity data sequence data set is used to train the prediction model, and the trained prediction model is used to predict the future battery capacity curve based on the battery historical data.

[0012] The present invention has the following beneficial effects: Data is decomposed using the CEEMDAN decomposition method to reduce the impact of irregular battery capacity rebound. To address the strong volatility and local mutations of the decomposed high-frequency components, a secondary decomposition is performed on the high-frequency components to reduce their volatility and improve prediction stability. For low-frequency components with gentle fluctuations and strong trends, an LSTM model suitable for simple time series is selected for prediction. This prevents complex models from overfitting to stable trend signals and capturing irrelevant small fluctuations, thereby improving prediction accuracy and ensuring model training efficiency and generalization. For high-frequency components with strong volatility, a secondary decomposition is used to more effectively reduce the sharp fluctuations in the high-frequency components while retaining key information. The decomposed sub-items are predicted separately using the informer model, which helps improve the model's prediction accuracy for information at different frequencies. For different components, the Pearson correlation coefficient is used to calculate the importance coefficient of each component to the original sequence. This is used as a weight, and each component is weighted and fused during the reconstruction phase to obtain the final result. In this way, effective information can be fully utilized and noise interference can be reduced, effectively improving the accuracy, stability, and generalization of the prediction model. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 It is a schematic flow chart of the prediction method of the present invention.

[0014] Figure 2 It is a structural diagram of the prediction model of the present invention.

[0015] Figure 3 It is a schematic diagram of the model structure of high-frequency component prediction of the present invention.

[0016] Figure 4 These are the capacity degradation curves of four sets of data in the NASA battery dataset.

[0017] Figure 5 It is the capacity degradation curve of two sets of data in the CALCE battery dataset.

[0018] Figure 6 This is the CEEMDAN decomposition result of the B0005 battery capacity data in the NASA battery dataset.

[0019] Figure 7 It is the CEEMDAN decomposition result of the CS2_35 battery capacity data in the CALCE battery dataset.

[0020] Figure 8 This is a comparison chart of the predicted results of B0005 battery capacity data.

[0021] Figure 9 This is a comparison chart of the predicted results of B0006 battery capacity data.

[0022] Figure 10 It is the model prediction fitting curve of CS2_35 battery capacity data.

[0023] Figure 11 It is the model prediction fitting curve of CS2_36 battery capacity data.

[0024] Figure 12 This is the evaluation index diagram of CS2_35 and CS2_36. DETAILED DESCRIPTION

[0025] The following is a clear and complete description of the technical solution of the present invention in conjunction with the accompanying drawings and specific embodiments. The specific contents listed in the following embodiments are not limited to the technical features required to solve the technical problems to be solved by the technical solutions described in the claims. At the same time, the examples listed are only part of the present invention, not all embodiments.

[0026] The process of the lithium battery life prediction method of the present invention is as follows: Figure 1 As shown in the figure, the original battery capacity sequence is processed through a series of processes to obtain the predicted battery life. The battery capacity sequence is a degradation data sequence that reflects the change in battery capacity with the number of charge and discharge cycles. This data sequence represents the battery capacity degradation curve. The prediction model is trained using the collected battery capacity data sequence dataset. The trained model can be used to predict the future battery capacity curve based on historical battery data, thereby predicting the battery life.

[0027] The prediction model processes the battery capacity data series as follows: First, the battery capacity data series is decomposed using the CEEMDAN method into multiple intrinsic mode components (IMFs) and a residual component (RES). Based on the adaptive noise complete ensemble empirical mode decomposition (EMD), CEEMDAN introduces adaptive white noise during each decomposition and uses the average results of multiple decompositions to suppress the impact of noise on the decomposition, effectively alleviating the problems of modal aliasing and noise sensitivity. The main steps are as follows: (1) Add a white noise sequence that obeys the normal distribution to the original capacity sequence: in, is the original capacity sequence, is the sequence after adding white noise, is the signal-to-noise ratio, is a white noise sequence.

[0028] (2) For the sequence with white noise added Use EMD to decompose, then take the average to get IMF1 and residual sequence : (3) For the residual sequence Add white noise and perform EMD decomposition again to obtain IMF2 and residual sequence in, It is the signal-to-noise ratio that is suitable for the current stage after adjustment.

[0029] (4) Calculate the subsequent IMF according to the above process. When the decomposed residual sequence is a monotonic function, the iteration ends.

[0030] (5) After CEEMDAN decomposition, the original signal can be expressed as: Among the components decomposed by CEEMDAN, the IMF component mainly captures the local oscillation characteristics of the original signal and can reflect the high-frequency and medium-frequency changes of different scales in the data, such as the temporary capacity recovery caused by factors such as material aging, SEI film thickening, and active material loss. Due to its more intense and frequent oscillations, it is regarded as a high-frequency component. The final The residual component (RES) represents the low-frequency trend in the original signal that is difficult to further decompose. It mainly reflects the overall degradation trend of battery capacity with increasing cycle number. It changes slowly and has a single trend, so it is considered a low-frequency component. This separates short-term mutations from long-term trends and effectively reduces the impact of capacity recovery on the overall prediction task.

[0031] In terms of processing low-frequency components, due to their gentle changes and strong trend, a simple model can be used to obtain good prediction results. If a complex model is used, it may cause the model to overfit the stable trend signal, capture irrelevant small fluctuations, and reduce the predictive ability of the model. At the same time, the long training time and high resource consumption of complex models can easily lead to a waste of computing resources. Therefore, the present invention selects an LSTM network model suitable for simple time series to predict the low-frequency components, inputs the low-frequency components into the LSTM network model to obtain the low-frequency component prediction results, which not only improves the prediction accuracy, but also ensures the training efficiency and generalization of the model.

[0032] Since high-frequency components oscillate violently and frequently, it is often difficult for the model to accurately capture their dynamic characteristics when directly predicting these high-frequency components, and the prediction error is large. This invention uses the high-frequency component secondary decomposition combined with the Informer network model for prediction. Figure 1 The D-Informer is represented in [1]. The secondary decomposition of high-frequency components uses the Dlinear method, decomposing each high-frequency component into a trend term and a remainder term. This method can reduce the sharp fluctuations and noise interference in high-frequency components while preserving the data's information integrity and improving the model's information utilization. This method can further extract the potential information within the high-frequency components, thereby improving the forecasting model's ability to predict complex nonlinear fluctuations and ultimately its accuracy.

[0033] The Dlinear decomposition formula is: in, is the trend item, is the residual term, is the input sequence, k is the sequence length, and i is the time position.

[0034] Using the informer network model to separately model and predict the trend term and residual term derived from the secondary decomposition of each high-frequency component helps improve the model's prediction accuracy for information at different frequencies. The trend term and residual term are each fed into the informer network model, extracting implicit features from the sequence and generating predictions for the two sub-terms. These predictions are then combined to produce the prediction for the high-frequency component.

[0035] The Informer network model is based on the transformer by introducing a probabilistic sparse self-attention mechanism to reduce computational complexity, and uses a self-attention distillation mechanism to halve the length of the original sequence to reduce the amount of computation and improve efficiency, making it more suitable for application in long time series. The Informer network model includes an encoder and a decoder. Its encoder consists of two stacks, each stack containing a self-attention layer and a distillation layer. The complete sequence is stacked in sequence, and the amount of computation is gradually reduced by selectively screening data. Finally, all the stacked feature maps are spliced ​​and input into the decoder. Informer uses a sparse probabilistic self-attention mechanism. For each query, only the dot products with the most important u are calculated and softmax normalized, which can reduce computational complexity. The calculation formula is as follows: Among them, Q, K, V are input feature matrices, and d is the vector dimension; is a sparse matrix of the same size as Q, which contains only the most important u queries and the rest are filled with 0.

[0036] The self-attention distillation mechanism optimizes the important information in the feature map and reduces redundant features through operations such as convolution, activation, and pooling, thereby reducing computational complexity and memory usage. The process of the distillation mechanism from layer j to layer j+1 is as follows: in, Represents the attention module; Conv1d uses the ELU activation function and a kernel of size 3 to perform a one-dimensional convolution operation on the time dimension; finally, a maximum pooling layer MaxPool with a step size of 2 is added to X t Downsample to half its size to reduce time complexity.

[0037] The decoder of Informer consists of two layers of multi-head self-attention and uses generative prediction results. The decoder input format is as follows: in, represents the representation vector of the token sequence processed by the decoder at time step t, is a placeholder for the target sequence and is set to 0. It is the real number field to which the vector after Concat belongs.

[0038] Each high-frequency component is decomposed twice and predicted by the Informer model to obtain the prediction results of multiple high-frequency components.

[0039] For different components, the amount of information they contain and the degree of importance to the original data are different. If the prediction results of different components are directly superimposed, it is easy for key information to be masked by noise, or the noise components are excessively introduced, thereby weakening the prediction ability of the model and affecting the prediction accuracy and stability of the model. The present invention uses a weighted fusion method to reconstruct the prediction results of low-frequency components and multiple high-frequency components to obtain the final prediction results. After the battery capacity data sequence is decomposed using the CEEMDAN method, the Pearson correlation coefficient of the multiple intrinsic mode components and one residual component obtained by decomposition and the original battery capacity data sequence is calculated respectively.

[0040] The formula for the Pearson correlation coefficient is: Where r is the correlation coefficient obtained, 、 For two sample variables to calculate the correlation, 、 They are 、 The sample mean of .

[0041] The Pearson correlation coefficient reflects the importance of each component to the original sequence and can be used as a weight for different components. Compared to random forests, the Pearson correlation coefficient is independent of the training model and is more suitable for small sample analysis. It preserves the importance of different components and helps improve the accuracy of the final fusion result. Using this weight to weight the low-frequency component prediction results and the prediction results of multiple high-frequency components to obtain the final prediction value can fully utilize effective information and reduce noise interference, thereby effectively improving the accuracy, stability, and generalization ability of the prediction model.

[0042] Four lithium battery datasets, B0005, B0006, B0007, and B0018, from the NASA lithium battery public dataset were selected as experimental datasets. These datasets include charge, discharge, and impedance information at room temperature (24°C). During charging, the batteries were charged at a constant current (CC) of 1.5A until the voltage reached 4.2V; charging then continued at a constant voltage (CV) of 4.2V until the charge current dropped to 20mA. During discharge, the batteries were discharged at a constant current (CC) of 2A until the voltages of the four battery groups dropped to 2.7V, 2.5V, 2.2V, and 2.5V, respectively. The battery life was considered terminated when its capacity dropped to 70% of its initial capacity, or 1.4Ah.

[0043] The degradation curves of the capacity of the four groups of batteries with the number of cycles are as follows Figure 4As shown in Figure 2, the battery capacity shows an overall downward trend, but there are different degrees of capacity recovery in the process. CEEMDAN is used to decompose the original capacity data. Figure 6 The CEEMDAN decomposition results for the B0005 battery data are shown. The B0005 battery capacity data is decomposed into four IMF components. IMF1-IMF3 exhibit significant oscillations and are high-frequency components. IMF4 is the final residual component (RES), which is automatically named the last IMF by the system code. IMF4 exhibits a clear and relatively stable trend and is a low-frequency component.

[0044] The data is decomposed into multiple IMF components using CEEMDAN. Different IMF components represent different frequency components of the signal and have varying degrees of influence on the original capacity data. Compared to simply summing the IMFs, this method can enhance specific frequency components, resulting in more accurate predictions. Therefore, the Pearson coefficient is used to calculate the relationship between each IMF component and the original capacity data. This relationship coefficient is used as a weight, and weighted fusion is performed after the model prediction is completed. Because IMF4 is a trend term, its weight is set to 1. The remaining weights are then calculated proportionally. The weights calculated for the NASA dataset are shown in Table 1.

[0045] To verify the effectiveness of the present invention, four sets of battery capacity data from the NASA battery data set were predicted using different models, and the prediction results were compared.

[0046] The model used is: LSTM: Use LSTM model for prediction.

[0047] Informer: Use the Informer model for prediction.

[0048] CEEMDAN-LSTM: The battery capacity data sequence is decomposed using CEEMDAN and then predicted using the LSTM model.

[0049] CEEMDAN-Informer: The battery capacity data series is decomposed using CEEMDAN and then predicted using the Informer model.

[0050] CEEMDAN-Informer-LSTM: After decomposing the battery capacity data sequence using CEEMDAN, the high-frequency components are predicted using the Informer model, and the low-frequency components are predicted using the LSTM model.

[0051] CEEMDAN-DInformer-LSTM: Using the method of the present invention, the battery capacity data sequence is decomposed by CEEMDAN, the high-frequency components are secondary decomposed and predicted by Informer, and the low-frequency components are predicted by the LSTM model.

[0052] The four evaluation indicators of each model are calculated and the results are shown in Table 2.

[0053] Among them, MAE is the mean absolute error, MAPE is the mean absolute percentage error, RMSE is the root mean square error, R 2 is the determination coefficient. From the overall results, the performance of the single model LSTM and Informer on different batteries fluctuates to a certain extent. The Informer model outperforms the LSTM model in most samples due to its modeling advantage in long time series. However, due to the capacity recovery phenomenon, the data fluctuates irregularly. These fluctuations make it impossible for a single model to achieve more accurate predictions. After introducing CEEMDAN to decompose the data, the overall performance of the model has also been greatly improved compared with the single model. Taking B0005 battery as an example, compared with the single LSTM model, the mean absolute error and mean absolute percentage error of CEEMDAN-LSTM are significantly reduced, the root mean square error is reduced by about 36%, and R 2 There has also been some improvement; compared with CEEMDAN-LSTM, CEEMDAN-Informer has a better R 2 Both are above 0.97, showing stronger fitting and prediction capabilities. However, the single network model still has the problem of insufficient feature extraction, and the accuracy that can be improved is limited. CEEMDAN-Informer-LSTM uses Informer and LSTM to process high-frequency components and low-frequency components respectively. The mean absolute error and root mean square error are both reduced compared to the single model CEEMDAN-Informer and CEEMDAN-LSTM. 2 The CEEMDAN-DInformer-LSTM model of the present invention achieved the best results in all indicators, and the goodness of fit was above 0.98, which was 1% to 2% higher than the other models. Among them, in the B0007 sample, its MAE was 0.0147, MAPE was only 0.0101, and R 2 As high as 0.9903, it reflects the comprehensive advantages of the composite model in error control, trend fitting and generalization ability.

[0054] In order to more intuitively reflect the pros and cons of the prediction results, take B0005 and B0006 batteries as examples. Figure 8 and Figure 9The comparison of the fitting effects of different prediction models on the battery capacity degradation process is demonstrated. In the figure, Origin is the original curve. It can be seen from the figure that all models can fit the early capacity decline trend of the battery well in the early stage of prediction, but the performance differences of the models gradually emerge in the middle and late stages. The predictions of the other models have large errors when approaching the failure threshold, and have obvious prediction lags; and the prediction results still have certain fluctuations, indicating that their ability to model the nonlinear characteristics of the battery capacity at the end of the decline is still insufficient, which may lead to unstable life prediction in practical applications and increase the risk of system operation. In contrast, the CEEMDAN-DInformer-LSTM model of the present invention enhances the modeling ability of the model for complex non-stationary signals by different processing of high-frequency components and low-frequency components and subsequent weighted fusion. From the perspective of evaluation indicators and the degree of fitting of the prediction curve, it is better than the other models, effectively improving the reliability of the remaining life prediction.

[0055] To further demonstrate the effectiveness of the proposed model, the experimental results of this model are compared with the recently proposed improved models HTS-GRU and CEEMDAN-BiLSTM-Transformer on the NASA dataset. The results are shown in Table 3: In the four battery data sets, the CEEMDAN-DInformer-LSTM model achieved the best results in all four indicators. Compared with the HTS-GRU model, the CEEMDAN-DInformer-LSTM reduced the MAE by 56.1% and the RMSE by 55.3% on the B0005 battery. On the B0007, the RMSE dropped from 0.0541 to 0.0174, a 67.8% decrease in error. On the B0006, compared with the CEEMDAN-BiLSTM-Transformer, the MAE dropped from 0.0244 to 0.0172, the RMSE dropped from 0.0296 to 0.0233, and the R 2 The performance of B0018 is also stable, with MAE and RMSE reduced by 19.4% and 20.2% respectively. 2 The prediction accuracy of CEEMDAN-DInformer-LSTM was improved from 0.9873 to 0.9908. CEEMDAN-DInformer-LSTM maintained high prediction accuracy on different batteries, proving that the proposed model has good performance in lithium battery life prediction.

[0056] To verify the generalization ability of the proposed model under diverse data conditions, batteries CS2_25 and CS2_26 from the lithium battery charge-discharge cycle life test dataset CALCE were selected. The battery operating temperature was set to a constant temperature. The charging process first used constant current (CC) mode until the battery voltage reached 4.2V, and then used constant voltage (CV) mode until the charging current dropped to 20mA. The discharge process used constant current (CC) mode until the battery voltage dropped to 2.7V. The battery life was considered to have ended when the battery capacity reached 0.77Ah, 70% of the rated capacity (1.1Ah).

[0057] The degradation curves of the two battery capacities with the number of cycles are as follows: Figure 5 As shown. Using CEEMDAN to decompose the original capacity data, the CEEMDAN decomposition results of the CS2_35 battery are shown as follows. Figure 7 The CS2_35 battery is decomposed into 7 components, of which IMF1-IMF6 are high-frequency components and IMF7 is a trend term, i.e., a low-frequency component.

[0058] The CS2_35 and CS2_36 batteries in the CALCE data set were selected. The first 50% of the data was divided into a training set and the second 50% of the data was divided into a prediction set. The model proposed in this invention was used for training. The prediction fitting curves of the two groups of batteries are shown in Figure 2. Figure 10 and 11 As shown in the figure, the evaluation indicators of the prediction results are as follows: Figure 12 As shown. From the results, the model performed well in various indicators on CS2_35, with MAE, MAPE and RMSE being 0.0120, 0.0191 and 0.0187 respectively, indicating that the model has a small error on this battery sample; the fitting coefficient is 0.9923, indicating that the model has strong fitting ability and can more accurately reflect the capacity degradation trend. On the CS2_36 battery, the model performed even better, with MAE, MAPE and RMSE being 0.0092, 0.0203 and 0.0160 respectively, and the fitting degree reached 99.62%. Although there are differences in the operating conditions of the batteries, the model still shows good adaptability and stability on different battery data, reflecting strong generalization ability. This result also shows that the proposed model can effectively cope with longer sequences and achieve good results in long sequence prediction.

[0059] Experimental results show that the model proposed in this paper obtains better results on different data sets, with an average goodness of fit of about 99%, effectively improving the prediction accuracy.

[0060] The above description of the specific embodiments is only used to help understand the technical concept and core ideas of the present invention. Although specific preferred embodiments are used herein to describe and illustrate the technical solutions, they should not be understood as limiting the present invention itself. Those skilled in the art may make various changes in form and details without departing from the technical concept of the present invention. These easily conceived changes or substitutions should all be included within the scope of protection of the present invention.

Claims

1. A lithium battery life prediction method based on modal decomposition and Informer-LSTM uses the CEEMDAN method to decompose the battery capacity data sequence into multiple intrinsic mode components as high-frequency components and a residual component as a low-frequency component. Different neural network models are used to extract the features of different components, perform predictions, and then fuse the prediction results to obtain the final prediction result. The following are the characteristics: The low-frequency component is input into the LSTM network model to obtain the low-frequency component prediction result; the high-frequency component is decomposed into trend terms and residual terms using the Dlinear method. The trend terms and residual terms of each high-frequency component are input into the Informer network model for prediction, and the results are superimposed to obtain the prediction result of the high-frequency component; Then, the low-frequency component prediction results and the prediction results of multiple high-frequency components are weighted and fused to obtain the final prediction result.

2. The lithium battery life prediction method based on modal decomposition and Informer-LSTM according to claim 1, characterized in that: After decomposing the battery capacity data sequence into multiple intrinsic mode components and one residual component, the Pearson correlation coefficient between each component and the original battery capacity data sequence is calculated and used as the weight of different components. The weight is used to weightedly fuse the prediction results of the low-frequency component and the prediction results of multiple high-frequency components.

3. The lithium battery life prediction method based on modal decomposition and Informer-LSTM according to claim 1, characterized in that: The formula for the Pearson correlation coefficient is: Where r is the correlation coefficient obtained, 、 For two sample variables to calculate the correlation, 、 They are 、 The sample mean of .

4. The lithium battery life prediction method based on modal decomposition and Informer-LSTM according to claim 2, characterized in that: The weight of the residual component was set to 1, and the weights of the other components were calculated in proportion to the obtained Pearson correlation coefficients.

5. The lithium battery life prediction method based on modal decomposition and Informer-LSTM according to claim 1, characterized in that: The Dlinear method is used to decompose the high-frequency component into trend terms and residual terms. The decomposition formula is: in, is the trend item, is the residual term, is the input sequence, k is the sequence length, and i is the time position.

6. The lithium battery life prediction method based on modal decomposition and Informer-LSTM according to claim 5, characterized in that: By performing secondary decomposition on the high-frequency components, the sharp fluctuations in the high-frequency components can be reduced while retaining key information.

7. The lithium battery life prediction method based on modal decomposition and Informer-LSTM according to claim 1, characterized in that: The battery capacity data sequence is a curve data sequence that reflects the change of battery capacity with the number of cycles. The collected battery capacity data sequence data set is used to train the prediction model, and the trained prediction model is used to predict the future battery capacity curve based on the battery historical data.

Images