Energy storage lithium battery life prediction method based on dice emd an and hybrid deep learning

Abstract

The application discloses a kind of energy storage lithium battery life prediction methods based on DICEEMDAN and hybrid deep learning, belong to data processing, lithium battery life prediction field, first, lithium battery capacity data is decomposed using DICEEMDAN and is reconstructed to process noise reduction, obtain multiple sub-sequences, the feature of high-frequency sub-sequence is extracted by improved bidirectional time series convolution network, different data characteristics existing for different frequency sub-sequences, a kind of different hybrid deep neural network is designed, the extracted feature data is used as the input of deep neural network, finally, the final prediction result is obtained by integrating multiple neural network prediction using the way of optimizing weight.The application is verified on actual lithium battery aging data set, and the experimental results show that the prediction of the remaining useful life of the energy storage lithium battery has high accuracy.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of research on the prediction of the remaining useful life of lithium batteries used in energy storage systems, and specifically relates to an energy storage lithium battery life prediction method based on dual improved complete ensemble empirical mode decomposition with adaptive noise (Dual Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise, DICEEMDAN) and hybrid deep learning. BACKGROUND

[0002] With the global energy structure transforming towards low carbonization, building a new type of power system has become the core task of energy strategy. New energy represented by wind and solar energy plays a key role in energy transformation, but its inherent intermittency and volatility characteristics may lead to imbalance between power supply and demand, threatening the stable operation of the power grid. To address this challenge, the development of efficient energy storage technology has become a top priority. Among them, lithium batteries have become the preferred solution in the global energy storage field due to their high energy density, low environmental pollution, lightweight, and excellent cycle performance. As the usage cycle of lithium batteries increases, its internal resistance increases, and the charging time is prolonged, exacerbating the complexity of power dispatching; as the usage time of lithium batteries increases, its safety performance decreases significantly. Therefore, accurately predicting the remaining useful life of lithium batteries plays a crucial role in the safe and stable operation of the entire power system.

[0003] The capacity data of lithium batteries has the characteristics of global nonlinearity, and there is a small amount of capacity regeneration phenomenon in the usage cycle of lithium batteries, so it is difficult to achieve high accuracy using original data for prediction, therefore more and more decomposition algorithms are used in the processing of capacity data to improve the accuracy of prediction, but the existing decomposition algorithms still have limitations such as mode mixing and endpoint effect in processing high-frequency noise, which cannot effectively extract the cause-and-effect relationship and internal time series characteristics in the data, resulting in ineffective improvement of prediction effect.

[0004] For high-order nonlinear, large-delay complex data systems, traditional prediction methods often have difficulty in accurate modeling. With the advancement of artificial intelligence technology, deep learning has shown significant potential in complex data prediction due to its powerful nonlinear fitting ability and adaptive feature extraction advantage. However, when a single neural network model is used to process such data with multiple complex characteristics, there are problems such as insufficient feature representation and loss of key information, which limits the prediction performance. SUMMARY

[0005] In order to overcome the shortcomings of existing prediction methods for the remaining service life of energy storage lithium batteries, such as poor feature extraction effect for complex systems and difficulty in achieving high-precision prediction, the present invention proposes a prediction method for the remaining service life of energy storage lithium batteries based on DICEEMDAN-BiTCN-DNN. First, the lithium battery capacity data is decomposed, denoised and reconstructed using DICEEMDAN to obtain multiple subsequences. The high-frequency subsequences are then feature extracted using an improved bidirectional temporal convolutional network (BiTCN). Based on the different data characteristics of subsequences with different frequencies, a differentiated hybrid deep neural network (DNN) is designed. The extracted feature data is used as the input of the deep neural network. Finally, the predictions of multiple neural networks are integrated by optimizing weights to obtain the final prediction result. The present invention has high accuracy in predicting the remaining service life of energy storage lithium batteries.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] A method for predicting the life of energy storage lithium batteries based on DICEEMDAN and hybrid deep learning, comprising the following steps:

[0008] S1, obtains the capacity data of the historical charge and discharge cycles of the energy storage lithium battery. The data contains capacity regeneration and global nonlinear characteristics. The double improved adaptive noise complete set empirical mode decomposition DICEEMDAN is used to reduce noise and reconstruct the data to obtain multiple characteristic subsequences;

[0009] S2, based on the different features of subsequences of different frequencies after decomposition, uses an improved bidirectional temporal convolutional network to extract features from high-frequency sequences, and recombines the extracted feature data with the low-frequency data in the original sequence as input data for the hybrid deep neural network;

[0010] S3, establishes a differentiated hybrid neural network. For high-frequency subsequences, a bidirectional long short-term memory network is used for prediction; for relatively low-frequency subsequences, a long short-term memory network is used for prediction. The final prediction result is obtained through an integration method with optimized weights.

[0011] Furthermore, in step S1, the data is subjected to denoising and reconstruction processing using the double-improved adaptive noise complete set empirical mode decomposition DICEEMDAN to obtain multiple feature subsequences, including the following sub-steps:

[0012] S1-1, obtain the capacity data of the energy storage lithium battery, use the adaptive noise complete set empirical mode decomposition on the data, and obtain the first decomposed modal componentIMF 1( t ) is then decomposed into M subband components { c j ( t )} M j=1 and residuals r ( t ):

[0013] (1);

[0014] S1-2, normalize the energy of the sub-band components, N Indicates the total number of component data for each subband:

[0015] (2);

[0016] Get the reconstructed high-frequency components:

[0017] (3);

[0018] S1-3, the first-order residual obtained after the adaptive noise complete set empirical mode decomposition e 1( t ), construct a sequence e ( t )= e 1( t )+ e 1 E 2( n i ), decompose it using the Empirical Mode Decomposition (EMD) algorithm, calculate the local mean, and take the average of the local means as the second-order residual e 2(t):

[0019] (4);

[0020] Where E2(·) is the operator of the second modal component obtained by EMD decomposition, e 1 is the expected signal-to-noise ratio when solving the second modal component, which is 0.2 times the standard deviation of the original signal; n i For the added i Gaussian white noise, M (⋅) represents the local mean after decomposition of the signal, m is the number of additions, and the second IMF Quantity IMF 2:

[0021] (5);

[0022] S1-4, repeat the same method as the above step S1-3 to decompose the k indivual IMF Quantity IMF k :

[0023] (6);

[0024] Where, e k-1 ( t ) indicates the k -1 order residual, e k ( t ) indicates the k Order residual, E k (·) is the first k modal component operators, e k-2 To solve the k -The expected signal-to-noise ratio when there is 1 modal component, e k-1 To solve the k The expected signal-to-noise ratio when the modal components are k =3,4,…, K ;

[0025] Finally, the capacity data of lithium batteries x ( t ) is decomposed into:

[0026] (7);

[0027] Where, K represents the final degree of the adaptive noise complete ensemble empirical mode decomposition, IMF k Indicates the first k modal components, e K ( t ) represents the trend component remaining after the final decomposition.

[0028] Furthermore, step S2 includes the following sub-steps:

[0029] S2-1: The input data is added to the Temporal Convolutional Network (TCN) input data to form a series of data connected end to end. To expand the receptive field of the convolution kernel and extract more sequence features, the network introduces a dilation factor that increases by multiples of two at each layer and uses dilated causal convolution to extract features from the expanded sequence, thereby effectively preserving the causal relationship of the time series.

[0030] S2-2, in order to ensure that the length of the sequence output by each convolution layer is consistent with the original sequence length, it is proposed to add zero padding items at the beginning and end of each sequence. The number of padding is p , the output of each layer is y ( t ), their mathematical expressions are:

[0031] (8);

[0032] Where, f ( i ) is the convolution kernel, x (t-i*d) is the first ( you * d ) values, t is the element index of the input data, i is the convolution kernel element index, k is the convolution kernel size, d is the expansion factor;

[0033] S2-3, as the receptive field continues to expand, in order to alleviate the problems of gradient disappearance and gradient explosion, residual blocks are added to the network, and two convolutions and one residual operation are added to the connection between layers.

[0034] Furthermore, in step S3, a differentiated hybrid neural network is established, trained using the divided data set, and an integration method for optimizing weights is proposed to obtain the final prediction result, which includes the following sub-steps:

[0035] S3-1. Decompose the training dataset into different batches based on the frequency characteristics of the data using the decomposition algorithm, and use different deep neural networks for prediction.

[0036] S3-2. For the modal component with a spectrum higher than one third of the sampling frequency, high-frequency data is recorded, and a bidirectional long short-term memory network is used. The network increases the reverse propagation path based on the traditional long short-term memory (Long Short-term Memory Networks, LSTM) network, so it can extract the forward and reverse features of the time series at the same time. The output calculation formula of each gate in the Bi-directional Long Short-Term Memory Networks (BiLSTM) is as follows:

[0037] (9);

[0038] wherein, i t d indicates the direction d up t the output of the input gate at time t, f t d indicates the direction d up t the output of the forget gate at time t, indicates the direction d up t the output of the candidate memory gate at time t, c t d indicates the direction d up t the output of the memory gate at time t, o t d indicates the direction d up t the output of the output gate at time t, h t d indicates the direction d up t the hidden state at time t, d indicates the direction of data transmission, forward or backward; x t is the input quantity at time step t , h t±1 d indicates the direction d up t the hidden state at time t±1, c t±1 d is the memory cell state in the direction d up t at time t±1, w xid It is the direction d The weight matrix of the upper input gate, w xf d It is the direction d The weight matrix of the upper forget gate, w xc d It is the direction d The weight matrix of the candidate memory gate, w xo d It is the direction d The weight matrix of the upper output gate, b i d It is the direction d The bias term of the upper input gate, b f d It is the direction d The bias term of the upper forget gate, b c d It is the direction d The bias term of the candidate memory gate, b o d It is the direction d The bias term of the upper output gate, ⊙, is Hadamard Product, σ is the sigmoid function, tanh is the hyperbolic tangent function, and the formulas are:

[0039] (10);

[0040] (11);

[0041] For data other than high-frequency data, traditional long short-term memory networks are used for prediction.

[0042] S3-3. This hierarchical processing method fully utilizes the short-term temporal dependencies of high-frequency sequences and the long-term trend characteristics of low-frequency sequences, thereby improving the overall prediction performance of the model. Finally, the prediction results of different networks are combined using a weighted combination based on the proportion of subsequences in the original sequence. The combination formula is as follows:

[0043] (12);

[0044] (13);

[0045] Where, w i t For the i The modal component is in thet The weight of the secondary cycle, IMF i t For the i The modal component is in the t The capacity value of the cycle, x t For the t Second cycle lithium battery capacity value, n is the number of test cycles, RUL The remaining service life of the lithium battery;

[0046] S3-4. The evaluation indicator used by the model is the mean absolute percentage error MAPE , root mean square error RMSE , mean absolute error MAE and absolute error AE , the calculation formula is:

[0047] (14);

[0048] (15);

[0049] (16);

[0050] (17);

[0051] Where, represents the predicted value, y i Indicates the actual value, n represents the predicted number of cycles, RUL re Indicates the number of cycles before the actual usage threshold is reached. RUL pr Indicates the number of cycles when the prediction reaches the usage threshold.

[0052] The beneficial effects of the present invention are mainly manifested in the following aspects: the present invention first decomposes the obtained energy storage lithium battery capacity data using DICEEMDAN, solves the modal aliasing problem of the traditional decomposition algorithm, and predicts the multiple decomposed sequences separately using hybrid deep learning. The hybrid deep learning mainly includes three modules: bidirectional temporal convolutional network, bidirectional long short-term memory network and long short-term memory network. The bidirectional temporal convolutional network adopts an expanded causal convolution structure. By expanding the receptive field of the convolution kernel and combining the bidirectional causal convolution mechanism, it can effectively extract the previous and next correlation information of high-frequency sequences. The extracted features are then input into the bidirectional long short-term memory network. For low-frequency sequences, a long short-term memory network is used for prediction. Experimental verification was carried out using real lithium battery life decay data. The results show that the invention has high accuracy in predicting the life of energy storage lithium batteries, and the model convergence time is shorter than other prediction algorithms. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 This is a flow chart of a method for predicting the life of an energy storage lithium battery based on DICEEMDAN and hybrid deep learning described in the present invention.

[0054] Figure 2 This is the effect diagram after using double-improved adaptive noise complete set empirical mode decomposition to process lithium battery capacity data.

[0055] Figure 3 This is a graph showing the predicted results of the invention for the remaining service life of a lithium battery.

[0056] Figure 4 It is a bar chart of various evaluation indicators of the model. DETAILED DESCRIPTION

[0057] The present invention will be further described below with reference to the accompanying drawings.

[0058] Reference Figure 1~Figure 4 , a method for predicting the life of energy storage lithium batteries based on DICEEMDAN and hybrid deep learning, comprising the following steps:

[0059] S1, obtains the capacity data of the historical charge and discharge cycles of the energy storage lithium battery. The data contains capacity regeneration and global nonlinear characteristics. Use DICEEMDAN to reduce noise and reconstruct the data to obtain multiple feature subsequences;

[0060] Use DICEEMDAN to perform noise reduction and reconstruction on the data. The results are as follows Figure 2 As shown, it includes the following sub-steps:

[0061] S1-1, obtain the capacity data of the energy storage lithium battery, use the adaptive noise complete set empirical mode decomposition on the data, and obtain the first decomposed modal component IMF 1(t ) is then decomposed into M subband components { c j ( t )} M j=1 and residuals r ( t ):

[0062] (1);

[0063] S1-2, normalize the energy of the sub-band components, N Indicates the total number of component data for each subband:

[0064] (2);

[0065] Get the reconstructed high-frequency components:

[0066] (3);

[0067] S1-3, the first-order residual obtained after the adaptive noise complete set empirical mode decomposition e 1( t ), construct a sequence e ( t )= e 1( t )+ e 1 E 2( n i ), decompose it using the empirical mode decomposition algorithm, calculate the local mean, and take the average of the local means as the second-order residual e 2(t):

[0068] (4);

[0069] Where, E 2(·) is the operator of the second modal component obtained by EMD decomposition, e 1 is the expected signal-to-noise ratio when solving the second modal component, which is 0.2 times the standard deviation of the original signal. n i For the added i Gaussian white noise, M (⋅) represents the local mean after decomposition of the signal, m is the number of additions, and the second IMF Quantity IMF 2:

[0070] (5);

[0071] S1-4, repeat the same method as the above step S1-3 to decompose the k indivual IMF Quantity IMF k :

[0072] (6);

[0073] Where, e k-1 ( t ) indicates the k -1 order residual, e k ( t ) indicates the k Order residual, E k (·) is the first k modal component operators, e k-2 To solve the k -The expected signal-to-noise ratio when there is 1 modal component, e k-1 To solve the k The expected signal-to-noise ratio when the modal components are k =3,4,…, K ;

[0074] The same method as above steps S1-3 is used, IMF 2 get IMF 3 , and then by IMF 3 get IMF 4 , and so on, until IMF k-1 get IMF k ;

[0075] Finally, the capacity data of lithium batteries x ( t ) is decomposed into:

[0076] (7);

[0077] Where, K represents the final degree of the adaptive noise complete ensemble empirical mode decomposition, IMF k Indicates the first k modal components, e K ( t ) represents the trend component remaining after the final decomposition.

[0078] S2, based on the different features of subsequences of different frequencies after decomposition, uses the improved bidirectional temporal convolutional network to extract features from high-frequency sequences, and recombines the extracted feature data with the low-frequency data in the original sequence as input data for the hybrid deep neural network;

[0079] In step S2, an improved bidirectional temporal convolutional network is used to extract features from data of different frequencies and divide the data set, which includes the following sub-steps:

[0080] S2-1: The input data is added to the temporal convolutional network input data to form a series of data connected end to end. To expand the receptive field of the convolution kernel and extract more sequence features, the network introduces a dilation factor that increases by multiples of two at each layer and uses dilated causal convolution to extract features from the expanded sequence, effectively preserving the causal relationship of the time series.

[0081] S2-2, in order to ensure that the length of the sequence output by each convolution layer is consistent with the original sequence length, it is proposed to add zero padding items at the beginning and end of each sequence. The number of padding is p , the output of each layer is y( t ), their mathematical expressions are:

[0082] (8);

[0083] Where, f ( i ) is the convolution kernel, x (t-i*d) is the first ( you * d ) values, t is the element index of the input data, i is the convolution kernel element index, k is the convolution kernel size, d is the expansion factor;

[0084] S2-3, as the receptive field continues to expand, in order to alleviate the problems of gradient disappearance and gradient explosion, residual blocks are added to the network, and two convolutions and one residual operation are added to the connection between layers.

[0085] S3, establish a differentiated hybrid neural network. For high-frequency subsequences, the extracted feature data is relatively complex, and a bidirectional long short-term memory network is used for prediction; for relatively low-frequency subsequences, due to their low complexity, a long short-term memory network is used for prediction. Finally, an integration method for optimizing weights is proposed to obtain the final prediction results as follows Figure 3 As shown in the figure, the evaluation indicators of the model are as follows Figure 4 As shown, it includes the following sub-steps:

[0086] S3-1, according to the decomposition algorithm, the different frequency characteristics of the data after the training data set are decomposed, divided into different batches, and different deep neural networks are used for prediction;

[0087] S3-2: For modal components whose spectrum is higher than one-third of the sampling frequency, they are recorded as high-frequency data. A bidirectional long short-term memory network is used. A backpropagation path is added to the traditional long short-term memory network, so that both forward and reverse features of the time series can be extracted simultaneously. The output calculation formula of each gate in the bidirectional long short-term memory network is as follows:

[0088] (9);

[0089] Where, i t d Indicates direction d superior t The output of the input gate at any moment, f t d Indicates direction d superior t Always forget the output of the gate, Indicates direction d superior t The output of the candidate memory gate at time, c t d Indicates direction d superior t The output of the memory gate at all times, o t d Indicates direction d superior t The output of the gate at all times, h t d Indicates direction d superior t The hidden state of the moment, d Indicates the direction of data transmission, forward or backward; x t is the time step t The input amount, h t±1 d Indicates direction d superior t The hidden state at ±1 time, c t±1 d It is the direction d superior t The memory cell state at ±1 moment, w xid It is the direction d The weight matrix of the upper input gate, w xf d It is the direction d The weight matrix of the upper forget gate, w xc d It is the direction d The weight matrix of the candidate memory gate, w xo d It is the direction d The weight matrix of the upper output gate, b i d It is the direction d The bias term of the upper input gate, b f d It is the direction d The bias term of the upper forget gate, b c d It is the direction d The bias term of the candidate memory gate, b o d It is the direction d The bias term of the upper output gate, ⊙, is Hadamard Product, σ is the sigmoid function, tanh is the hyperbolic tangent function, and the formulas are:

[0090] (10);

[0091] (11);

[0092] For data other than high-frequency data, traditional long-short-term memory networks are used for prediction;

[0093] S3-3, this hierarchical processing method fully utilizes the short-term temporal dependency of high-frequency sequences and the long-term trend characteristics of low-frequency sequences, thereby improving the overall prediction performance of the model; finally, according to the proportion of subsequences in the original sequence, the prediction results of different networks are combined using the proportion weighting method. The combination formula is as follows:

[0094] (12);

[0095] (13);

[0096] Where, w i t For the i The modal component is in thet The weight of the secondary cycle, IMF i t For the i The modal component is in the t The capacity value of the cycle, x t For the t Second cycle lithium battery capacity value, n is the number of test cycles, RUL The remaining service life of the lithium battery;

[0097] S3-4, the evaluation indicator used by the model is the mean absolute percentage error MAPE , root mean square error RMSE , mean absolute error MAE and absolute error AE , the calculation formula is:

[0098] (14);

[0099] (15);

[0100] (16);

[0101] (17);

[0102] Where, represents the predicted value, y i Indicates the actual value, n represents the predicted number of cycles, RUL re Indicates the number of cycles before the actual usage threshold is reached. RUL pr Indicates the number of cycles when the prediction reaches the usage threshold.

[0103] To enable researchers in this field to better understand the prediction process, Figure 1 The flowchart is intuitive and easy to understand.

[0104] The present invention first uses DICEEMDAN to decompose, de-noise and reconstruct the lithium battery capacity data to obtain multiple subsequences, and then uses an improved bidirectional temporal convolutional network to extract features from the high-frequency subsequences. In view of the different data characteristics of subsequences with different frequencies, a differentiated hybrid deep neural network is designed, and the extracted feature data is used as the input of the deep neural network. Finally, the multiple neural network predictions are integrated using the optimized weight method to obtain the final prediction result. The present invention has been verified on an actual lithium battery aging data set, and the experimental results show that it has a high accuracy in predicting the remaining service life of energy storage lithium batteries. Therefore, the present invention can accurately predict the remaining service life of energy storage lithium batteries and further improve the stability of the energy storage system operation. In summary, the method for predicting the remaining service life of energy storage lithium batteries based on DICEEMDAN-BiTCN-DNN established by the present invention provides a certain reference for predicting the remaining service life of lithium batteries in energy storage systems.

[0105] The embodiments of this specification are merely examples of implementations of the invention and are provided for illustrative purposes only. The scope of protection of the present invention should not be considered limited to the specific embodiments described in these embodiments. The scope of protection of the present invention also extends to equivalent technical means that can be conceived by a person of ordinary skill in the art based on the invention.

Claims

1. A method for predicting the life of energy storage lithium batteries based on DICEEMDAN and hybrid deep learning, characterized in that: The method comprises the following steps: S1, obtains the capacity data of the historical charge and discharge cycles of the energy storage lithium battery. The data contains capacity regeneration and global nonlinear characteristics. The double improved adaptive noise complete set empirical mode decomposition DICEEMDAN is used to reduce noise and reconstruct the data to obtain multiple characteristic subsequences; S2, based on the different features of subsequences of different frequencies after decomposition, uses an improved bidirectional temporal convolutional network to extract features from high-frequency sequences, and recombines the extracted feature data with the low-frequency data in the original sequence as input data for the hybrid deep neural network; S3, establishes a differentiated hybrid neural network. For high-frequency subsequences, a bidirectional long short-term memory network is used for prediction; for relatively low-frequency subsequences, a long short-term memory network is used for prediction. The final prediction result is obtained through an integration method with optimized weights.

2. The energy storage lithium battery life prediction method based on DICEEMDAN and hybrid deep learning as claimed in claim 1, characterized in that In step S1, the data is subjected to denoising and reconstruction processing using the double-improved adaptive noise complete ensemble empirical mode decomposition (DICEEMDAN) to obtain multiple feature subsequences, including the following sub-steps: S1-1, obtain the capacity data of the energy storage lithium battery, use the adaptive noise complete set empirical mode decomposition on the data, and obtain the first decomposed modal component IMF 1( t ) is then decomposed into M subband components { c j ( t )} M j=1 and residuals r ( t ): (1)； S1-2, normalize the energy of the sub-band components, N Indicates the total number of component data for each subband: (2)； Get the reconstructed high-frequency components: (3)； S1-3, the first-order residual obtained after the adaptive noise complete set empirical mode decomposition e 1( t ), construct a sequence e ( t )= e 1 ( t )+ ε 1 E 2( n i ), decompose it using the empirical mode decomposition algorithm EMD, calculate the local mean, and take the average of the local means as the second-order residual e 2(t): (4)； Where E2(·) is the operator of the second modal component obtained by EMD decomposition, ε 1 is the expected signal-to-noise ratio when solving the second modal component, which is 0.2 times the standard deviation of the original signal. n i For the added i Gaussian white noise, M (⋅) represents the local mean after decomposition of the signal, m is the number of additions, and the second IMF Quantity IMF 2: (5)； S1-4, repeat the same method as the above step S1-3 to decompose the k indivual IMF Quantity IMF k : (6)； Where, e k-1 ( t ) indicates the k -1 order residual, e k ( t ) indicates the k Order residual, E k (·) is the first k modal component operators, ε k-2 To solve the k -The expected signal-to-noise ratio when there is 1 modal component, ε k-1 To solve the k The expected signal-to-noise ratio when the modal components are k =3,4,…, K ; Finally, the capacity data of lithium batteries x ( t ) is decomposed into: (7)； Where, K represents the final degree of the adaptive noise complete ensemble empirical mode decomposition, IMF k Indicates the first k modal components, e K ( t ) represents the trend component remaining in the final decomposition.

3. The energy storage lithium battery life prediction method based on DICEEMDAN and hybrid deep learning according to claim 1 or 2, characterized in that The step S2 includes the following sub-steps: S2-1, based on the input data of the temporal convolutional network (TCN), the input data is added again to form a series of data connected end to end. In order to expand the receptive field of the convolution kernel and extract more sequence features, the network introduces a dilation factor that increases by multiples of two at each layer, and uses dilated causal convolution to extract features from the expanded sequence, thereby effectively preserving the causal relationship of the time series; S2-2, in order to ensure that the length of the sequence output by each convolution layer is consistent with the original sequence length, it is proposed to add zero padding items at the beginning and end of each sequence. The number of padding is p , the output of each layer is y ( t ), their mathematical expressions are: (8)； Where, f ( i ) is the convolution kernel, x (t-i*d) is the first ( ti * d ) values, t is the element index of the input data, i is the convolution kernel element index, k is the convolution kernel size, d is the expansion factor; S2-3. As the receptive field continues to expand, to alleviate the problems of gradient vanishing and gradient exploding, residual blocks are added to the network, and two convolutions and one residual operation are added to the connection between layers.

4. The energy storage lithium battery life prediction method based on DICEEMDAN and hybrid deep learning as claimed in claim 3, characterized in that In step S3, a differentiated hybrid neural network is established, and an integration method for optimizing weights is proposed to obtain the final prediction result, which includes the following sub-steps: S3-1, according to the decomposition algorithm, the different frequency characteristics of the data after the training data set are decomposed, divided into different batches, and different deep neural networks are used for prediction; In S3-2, modal components whose spectrum is higher than one-third of the sampling frequency are recorded as high-frequency data. A bidirectional long short-term memory (BiLSTM) network is used. This adds a backpropagation path to the traditional unidirectional long short-term memory (LSTM) network, thereby simultaneously extracting both forward and reverse features of the time series. The output calculation formulas for each gate in the BiLSTM are as follows: (9)； Where, i t d Indicates direction d superior t The output of the input gate at any moment, f t d Indicates direction d superior t Always forget the output of the gate, Indicates direction d superior t The output of the candidate memory gate at time, c t d Indicates direction d superior t The output of the memory gate at all times, o t d Indicates direction d superior t The output of the gate at all times, h t d Indicates direction d superior t The hidden state of the moment, d Indicates the direction of data transmission, forward or backward; x t is the time step t The input amount, h t±1 d Indicates direction d superior t The hidden state at ±1 time, c t±1 d It is the direction d superior t The memory cell state at ±1 moment, w xi d It is the direction d The weight matrix of the upper input gate, w xf d It is the direction d The weight matrix of the upper forget gate, w xc d It is the direction d The weight matrix of the candidate memory gate, w xo d It is the direction d The weight matrix of the upper output gate, b i d It is the direction d The bias term of the upper input gate, b f d It is the direction d The bias term of the upper forget gate, b c d It is the direction d The bias term of the candidate memory gate, b o d It is the direction d The bias term of the upper output gate, ⊙, is Hadamard Product, σ is the sigmoid function, tanh is the hyperbolic tangent function, and the formulas are: (10)； (11)； For data other than high-frequency data, the traditional LSTM network is used for prediction; S3-3, finally, according to the proportion of the subsequence in the original sequence, the prediction results of different networks are combined using the proportion weight. The combination formula is as follows: (12)； (13)； Where, w i t For the i The modal component is in the t The weight of the secondary cycle, IMF i t For the i The modal component is in the t The capacity value of the cycle, x t For the t Second cycle lithium battery capacity value, n is the number of test cycles, RUL The remaining service life of the lithium battery; S3-4, the evaluation indicator used by the model is the mean absolute percentage error MAPE , root mean square error RMSE , mean absolute error MAE and absolute error AE , the calculation formula is: (14)； (15)； (16)； (17)； Where, represents the predicted value, y i Indicates the actual value, n represents the predicted number of cycles, RUL re Indicates the number of cycles before the actual usage threshold is reached. RUL pr Indicates the number of cycles when the prediction reaches the usage threshold.

Images