A method for predicting the cycle life of energy storage power supply based on data analysis

Abstract

The present invention provides a method for predicting the cycle life of an energy storage power supply based on data analysis, and relates to the technical field of energy storage power supply life analysis. The method includes: collecting historical status monitoring data of a power module, performing a single-cell aging feature analysis on each battery cell in the power module, and constructing a local cycle aging path tensor for each battery cell; performing an interactive feature analysis on the power module, extracting multiple interactive feature parameters and generating a first global cycle aging path tensor for each charge and discharge cycle; performing DTW clustering on the local cycle aging path tensors of multiple battery cells to obtain multiple cluster categories, adding cluster labels to the multiple cluster categories, constructing a training data set based on multiple first global cycle aging path tensors and cluster labels, training to obtain a cycle life prediction model, and predicting the cycle life of the power module through the cycle life prediction model. The present invention achieves the improvement of the prediction accuracy of the cycle life of the energy storage power supply.

Description

Technical Field

[0001] The present invention relates to the technical field of energy storage power supply life analysis, and in particular to a method for predicting the cycle life of an energy storage power supply based on data analysis. Background Art

[0002] With the widespread application of energy storage technology, energy storage power supply has become an important component in the fields of power grid peak regulation, renewable energy grid connection and electric vehicles. Energy storage power supply is composed of multiple battery cells. During the long-term charge and discharge cycle, due to the influence of factors such as the position difference of each battery cell in the system, heat dissipation conditions, and uneven load distribution, the aging process of different battery cells is prone to inconsistency.

[0003] Some energy storage power cycle life prediction methods usually rely on the overall macro characteristics of the system, such as overall capacity decay, internal resistance change, etc., while ignoring the differences between individual battery cells within the system, making this overall life prediction strategy unable to capture the decay mode of a single battery cell and its impact on the health status of the entire system. In practical applications, this can easily lead to low prediction accuracy and inaccurate assessment of the remaining life of the energy storage power system. Summary of the invention

[0004] In order to solve the above technical problems, the present invention proposes a method for predicting the cycle life of energy storage power supply based on data analysis, which improves the prediction accuracy of the cycle life of energy storage power supply by fully considering the inconsistency of aging modes of each battery cell inside the module.

[0005] To achieve the above object, the present invention provides a method for predicting the cycle life of an energy storage power source based on data analysis, comprising:

[0006] Collect historical status monitoring data of the power module, including the status data of each battery cell in the power module during multiple charge and discharge cycles;

[0007] Based on the historical status monitoring data, the aging characteristics of each battery cell in the power module are analyzed, and the local cycle aging path tensor of each battery cell is constructed according to the multiple state characteristics of each battery cell in each charge and discharge cycle;

[0008] Perform interactive feature analysis on the power module according to multiple local cycle aging path tensors, extract multiple interactive feature parameters of the power module in each charge and discharge cycle, construct an interactive feature matrix corresponding to each interactive feature parameter, and perform feature splicing on multiple interactive feature matrices in each charge and discharge cycle to generate the first global cycle aging path tensor of each charge and discharge cycle;

[0009] DTW clustering is performed on the local cycle aging path tensors of multiple battery cells to obtain multiple cluster categories, cluster labels are added to the multiple cluster categories based on multiple preset statistical feature thresholds, and a training data set is constructed according to the first global cycle aging path tensor of multiple charge and discharge cycles and the cluster label of each battery cell. The cycle life prediction model is trained through the training data set. The cycle life prediction model includes an input layer, an interactive feature reconstruction layer, a timing feature extraction layer, a pattern feature separation layer, a residual path network layer, a pattern feature fusion layer, a timing feature fusion layer and an output layer. The cycle life of the power module is predicted through the trained cycle life prediction model.

[0010] Preferably, the cycle life prediction model further includes:

[0011] The input layer is used to receive multiple groups of local sample data in the training data set, wherein for the training data set, a cluster label is added to each battery cell in the first global cycle aging path tensor of multiple charge and discharge cycles to construct a training data set;

[0012] The interactive feature reconstruction layer is used to perform interactive feature reconstruction on multiple first global cycle aging path tensors in the training data set to obtain a second global cycle aging path tensor corresponding to each first global cycle aging path tensor, and to obtain a temporal aging tensor after feature splicing of multiple second global cycle aging path tensors. The temporal aging tensor is segmented by sliding windows based on a preset sliding window strategy to obtain multiple groups of local sample data;

[0013] The time series feature extraction layer is used to extract the time series features between multiple groups of local sample data to obtain a time series feature sequence, and the pattern feature separation layer is used to map the time series feature sequence to multiple pattern feature subspaces;

[0014] The residual path network layer is used to perform residual learning on the mapping features of multiple pattern feature subspaces to achieve feature reconstruction. The pattern feature fusion layer is used to fuse the mapping features after feature reconstruction of different pattern feature subspaces to obtain the global aging path features. The time series feature fusion layer is used to perform time series feature fusion on the global aging path features. The output layer is used to output the life prediction results.

[0015] Preferably, interactive feature reconstruction is performed on multiple first global cycle aging path tensors in the training data set to obtain a second global cycle aging path tensor corresponding to each first global cycle aging path tensor, including:

[0016] The interactive feature reconstruction layer includes a label embedding layer and a feature reconstruction layer. The label embedding layer is used to convert multiple cluster labels in the training data set into label embedding vectors. The feature reconstruction layer is used to perform interactive feature reconstruction on multiple first global cycle aging path tensors based on the multiple label embedding vectors, wherein:

[0017] ;

[0018] In the formula, For the The second global loop aging path tensor and The interaction characteristics between the battery cells, Represents the weight mapping function , Respectively and The label embedding vector of each battery cell, For the The first global loop aging path tensor and The interaction characteristics between the battery cells.

[0019] Preferably, the construction of the first global cycle aging path tensor for each charge and discharge cycle includes:

[0020] For any charge and discharge cycle, calculate the average voltage change rate of multiple battery cells And the standard deviation of voltage change rate , the voltage interaction characteristic parameters between any two battery cells are calculated according to the mean value and standard deviation of the voltage change rate, and the voltage interaction characteristic matrix is ​​constructed:

[0021] ;

[0022] in, For the , and Voltage interaction characteristic parameters between battery cells, , Respectively , and The voltage change rate of each battery cell;

[0023] Calculate the internal resistance change value between any two battery cells, calculate the internal resistance interaction characteristic parameters between any two battery cells based on multiple internal resistance change values, and construct the internal resistance interaction characteristic matrix:

[0024] ;

[0025] In the formula, For the , and The internal resistance interaction characteristic parameters between battery cells, , Respectively , and The change in internal resistance of each battery cell;

[0026] The average temperature of each battery cell and the average temperature of multiple battery cells are calculated. The temperature interaction characteristic parameters between any two battery cells are calculated based on the average temperature difference of multiple battery cells, and the temperature interaction characteristic matrix is ​​constructed:

[0027] ;

[0028] In the formula, For the , and The temperature interaction characteristic parameters between battery cells, , Respectively , and The average temperature of each battery cell, is the standard deviation of the mean temperature;

[0029] Calculate the capacity attenuation value of each battery cell, calculate the capacity interaction characteristic parameters between any two battery cells based on multiple capacity attenuation values, and construct the capacity interaction characteristic matrix:

[0030] ;

[0031] In the formula, For the , and The capacity interaction characteristic parameters between battery cells, , Respectively , and The capacity decay value of each battery cell, is the adjustment coefficient;

[0032] The first global cycle aging path tensor of each charge and discharge cycle is generated by feature concatenation of multiple interactive feature matrices in each charge and discharge cycle.

[0033] Preferably, the cycle life prediction model further includes:

[0034] According to the predicted loss , interactive feature reconstruction loss and pattern feature reconstruction loss Constructing the loss function of the cycle life prediction model :

[0035] ;

[0036] In the formula, , and They are prediction loss weight, interaction feature reconstruction loss weight and pattern feature reconstruction loss weight respectively;

[0037] For interactive feature reconstruction loss :

[0038] ;

[0039] In the formula, is the second global loop aging path tensor after interactive feature reconstruction in the interactive feature reconstruction layer, is the first global loop aging path tensor without interactive feature reconstruction, express and The cosine similarity between ;

[0040] For the mode feature reconstruction loss :

[0041] ;

[0042] In the formula, Reconstruct features for the pattern map features generated in the residual path network layer, The pattern mapping features generated in the pattern feature separation layer, is the L2 norm.

[0043] Preferably, the pattern feature subspace uses the following formula for feature mapping:

[0044] ;

[0045] In the formula, For the The mapping characteristics of the pattern feature subspace, For the The mapping matrix of the pattern feature subspace is The temporal feature sequence generated by the temporal feature extraction layer; For the The bias term of the pattern feature subspace.

[0046] Preferably, the time series feature extraction layer is constructed using LSTM units, and the residual path network layer includes multiple residual blocks.

[0047] Preferably, the output layer uses a fully connected layer to output the life prediction result.

[0048] The present invention has the following beneficial effects:

[0049] By introducing multi-level feature extraction, label-guided feature reconstruction strategy and multi-objective loss function design, the present invention can effectively capture the complex aging pattern differences within the energy storage power supply module, and realize accurate life prediction in multimodal feature learning and fusion, which can significantly improve the accuracy of life prediction and model robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 A schematic flow chart of a method for predicting the cycle life of an energy storage power source based on data analysis provided in an embodiment of the present invention.

[0051] Figure 2 A schematic diagram of the structure of a cycle life prediction model provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0052] In order to make those skilled in the art better understand the technical solutions in the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0053] See also Figure 1 and Figure 2 The present invention provides a method for predicting the cycle life of an energy storage power source based on data analysis, which specifically includes the following steps:

[0054] Step S1, collecting historical status monitoring data of the power module, performing single-cell aging characteristic analysis on each battery cell in the power module, and constructing a local cycle aging path tensor of each battery cell according to multiple state characteristics of each battery cell in each charge and discharge cycle.

[0055] Specifically, historical status monitoring data can be the status data of multiple battery cells during the long-term use of the power module. These data usually include but are not limited to information such as voltage, current, temperature, internal resistance, battery state of charge, and remaining capacity. Each battery cell will undergo multiple cycles during the charging and discharging process, so the status data of each battery cell in different charging and discharging cycles is collected. In order to better capture the aging pattern of each battery cell, feature analysis is performed on the data of each charging and discharging cycle, and multiple state features in each charging and discharging cycle are extracted, such as capacity decay features, voltage features, internal resistance features, temperature features, current features, etc., so as to construct a local cycle aging path tensor for each battery cell, which is used to represent the changes in multiple state features of the battery cell in a series of charging and discharging cycles, reflecting the aging process and health status of the battery cell.

[0056] Step S2: performing interactive feature analysis on the power module according to multiple local cycle aging path tensors, extracting multiple interactive feature parameters of the power module in each charge and discharge cycle, constructing an interactive feature matrix corresponding to each interactive feature parameter, and performing feature splicing on the multiple interactive feature matrices in each charge and discharge cycle to generate the first global cycle aging path tensor for each charge and discharge cycle.

[0057] Specifically, in energy storage power modules, different battery cells may show obvious aging differences due to factors such as location, heat dissipation conditions and load. Analyzing the aging characteristics of each battery cell alone cannot accurately assess the overall health of the module. In order to improve the prediction accuracy, it is necessary to analyze the interactive characteristics of the battery cells inside the power module to capture the relationship between different battery cells and the consistency of the attenuation mode.

[0058] For any charge and discharge cycle, an interactive feature analysis is performed on the relationship between the battery cells of the power module. In this embodiment, the interactive feature analysis is performed based on state characteristics such as voltage, internal resistance, temperature, and capacity to obtain multiple interactive feature parameters between any two battery cells. An interactive feature matrix corresponding to each interactive feature parameter is constructed based on these interactive feature parameters, and multiple interactive feature matrices in the same charge and discharge cycle are feature spliced ​​to generate a first global cycle aging path tensor that characterizes the interactive relationship and aging pattern between multiple battery cells in the power module in the charge and discharge cycle.

[0059] Step S3, performing DTW clustering on the local cycle aging path tensors of multiple battery cells to obtain multiple cluster categories, adding cluster labels to the multiple cluster categories based on multiple preset statistical feature thresholds, constructing a training data set according to the first global cycle aging path tensor of multiple charge and discharge cycles and the cluster label of each battery cell, training a cycle life prediction model through the training data set, and predicting the cycle life of the power module through the trained cycle life prediction model.

[0060] Specifically, the first global cycle aging path tensor represents the interaction between different state characteristics in one charge and discharge cycle. Due to the differences in aging patterns of different battery cells, in order to better understand the aging trends of each battery cell in the module, cluster analysis is performed on multiple cells based on the local cycle aging path tensors of the battery cells. In this embodiment, the dynamic time warping (DTW) method is used to perform pattern matching and clustering on multiple local cycle aging path tensors.

[0061] Specifically, the DTW algorithm is used to measure the similarity between the aging path tensors of each battery cell, and cluster analysis is performed on all battery cells according to the similarity matrix. Finally, each battery cell is assigned to a specific clustering category. Different clustering categories reflect different aging modes, such as fast decay mode, stable decay mode, abnormal fluctuation mode, etc. Different statistical feature thresholds can be pre-designed according to the characteristics of different aging modes. For example, if the average capacity decay rate of a certain category is significantly higher than that of other categories, it can be labeled as "fast decay type". If the capacity mean of a certain category changes more slowly with the number of cycles, it can be labeled as "stable decay type". Technicians in this field can set multiple preset statistical feature thresholds according to actual needs, so as to realize automatic label allocation based on feature statistics to determine cluster labels of different clustering categories.

[0062] Then, a training data set is constructed based on the first global cycle aging path tensor of multiple charge and discharge cycles and the clustering label of each battery cell. Specifically, a clustering label is added to each battery cell in the first global cycle aging path tensor of multiple charge and discharge cycles to construct a training data set. The pre-constructed cycle life prediction model is trained with the training data set. After inputting the new global path representation, the trained cycle life prediction model is used to predict the remaining life of the power module to obtain an accurate cycle life prediction result.

[0063] In one embodiment, regarding step S2, constructing a first global cycle aging path tensor for each charge and discharge cycle specifically includes:

[0064] Taking any charge and discharge cycle as an example, based on multiple local cycle aging path tensors, interactive feature analysis is performed on the state parameters of the battery cell, such as voltage, internal resistance, temperature and capacity, to construct multiple interactive feature matrices, specifically:

[0065] For voltage characteristics, the purpose of constructing a voltage interaction feature matrix is ​​to measure the consistency or difference in the voltage change rate between different battery cells in the module. In order to capture this interactive relationship, the voltage change rate of all battery cells in the charge and discharge cycle is analyzed, and the relationship between battery cells is quantified through the interactive feature parameter matrix. The voltage interaction feature matrix is ​​composed of the voltage interaction feature parameters between two battery cells. By calculating the average voltage change rate of multiple battery cells in the charge and discharge cycle, the voltage interaction feature matrix is ​​constructed. And the standard deviation of voltage change rate , the voltage interaction characteristic parameters between any two battery cells are calculated based on the mean value and standard deviation of the voltage change rate:

[0066] ;

[0067] in, For the , and Voltage interaction characteristic parameters between battery cells, , Respectively , and The voltage change rate of each battery cell.

[0068] The voltage interaction characteristic matrix can measure the similarity of voltage change rates of different battery cells within the battery module. When the value is close to 0, it means , and The voltage changes of individual battery cells vary greatly. When the value is close to 1, it means that the patterns of voltage change of the two are very similar, reflecting their consistency in the aging process.

[0069] The change in internal resistance is one of the important indicators for measuring battery aging. During different charge and discharge cycles, the internal resistance of a battery cell may increase significantly with the increase in the number of cycles. In order to better capture the impact of changes in the internal resistance of a battery cell on the overall health of the module, the internal resistance interaction characteristics between battery cells can be calculated to reflect the aging consistency within the battery module. The internal resistance interaction feature matrix is ​​composed of the internal resistance interaction feature parameters between any two battery cells. The internal resistance change value between any two battery cells is calculated, and the internal resistance interaction feature parameters between any two battery cells are calculated based on multiple internal resistance change values:

[0070] ;

[0071] In the formula, For the , and The internal resistance interaction characteristic parameters between battery cells, , Respectively , and The internal resistance change of each battery cell.

[0072] The internal resistance interaction characteristic matrix can reflect whether the internal resistance increase rate of different battery cells is consistent during the aging process. Battery cells with relatively consistent internal resistance changes have less impact on the overall health of the module; while battery cells with large internal resistance changes, such as "bottleneck batteries" with too fast internal resistance increases, may have an adverse effect on the overall life of the module. Therefore, this matrix can help identify potential abnormal batteries inside the module.

[0073] The uniformity of temperature distribution inside the battery module is an important factor affecting the life of the module. Uneven temperature distribution can cause some battery cells in the module to overheat or overcool, thereby accelerating aging. The temperature interaction feature can measure the temperature consistency between battery cells. The temperature interaction feature matrix is ​​composed of temperature interaction feature parameters between two battery cells. The temperature interaction feature parameters between any two battery cells are calculated by calculating the average temperature of each battery cell and the average temperature of multiple battery cells, and the temperature interaction feature parameters between any two battery cells are calculated based on the average temperature difference of multiple battery cells:

[0074] ;

[0075] In the formula, For the , and The temperature interaction characteristic parameters between battery cells, , Respectively , and The average temperature of each battery cell, is the standard deviation of the mean temperature.

[0076] when When the value is close to 1, it means that the temperature distribution of the two is highly consistent. When the value is close to 0, it means that there is a large difference in temperature distribution between the two. The temperature interaction feature matrix can reflect the thermal balance state inside the battery module, thereby helping to identify the problem of increased battery cell aging caused by uneven heat dissipation.

[0077] Capacity decay is one of the core indicators for measuring the health status of batteries. In order to better capture the capacity decay patterns of different battery cells within the module, the capacity interaction characteristics can be calculated to reflect their consistency in capacity decay. The capacity interaction characteristic matrix is ​​composed of the capacity interaction characteristic parameters between two battery cells. The capacity decay value of each battery cell is calculated, and the capacity interaction characteristic parameters between any two battery cells are calculated based on multiple capacity decay values:

[0078] ;

[0079] In the formula, For the , and The capacity interaction characteristic parameters between battery cells, , Respectively , and The capacity decay value of each battery cell, is the adjustment coefficient.

[0080] The capacity interaction characteristic matrix can measure the consistency of different battery cells during the capacity decay process. When the value is close to 1, it means that the two are very similar in capacity decay mode; When the value is close to 0, it means that there is a significant difference in the capacity decay pattern between the two. This matrix can help identify the "bottleneck cells" within the module, that is, those cells that decay much faster than other battery cells.

[0081] In each charge and discharge cycle, the above steps can generate the interactive feature matrix of voltage, internal resistance, temperature and capacity. In order to more comprehensively reflect the aging mode inside the battery module, these interactive feature matrices are feature spliced ​​to generate the first global cycle aging path tensor of each charge and discharge cycle, which is used to comprehensively reflect the interactive relationship and aging mode between multiple battery cells inside the module, thereby providing high-quality input features for subsequent life prediction.

[0082] In one embodiment, regarding step S3, the cycle life prediction model includes an input layer, an interactive feature reconstruction layer, a timing feature extraction layer, a pattern feature separation layer, a residual path network layer, a pattern feature fusion layer, a timing feature fusion layer and an output layer, wherein:

[0083] The input layer is used to receive multiple groups of local sample data in the training data set;

[0084] The interactive feature reconstruction layer is used to perform interactive feature reconstruction on multiple first global cycle aging path tensors in the training data set to obtain the second global cycle aging path tensor corresponding to each first global cycle aging path tensor, and to obtain a temporal aging tensor after feature splicing of multiple second global cycle aging path tensors. The temporal aging tensor is segmented by sliding window based on a preset sliding window strategy to obtain multiple groups of local sample data.

[0085] It is worth noting that during the aging analysis of energy storage power modules, the relationships between different battery cells in the module, such as voltage changes, internal resistance deviations, and capacity attenuation differences, will change dynamically over time. In order to more accurately capture these complex aging pattern differences, a dynamic interactive feature reconstruction strategy based on label embedding vectors is introduced.

[0086] Specifically, for the input data received by the input layer, only the original interactive feature tensor, that is, multiple first global cycle aging path tensors, is considered first, and the feature reconstruction guidance based on cluster labels is not considered temporarily. By introducing the interactive feature reconstruction layer, the cluster labels corresponding to different cluster categories are converted into label embedding vectors. During the model training process, these embedding vectors are used as guiding factors to adjust the interactive feature weights of each pair of battery cells in the module, thereby dynamically optimizing the interactive relationship between different battery cells and improving the model's ability to express complex aging patterns within the module.

[0087] The interactive feature reconstruction layer includes a label embedding layer and a feature reconstruction layer. The label embedding layer is used to convert multiple cluster labels in the training data set into label embedding vectors. The feature reconstruction layer is used to perform interactive feature reconstruction on multiple first global cycle aging path tensors based on multiple label embedding vectors. Specifically, in the label embedding layer, each cluster label is mapped to a low-dimensional feature representation based on the label embedding technology. In the interactive feature reconstruction layer, the interactive feature parameters between each pair of battery cells in the module are dynamically adjusted based on the label embedding vector, thereby generating a second global cycle aging path tensor guided by labels. Among them:

[0088] ;

[0089] In the formula, For the The second global loop aging path tensor and The interaction features between battery cells are used to characterize the interaction features reconstructed after label embedding guidance. Represents the weight mapping function , Respectively and The label embedding vector of each battery cell, For the The first global loop aging path tensor and The interaction features between battery cells are used to characterize the interaction features of the original input.

[0090] After the interactive feature reconstruction of multiple first global cycle aging path tensors is performed through the interactive feature reconstruction layer, the interactive feature parameters of each pair of battery cells are dynamically weighted and adjusted according to the similarity of their aging patterns, so that the differences in aging patterns within the module can be better captured. The reconstructed second global cycle aging path tensor not only contains the interactive features between battery cells within the module, but also reflects the relationship between each aging pattern in the feature space through the label embedding vector, which contains rich feature information.

[0091] The timing aging tensor is obtained by feature splicing of multiple second global cycle aging path tensors, which can characterize the characteristic changes in the global aging process of the power module. The timing aging tensor is segmented by sliding windows using a preset sliding window strategy to obtain multiple groups of local sample data representing the aging pattern characteristics in different time windows.

[0092] The time series feature extraction layer is used to extract the time series features between multiple groups of local sample data to obtain a time series feature sequence. In this embodiment, an LSTM unit is used to construct the time series feature extraction layer to capture the time dependency between multiple groups of local sample data. The dynamic evolution of the local path representation can be captured in the time series, and the context feature representation of each time segment can be generated.

[0093] The pattern feature separation layer is used to map the temporal feature sequence containing context feature information to multiple pattern feature subspaces. The number of pattern feature subspaces can be set based on the number of clustering categories obtained by DTW clustering. Each pattern feature subspace is set with an independent linear projection matrix, and multiple mapping features can be output in the end. The pattern feature subspace uses the following formula for feature mapping:

[0094] ;

[0095] In the formula, For the The mapping characteristics of the pattern feature subspace, For the The mapping matrix of the pattern feature subspace is The temporal feature sequence generated by the temporal feature extraction layer; For the The bias term of the pattern feature subspace.

[0096] The residual path network layer is used to perform residual learning on the mapping features of multiple pattern feature subspaces to achieve feature reconstruction. The residual path network layer contains multiple residual blocks to improve the model's feature reconstruction capabilities under different modes. Through residual learning, the model can remove redundant information in the features and improve the discrimination of pattern feature representation.

[0097] The pattern feature fusion layer is used to fuse the mapping features after feature reconstruction of different pattern feature subspaces, that is, multiple features generated by the residual path network layer, to obtain a global aging path feature containing different pattern features. Its goal is to effectively integrate the information in each pattern feature subspace. In this embodiment, pattern feature fusion is performed based on the attention mechanism to improve the expressiveness of the global aging path feature representation.

[0098] The time series feature fusion layer is used to perform time series feature fusion on the global aging path features to generate a comprehensive context feature representation. Finally, the life prediction result is output through the output layer, and a scalar value is generated as the remaining life prediction value using the fully connected layer.

[0099] During the training process of the cycle life prediction model, according to the prediction loss , interactive feature reconstruction loss and pattern feature reconstruction loss Constructing the loss function of the cycle life prediction model , predict the loss Used to measure the performance of the model on the cycle life prediction task, interactive feature reconstruction loss Used to constrain the consistency between the reconstructed interaction features and the original features, mode feature reconstruction loss Used to measure the reconstruction effect of pattern features in different subspaces.

[0100] Loss function for cycle life prediction model :

[0101] ;

[0102] In the formula, , and They are the prediction loss weight, the interaction feature reconstruction loss weight, and the pattern feature reconstruction loss weight. The weight coefficient can be determined by the hyperparameter tuning strategy to balance the impact of different loss items on model training.

[0103] For interactive feature reconstruction loss :

[0104] ;

[0105] In the formula, is the second global loop aging path tensor after interactive feature reconstruction in the interactive feature reconstruction layer, is the first global loop aging path tensor without interactive feature reconstruction, express and The cosine similarity between .

[0106] The interactive feature reconstruction loss is used to measure the consistency between the reconstructed second global loop aging path tensor and the original first global loop aging path tensor in the interactive feature reconstruction layer guided by label embedding. Its core purpose is to ensure that after the introduction of the label embedding vector, the interactive features can maintain consistency with the original features, thereby avoiding feature loss or overfitting problems in the model during the label guidance process.

[0107] For the mode feature reconstruction loss :

[0108] ;

[0109] In the formula, Reconstruct features for the pattern map features generated in the residual path network layer, The pattern mapping features generated in the pattern feature separation layer, is the L2 norm.

[0110] The pattern feature reconstruction loss is used to measure the reconstruction effect of the model in the pattern feature separation and residual path network layer. Its core purpose is to ensure that the representation of pattern features in different pattern subspaces can be effectively reconstructed, thereby removing redundant information in the features and improving the expression ability of the model under different aging patterns.

[0111] The prediction task loss is used to measure the performance of the model in the cycle life prediction task. Its core purpose is to improve the accuracy of the model in the cycle life prediction task by minimizing the difference between the model prediction value and the true value. In this embodiment, the mean squared error (MSE) is used to measure the difference between the model prediction value and the true value.

[0112] Based on the above loss function The model can achieve joint optimization on multiple objectives such as interactive feature reconstruction, pattern feature learning and task prediction, thereby improving the ability to express and predict complex aging patterns within the battery module, and improving the accuracy and robustness of the model in cycle life prediction tasks.

[0113] The traditional partial cycle life prediction scheme is based on the overall characteristics of the module, ignoring the inconsistency of the aging patterns of each battery cell inside due to differences in position, heat dissipation conditions and load. The present invention can identify the differences between each battery cell under different aging modes by extracting and analyzing the local aging path tensor of each battery cell inside the module, thereby more accurately capturing the complex aging characteristics and dynamic evolution laws inside the module. By introducing label embedding vectors and interactive feature reconstruction strategies, the interaction weights between different battery cells can be dynamically adjusted to capture the complex relationships between different battery cells under different aging modes, so as to better reflect the pattern differences of each battery cell in the reconstructed interactive features, so that the model can better identify the aging laws under different modes. By introducing residual path networks and pattern feature separation layers, the features of different aging modes are mapped to independent subspaces for learning, and the pattern feature fusion strategy is used to fuse the feature representations of each pattern subspace into a global path representation, so that the model can maintain the distinction between different modes during the feature fusion process, thereby improving the generalization ability of the global path representation in complex aging modes and preventing pattern confusion or feature loss during the fusion process. A multi-objective loss function based on prediction task loss, interaction feature reconstruction loss and pattern feature reconstruction loss is designed, so that the model can be jointly optimized at multiple levels such as global task objectives, interaction feature consistency, and pattern feature reconstruction, ensuring that the model can effectively learn feature representations at each level under complex aging patterns and improve the model's ability to express multimodal aging patterns.

[0114] The above is only a specific embodiment of the present invention, so that those skilled in the art can understand or implement the present invention. Various modifications to these embodiments will be obvious to those skilled in the art. Parts not described in detail in this specification belong to the prior art known to those skilled in the art.

Claims

1. A method for predicting the cycle life of energy storage power source based on data analysis, characterized in that: include: Collect historical status monitoring data of the power module, including the status data of each battery cell in the power module during multiple charge and discharge cycles; Based on the historical status monitoring data, the aging characteristics of each battery cell in the power module are analyzed, and the local cycle aging path tensor of each battery cell is constructed according to the multiple state characteristics of each battery cell in each charge and discharge cycle; Perform interactive feature analysis on the power module according to multiple local cycle aging path tensors, extract multiple interactive feature parameters of the power module in each charge and discharge cycle, construct an interactive feature matrix corresponding to each interactive feature parameter, and perform feature splicing on multiple interactive feature matrices in each charge and discharge cycle to generate the first global cycle aging path tensor of each charge and discharge cycle; DTW clustering is performed on the local cycle aging path tensors of multiple battery cells to obtain multiple cluster categories, cluster labels are added to the multiple cluster categories based on multiple preset statistical feature thresholds, and a training data set is constructed according to the first global cycle aging path tensor of multiple charge and discharge cycles and the cluster label of each battery cell. The cycle life prediction model is trained through the training data set. The cycle life prediction model includes an input layer, an interactive feature reconstruction layer, a timing feature extraction layer, a pattern feature separation layer, a residual path network layer, a pattern feature fusion layer, a timing feature fusion layer and an output layer. The cycle life of the power module is predicted through the trained cycle life prediction model.

2. The method for predicting the cycle life of energy storage power source based on data analysis according to claim 1 is characterized in that: For cycle life prediction models, also included: The input layer is used to receive multiple groups of local sample data in the training data set, wherein for the training data set, a cluster label is added to each battery cell in the first global cycle aging path tensor of multiple charge and discharge cycles to construct a training data set; The interactive feature reconstruction layer is used to perform interactive feature reconstruction on multiple first global cycle aging path tensors in the training data set to obtain a second global cycle aging path tensor corresponding to each first global cycle aging path tensor, and to obtain a temporal aging tensor after feature splicing of multiple second global cycle aging path tensors. The temporal aging tensor is segmented by sliding windows based on a preset sliding window strategy to obtain multiple groups of local sample data; The time series feature extraction layer is used to extract the time series features between multiple groups of local sample data to obtain a time series feature sequence, and the pattern feature separation layer is used to map the time series feature sequence to multiple pattern feature subspaces; The residual path network layer is used to perform residual learning on the mapping features of multiple pattern feature subspaces to achieve feature reconstruction. The pattern feature fusion layer is used to fuse the mapping features after feature reconstruction of different pattern feature subspaces to obtain the global aging path features. The time series feature fusion layer is used to perform time series feature fusion on the global aging path features. The output layer is used to output the life prediction results.

3. The method for predicting the cycle life of energy storage power source based on data analysis according to claim 1 is characterized in that: Interactive feature reconstruction is performed on multiple first global cyclic aging path tensors in the training data set to obtain a second global cyclic aging path tensor corresponding to each first global cyclic aging path tensor, including: The interactive feature reconstruction layer includes a label embedding layer and a feature reconstruction layer. The label embedding layer is used to convert multiple cluster labels in the training data set into label embedding vectors. The feature reconstruction layer is used to perform interactive feature reconstruction on multiple first global cycle aging path tensors based on the multiple label embedding vectors, wherein: ； In the formula, For the The second global loop aging path tensor and The interaction characteristics between the battery cells, represents the weight mapping function, , Respectively and The label embedding vector of each battery cell, For the The first global loop aging path tensor and The interaction characteristics between the battery cells.

4. The method for predicting the cycle life of energy storage power source based on data analysis according to claim 3 is characterized in that: The construction of the first global cycle aging path tensor for each charge and discharge cycle includes: For any charge and discharge cycle, calculate the average voltage change rate of multiple battery cells And the standard deviation of voltage change rate , the voltage interaction characteristic parameters between any two battery cells are calculated according to the mean value and standard deviation of the voltage change rate, and the voltage interaction characteristic matrix is ​​constructed: ； in, For the and Voltage interaction characteristic parameters between battery cells, , Respectively and The voltage change rate of each battery cell; Calculate the internal resistance change value between any two battery cells, calculate the internal resistance interaction characteristic parameters between any two battery cells based on multiple internal resistance change values, and construct the internal resistance interaction characteristic matrix: ； In the formula, For the and The internal resistance interaction characteristic parameters between battery cells, , Respectively and The change in internal resistance of each battery cell; The average temperature of each battery cell and the average temperature of multiple battery cells are calculated. The temperature interaction characteristic parameters between any two battery cells are calculated based on the average temperature difference of multiple battery cells, and the temperature interaction characteristic matrix is ​​constructed: ； In the formula, For the and The temperature interaction characteristic parameters between battery cells, , Respectively and The average temperature of each battery cell, is the standard deviation of the mean temperature; Calculate the capacity attenuation value of each battery cell, calculate the capacity interaction characteristic parameters between any two battery cells based on multiple capacity attenuation values, and construct the capacity interaction characteristic matrix: ； In the formula, For the and The capacity interaction characteristic parameters between battery cells, , Respectively and The capacity decay value of each battery cell, is the adjustment coefficient; The first global cycle aging path tensor of each charge and discharge cycle is generated by feature concatenation of multiple interactive feature matrices in each charge and discharge cycle.

5. The method for predicting the cycle life of energy storage power source based on data analysis according to claim 4 is characterized in that: For cycle life prediction models, also included: According to the predicted loss , interactive feature reconstruction loss and pattern feature reconstruction loss Constructing the loss function of the cycle life prediction model : ； In the formula, , and They are prediction loss weight, interaction feature reconstruction loss weight and pattern feature reconstruction loss weight respectively; For interactive feature reconstruction loss : ； In the formula, is the second global loop aging path tensor after interactive feature reconstruction in the interactive feature reconstruction layer, is the first global loop aging path tensor without interactive feature reconstruction, express and The cosine similarity between ; For the mode feature reconstruction loss : ； In the formula, Reconstruct features for the pattern map features generated in the residual path network layer, The pattern mapping features generated in the pattern feature separation layer, is the L2 norm.

6. The method for predicting the cycle life of energy storage power source based on data analysis according to claim 5 is characterized in that: The pattern feature subspace uses the following formula for feature mapping: ； In the formula, For the The mapping characteristics of the pattern feature subspace, For the The mapping matrix of the pattern feature subspace is The temporal feature sequence generated by the temporal feature extraction layer; For the The bias term of the pattern feature subspace.

7. The method for predicting the cycle life of energy storage power source based on data analysis according to claim 6 is characterized in that: The LSTM unit is used to construct the temporal feature extraction layer, and the residual path network layer contains multiple residual blocks.

8. The method for predicting the cycle life of energy storage power source based on data analysis according to claim 7 is characterized in that: The output layer uses a fully connected layer to output the life prediction results.

Images