A new energy storage lithium battery life prediction method based on deep learning

Abstract

The application discloses a new energy energy storage lithium battery life prediction method based on deep learning, and relates to the technical field of energy storage battery management, and comprises the following steps: generating a lithium battery life prediction constraint set, normalizing and equivalent cycle converting energy storage lithium battery operation records to form an energy storage lithium battery operation observation sequence; generating a single operation evidence sequence through state normalization and feature coding, generating a battery physical correlation matrix according to the physical correlation of the battery monomer; inputting the two into an improved Graph WaveNet model to generate a battery degradation trajectory matrix; constructing a joint degradation segmentation cost by using an improved PELT algorithm to generate a degradation stage boundary set within the continuous degradation candidate boundary range; determining a life termination boundary according to a life termination criterion to generate a lithium battery life prediction result and a life prediction feedback record. The application can combine the single operation difference, the physical correlation and the multi-state degradation stage to complete life prediction.

Description

Technical Field

[0001] This invention relates to the field of energy storage battery management technology, and in particular to a method for predicting the lifespan of new energy storage lithium batteries based on deep learning. Background Technology

[0002] As the installed capacity of new energy storage systems continues to expand, the operating time of energy storage lithium batteries in grid peak shaving, distributed energy access and emergency power supply continues to increase. The manufacturing differences, connection relationships, thermal environment and equilibrium control status between battery cells will form different degradation paths during long-term charging and discharging.

[0003] Existing lifetime prediction technologies typically rely on historical data such as capacity, internal resistance, voltage, or temperature to build degradation models and then combine this with cycle count or operating time to estimate remaining lifetime. These technologies have received widespread attention. However, the following problems commonly exist in practical applications:

[0004] The operation records of energy storage lithium batteries have inconsistent sampling intervals, discontinuous charging and discharging intervals, and inconsistent cycle diameters under different operating conditions, making it difficult to directly form a continuous and comparable degradation observation sequence. Existing methods mostly model the state of individual cells independently, failing to fully utilize the physical connections formed by electrical connections, thermal management partitions, and equalization channels, making it difficult to characterize the asynchronous transmission of degradation states between associated cells.

[0005] Some deep learning models directly output a single lifetime value without jointly segmenting the multi-stage changes in capacity, internal resistance, and battery cluster consistency. Short-term fluctuations are easily identified as degradation transitions. At the same time, the prediction results lack consistent deviation records from the subsequent actual operating conditions, making it difficult to form traceable lifetime prediction feedback data.

[0006] Therefore, how to provide a deep learning-based method for predicting the lifespan of new energy storage lithium batteries is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] One objective of this invention is to propose a deep learning-based method for predicting the lifespan of new energy storage lithium batteries. This invention fully utilizes deep learning and degradation stage identification technologies, and describes in detail the methods for encoding the operating state of energy storage lithium batteries, physical correlation modeling, and determining the lifespan termination boundary. It has the advantages of continuous prediction process, clear degradation relationship, and traceable results.

[0008] A method for predicting the lifespan of a new energy storage lithium battery based on deep learning according to an embodiment of the present invention includes the following steps:

[0009] Step 1: Receive lithium battery life prediction task information, parse the prediction object and prediction conditions, and generate a lithium battery life prediction constraint set;

[0010] Step 2: Read the operation records of the energy storage lithium battery according to the lithium battery life prediction constraint set, perform data normalization and equivalent cyclic transformation on the operation records, and generate the energy storage lithium battery operation observation sequence;

[0011] Step 3: Perform state regularization and feature encoding on the operational observation sequence of the energy storage lithium battery to generate a single-cell operational evidence sequence;

[0012] Step 4: Obtain the structural association information of the energy storage lithium battery, establish association codes based on the physical association relationships between battery cells, and generate a battery physical association matrix;

[0013] Step 5: Input the single-unit running evidence sequence and battery physical correlation matrix into the improved Graph WaveNet model. After running evidence encoding, dynamic degradation relationship construction, multi-scale degradation propagation and lifetime trajectory output, the battery degradation trajectory matrix is ​​generated.

[0014] Step 6: Input the battery degradation trajectory matrix into the improved PELT algorithm, construct the joint degradation segmentation cost based on the multi-state degradation changes, perform boundary screening, cumulative cost calculation and boundary backtracking within the continuous degradation candidate boundary range, and generate the degradation stage boundary set;

[0015] Step 7: Process the degradation stage boundary set according to the lithium battery life prediction constraint set, determine the target life termination boundary, generate the lithium battery life prediction result, and generate life prediction feedback record based on the deviation between the actual operating state and the lithium battery life prediction result.

[0016] Optionally, step one specifically includes:

[0017] Read the prediction object description in the lithium battery life prediction task information, compare the object code in the prediction object description with the battery code in the energy storage lithium battery resource index item by item, and determine the resource record with the same battery code as the prediction object.

[0018] When the prediction target is a single battery cell, the resource record of the corresponding battery cell is included in the prediction target range.

[0019] When the prediction object is a battery module or battery cluster, the battery cells whose resource level paths belong to the corresponding paths of the prediction object are included in the prediction object scope.

[0020] Read the running data index corresponding to the predicted object range, retain the running data index where the bound battery code belongs to the predicted object range and the recording time overlaps with the task call period, and determine the overlapping period as the running record call range;

[0021] Read the prediction target from the lithium battery life prediction task information, and filter the configuration records in the life prediction configuration library that have the same target type and the same applicable battery type according to the prediction target and the battery type corresponding to the prediction object.

[0022] Extract equivalent cyclic conversion rules, degradation state determination rules, prediction range rules, and lifetime termination criteria;

[0023] The prediction object range, the running record call range, and the extracted rules are combined according to the preset field order to generate a lithium battery life prediction constraint set.

[0024] Optionally, step two specifically includes:

[0025] Read the energy storage lithium battery operation records according to the prediction object range and operation record call range in the lithium battery life prediction constraint set;

[0026] Sort the data in ascending order by the acquisition timestamp, delete duplicate records where the prediction object, acquisition timestamp, and record content are all the same, and remove records that are missing acquisition timestamps or current values ​​to generate a regular operation record.

[0027] According to the effective sampling interval in the equivalent cycle conversion rule, check the adjacent regular operation records, disconnect the adjacent records whose time interval exceeds the effective sampling interval, and divide the continuous charging and discharging interval according to the change of current direction.

[0028] For each continuous charge and discharge interval, calculate the product of the average absolute value of the current of adjacent regular operation records and the time interval converted to hours;

[0029] The products within the interval are summed to generate the interval charge and discharge capacity; the interval charge and discharge capacity is divided by twice the rated capacity bound in the equivalent cycle conversion rule to generate the interval equivalent cycle increment.

[0030] Based on the equivalent cyclic increment of the accumulated interval according to the collection timestamp, the cumulative results are written into the corresponding regularized operation record, and the regularized operation records with the cumulative results are arranged according to the collection timestamp to generate the energy storage lithium battery operation observation sequence.

[0031] Optionally, step three specifically includes:

[0032] The equivalent cycle window is divided according to the integer range of the cumulative equivalent cycle value in the energy storage lithium battery operation observation sequence, and the operation records of the same battery cell in each equivalent cycle window are arranged according to the collection timestamp.

[0033] From the operation records where the current direction is consistent and the state of charge changes continuously, the product of the average absolute value of the current of adjacent records and the time interval is accumulated. The accumulated result is divided by the corresponding state of charge change ratio, and the median of the calculation results within the same equivalent cyclic window is taken to generate the single-cell capacity state.

[0034] For adjacent operating records with non-zero current differences, calculate the ratio of the absolute value of the voltage difference to the absolute value of the current difference between the individual units, and take the median of each ratio to generate the internal resistance status of the individual unit.

[0035] Calculate the median voltage and median temperature of each battery cell within the same equivalent cycle window, and subtract the corresponding median within the prediction range from the median voltage and median temperature, and combine them to generate a cell state deviation code.

[0036] Calculate the ratio of the number of valid records in the balance control to the number of valid records in the window, and generate the balance occupancy.

[0037] When the individual unit capacity state, individual unit internal resistance state, and individual unit state deviation coding are all generated, the state trust flag is set to one; when any state is not generated, the state trust flag is set to zero.

[0038] Based on the feature dimension, robust standardization is performed using the median and median absolute deviation of the cell capacity state, cell internal resistance state, cell state deviation code, and balanced occupancy of each battery cell within the prediction target range. When the median absolute deviation is zero, the corresponding standardization result is set to zero.

[0039] The standardized unit capacity state, unit internal resistance state, unit state deviation code, and balanced occupancy are combined with the state credibility marker in a preset order and arranged according to the cumulative equivalent cyclic value to generate a unit operation evidence sequence.

[0040] Optionally, step four specifically involves:

[0041] Read the energy storage lithium battery structure association information according to the prediction object range in the lithium battery life prediction constraint set, and extract the electrical structure path, hot zone identifier and equalization channel identifier corresponding to each battery cell from the structure association information.

[0042] The battery cells are paired up according to the order of the battery cells in the single cell operation evidence sequence; when the electrical structure paths of two battery cells point to the same module or the same electrical branch, an electrical association code is written in the corresponding single cell pair; when the hot zone identifiers are consistent, a hot zone association code is written.

[0043] When the equalization channel identifiers are consistent, the control association code is written; when the corresponding judgment condition is not met, the corresponding unassociated code is written, and a single-unit pair association code is generated.

[0044] According to the arrangement position of each battery cell in the cell operation evidence sequence, the cell pair association code is written into the corresponding row and column position of the matrix, and the self-association code is written into the main diagonal position.

[0045] When structural association information is missing, only the self-association code corresponding to the main diagonal position is retained to generate the battery physical association matrix.

[0046] Optionally, the improved Graph WaveNet model includes a running evidence encoding layer, a state graph construction layer, a multi-scale diffusion layer, and a lifetime trajectory output layer, wherein an asynchronous weakening propagation mechanism is embedded in the state graph construction layer;

[0047] The operational evidence coding layer divides the individual operational evidence sequence according to a set historical window, performs linear projection on the individual operational evidence within the window, and generates individual state codes.

[0048] Based on the battery physical correlation matrix, the states of the associated neighborhood are collected, and a common state code is generated by median merging. The difference between the co-position of the individual state code and the common state code is calculated, and layer normalization is performed to generate the individual degradation code.

[0049] The state diagram construction layer performs a linear mapping on the battery physical correlation matrix to generate a priori correlation tensor; it concatenates the pairwise individual degenerate codes and performs a linear projection to generate a state correlation tensor.

[0050] The asynchronous weakening propagation mechanism processes prior correlation tensors and single-entity degenerate encodings to generate directed propagation gating matrices;

[0051] The directed transmission gate matrix is ​​multiplied element-wise with the state association tensor, and row normalization is performed using the SoftMax function to generate a dynamically degenerate adjacency matrix.

[0052] The multi-scale diffusion layer uses gated dilated causal convolution to process the monomer degradation encoding, and inputs the convolution results into the ordinary diffusion channel and the key monomer channel respectively.

[0053] The ordinary diffusion channel performs directed graph convolution according to the dynamic degenerate adjacency matrix, while the key single-unit channel compares the single-unit capacity state, single-unit internal resistance state, and single-unit state deviation encoding with the corresponding judgment conditions in the degenerate state judgment rules item by item.

[0054] Battery cells that meet the judgment criteria are marked as key degraded cells. The convolution features corresponding to the key degraded cells are retained, and zero values ​​are written at the positions of the remaining battery cells.

[0055] The results from the two channels are concatenated and linearly mapped, and multi-scale degradation propagation features are generated through residual connections and skip connections.

[0056] The lifetime trajectory output layer uses one-dimensional convolution to process multi-scale degradation propagation features, maps and generates the single-cell capacity state, single-cell internal resistance state and battery cluster consistency state corresponding to each predicted equivalent cycle position, and combines them according to the predicted equivalent cycle position and the battery single-cell arrangement order to generate the battery degradation trajectory matrix.

[0057] Optionally, the asynchronous weakening propagation mechanism specifically includes:

[0058] Read the non-zero correlation positions in the prior correlation tensor to determine candidate transmission singleton pairs;

[0059] According to the positional order in the historical window, calculate the difference between adjacent positions of the individual cell degradation code of each battery cell, and take the median of the difference of each coding dimension at the same position to generate the positional change.

[0060] The median of the absolute value of the position change and the sum of the median absolute deviation are used as the deviation threshold. Continuous positions whose absolute value of position change exceeds the deviation threshold are grouped into deviation intervals, and the direction of change is determined based on the sign of the cumulative result of the position change within the deviation interval.

[0061] Compare the first deviation intervals of the two cells in the candidate conductive cell pair; when the two deviation intervals change in the same direction and overlap, the cell with the earlier starting position of the deviation interval is determined as the conductive end, and the other cell is determined as the response end;

[0062] If two deviation intervals have the same starting position, different directions of change, or no overlap, the corresponding gate value is set to zero.

[0063] Calculate the ratio of the overlap length of the two deviation intervals to the length of the shorter deviation interval to generate a continuous matching value; calculate the reciprocal of the interval between the starting positions of the two deviation intervals plus one to generate a hysteresis decay value.

[0064] Multiply the corresponding prior correlation value, continuous matching value and hysteresis decay value to generate a directed conduction gating value, and write it into the matrix position pointing from the conduction end to the response end according to the battery cell arrangement order, and write zero value into the other positions to generate a directed conduction gating matrix.

[0065] Optionally, step six specifically includes:

[0066] The battery degradation trajectory matrix is ​​read according to the predicted equivalent cycle position. The decrease in single cell capacity state, the increase in single cell internal resistance state, and the decrease in battery cluster consistency state are calculated respectively for adjacent predicted equivalent cycle positions. The three types of state changes are uniformly converted into a unidirectional degradation sequence in which the numerical value increases to indicate the aggravation of degradation.

[0067] Robust standardization is performed using the median and median absolute deviation according to the state type. When the median absolute deviation is zero, the corresponding standardization result is set to zero.

[0068] For any two candidate boundaries, calculate the squared difference between the decrease in the capacity state of each battery cell, the increase in the internal resistance state of each battery cell and the mean of the corresponding interval, and calculate the squared difference between the decrease in the uniformity state of the battery cluster and the mean of the corresponding interval.

[0069] The squared differences corresponding to each individual battery cell and the squared differences corresponding to the consistency state of the battery cluster are summed to generate the joint degradation segmentation cost.

[0070] The continuous degradation candidate boundary rule is invoked to process the same-direction degradation sequence, generate the continuous degradation candidate boundary range, and the first predicted equivalent cycle position and the last predicted equivalent cycle position are added to the continuous degradation candidate boundary range;

[0071] Set the cumulative cost of the first predicted equivalent cycle position to zero, write the first predicted equivalent cycle position into the candidate predecessor set, and read the candidate boundaries within the continuously degenerate candidate boundary range in order of position.

[0072] For the current candidate boundary, calculate the cumulative cost of each candidate predecessor position, the sum of the joint degradation segment cost from the candidate predecessor position to the current candidate boundary and the segment penalty value, where the segment penalty value is the product of the number of degradation state types and the natural logarithm of the number of predicted equivalent cycle positions.

[0073] Select the candidate predecessor position with the smallest sum, and write the corresponding sum and candidate predecessor position into the cumulative cost and predecessor index of the current candidate boundary, respectively.

[0074] For candidate predecessor positions that are not selected, calculate the sum of the corresponding cumulative cost and the joint degenerate segment cost extending to the current candidate boundary. Remove candidate predecessor positions whose calculation results are not less than the cumulative cost of the current candidate boundary from the candidate predecessor set, and write the current candidate boundary into the candidate predecessor set.

[0075] Starting from the last predicted equivalent cycle position, candidate boundaries are traced in reverse order according to the predecessor index. The traced candidate boundaries are then arranged in ascending order according to the predicted equivalent cycle position to generate the degradation stage boundary set.

[0076] Optionally, the persistent degradation candidate boundary rule specifically refers to:

[0077] Read the same-direction degradation sequence according to the predicted equivalent cycle position, take the median of the decrease in cell capacity state and the increase in cell internal resistance state for each cell at the same position, and read the corresponding decrease in cell cluster consistency state to generate three types of position degradation.

[0078] When at least two of the three types of positional degradation are greater than zero, the corresponding position is marked as a position of degradation in the same direction.

[0079] When the same-direction degradation position is empty, an empty candidate boundary range for continuous degradation is generated; when the same-direction degradation position is not empty, the sum of the degradation amount of the three types of positions at each same-direction degradation position is calculated, and the sum of the median and the median absolute deviation of the corresponding sum of all same-direction degradation positions is used as the threshold for continuous degradation judgment.

[0080] Retain the same-direction degradation positions whose sum is not less than the continuous degradation judgment threshold, merge the retained positions adjacent to the predicted equivalent cycle positions into continuous degradation intervals, and delete isolated positions that are not continuous with the previous or next position.

[0081] Write the start and end positions of each continuously degenerate interval into the candidate boundary. For each intermediate position within the continuously degenerate interval, calculate the first difference of the degenerate sum of adjacent positions, and then calculate the absolute value of the difference between adjacent first differences.

[0082] The median and the sum of the absolute values ​​of each absolute value within the interval are used as the turning point threshold. Positions with absolute values ​​not less than the turning point threshold are written into candidate boundaries. The positions are sorted according to the predicted equivalent cyclic positions and duplicate positions are deleted to generate the range of candidate boundaries for continuous degradation.

[0083] Optionally, step seven specifically includes:

[0084] The equivalent cycle position of the prediction is limited according to the prediction range rules in the lithium battery life prediction constraint set, and the degradation stage boundary set is read within the limited equivalent cycle position of the prediction.

[0085] Extract the individual cell capacity state, individual cell internal resistance state, and battery cluster consistency state corresponding to the consecutive positions from the boundary of each degradation stage in the battery degradation trajectory matrix.

[0086] The capacity status and internal resistance status of each cell are compared with the corresponding threshold in the lifespan termination criterion. The number of cell cells that meet the corresponding threshold is counted, and the consistency status of the battery cluster is compared with the corresponding threshold.

[0087] Count the consecutive positions where the number of individual cells and the consistency of cell clusters in both types of batteries reach the life end criterion, and mark the boundary of the degradation stage where the number of consecutive positions reaches the life end criterion.

[0088] The marked degradation stage boundaries are arranged in ascending order according to the predicted equivalent cycle positions, and the first marked degradation stage boundary is determined as the target lifetime termination boundary.

[0089] Read the predicted equivalent cycle value corresponding to the target lifespan termination boundary, and read the cumulative equivalent cycle value corresponding to the last position of the energy storage lithium battery operation observation sequence. Calculate the difference between the predicted equivalent cycle value and the cumulative equivalent cycle value to generate the predicted remaining equivalent cycle value.

[0090] The prediction object, target life end boundary, predicted equivalent cycle value, predicted remaining equivalent cycle value and battery degradation trajectory matrix are combined to generate lithium battery life prediction results.

[0091] Read the actual running state of the predicted object, and extract the actual cumulative equivalent loop value and the corresponding three types of actual states from the actual running state;

[0092] When the actual operating state does not meet the lifespan termination criterion, the three types of predicted states corresponding to the actual cumulative equivalent cycle value are extracted from the battery degradation trajectory matrix in the lithium battery lifespan prediction results, and the difference between the three types of actual states and the corresponding predicted states is calculated.

[0093] When the actual operating condition meets the end-of-life criterion, calculate the difference between the actual cumulative equivalent cycle value and the predicted equivalent cycle value;

[0094] The prediction object, corresponding difference, actual cumulative equivalent cyclic value, and target lifetime termination boundary are combined to generate a lifetime prediction feedback record.

[0095] The beneficial effects of this invention are:

[0096] Compared with existing technologies, this invention first performs data normalization and equivalent cycle transformation on the operation records of energy storage lithium batteries, so that the operation data under different sampling intervals, different charging and discharging processes and different object levels are arranged according to a unified equivalent cycle caliber. It also extracts the single-cell capacity status, single-cell internal resistance status, single-cell status deviation code and equalization occupancy to form a continuous and comparable single-cell operation evidence sequence. At the same time, it constructs a battery physical correlation matrix by combining electrical structure path, hot zone identification and equalization channel identification, providing a data foundation and structural basis for degradation relationship analysis.

[0097] Based on this, the present invention employs an improved Graph WaveNet model to jointly process the single-unit running evidence sequence and the battery physical correlation matrix. It forms a single-unit degradation code by the difference between the common state code and the single-unit state code, and constructs a dynamic degradation adjacency matrix by combining the prior correlation tensor and the state correlation tensor. The asynchronous weakening propagation mechanism characterizes the directed propagation of the degradation state according to the sequential relationship, change direction and continuous overlap of the deviation interval of the associated single units. It retains the overall propagation characteristics and key degradation characteristics in conjunction with the ordinary diffusion channel and the key single-unit channel, forming the battery degradation trajectory matrix.

[0098] Furthermore, this invention employs an improved PELT algorithm to divide the battery degradation trajectory matrix into stages, converting multiple state changes into a unidirectional degradation sequence. By combining degradation segmentation costs and continuous degradation candidate boundary rules, candidate positions with continuous degradation basis are defined, allowing cumulative cost calculation and boundary backtracking to avoid isolated fluctuations, thus obtaining a degradation stage boundary set. Then, the target lifespan termination boundary is determined according to the lifespan termination criterion, generating lithium battery lifespan prediction results and lifespan prediction feedback records, forming a complete processing chain from operating data, physical correlation, degradation trajectory to lifespan results. Attached Figure Description

[0099] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0100] Figure 1 This is an overall flowchart of a deep learning-based method for predicting the lifespan of new energy storage lithium batteries proposed in this invention.

[0101] Figure 2 This is a schematic diagram of the improved Graph WaveNet model structure proposed in this invention;

[0102] Figure 3 This is a flowchart of the improved PELT algorithm proposed in this invention. Detailed Implementation

[0103] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0104] refer to Figures 1-3 A deep learning-based method for predicting the lifespan of new energy storage lithium batteries includes the following steps:

[0105] Step 1: Receive lithium battery life prediction task information, parse the prediction object and prediction conditions, and generate a lithium battery life prediction constraint set;

[0106] Step 2: Read the operation records of the energy storage lithium battery according to the lithium battery life prediction constraint set, perform data normalization and equivalent cyclic transformation on the operation records, and generate the energy storage lithium battery operation observation sequence;

[0107] Step 3: Perform state regularization and feature encoding on the operational observation sequence of the energy storage lithium battery to generate a single-cell operational evidence sequence;

[0108] Step 4: Obtain the structural association information of the energy storage lithium battery, establish association codes based on the physical association relationships between battery cells, and generate a battery physical association matrix;

[0109] Step 5: Input the single-unit running evidence sequence and battery physical correlation matrix into the improved Graph WaveNet model. After running evidence encoding, dynamic degradation relationship construction, multi-scale degradation propagation and lifetime trajectory output, the battery degradation trajectory matrix is ​​generated.

[0110] Step 6: Input the battery degradation trajectory matrix into the improved PELT algorithm, construct the joint degradation segmentation cost based on the multi-state degradation changes, perform boundary screening, cumulative cost calculation and boundary backtracking within the continuous degradation candidate boundary range, and generate the degradation stage boundary set;

[0111] Step 7: Process the degradation stage boundary set according to the lithium battery life prediction constraint set, determine the target life termination boundary, generate the lithium battery life prediction result, and generate life prediction feedback record based on the deviation between the actual operating state and the lithium battery life prediction result.

[0112] In this embodiment, step one specifically includes:

[0113] The lithium battery life prediction task information includes a description of the prediction object, the prediction target, and the task invocation period. The description of the prediction object adopts a hierarchical coding path of "battery cluster coding - battery module coding - battery cell coding". The task invocation period is uniformly converted into a millisecond-level timestamp. If any item is missing, the task information is recorded as incomplete and the current task is terminated.

[0114] The hierarchical coding path in the description of the predicted object is matched with the energy storage lithium battery resource index. When the predicted object is a battery cell, it is included in the corresponding battery cell. When the predicted object is a battery module or battery cluster, the subordinate battery cells are collected according to the hierarchical coding path prefix, and the predicted object range is formed in ascending order of battery cell coding.

[0115] The intersection of the record interval of the running data index corresponding to the predicted object range and the task call time period is calculated. The intersection start point is the later value of the two start times, and the intersection end point is the earlier value of the two end times. When the intersection start point is earlier than the intersection end point, the running record call range is formed. When there is no valid intersection, the running data missing state is recorded and the current task is terminated.

[0116] Based on the prediction target and battery type, filter the configuration records that have the same target type and applicable battery type and whose configuration validity period covers the start of the task call period. When multiple records meet the conditions, select the configuration record with the latest effective time, and read the equivalent cyclic conversion rules, degradation state judgment rules, prediction range rules and lifespan termination criteria.

[0117] By combining the predicted object range, the operation record call range, the equivalent cycle conversion rule, the degradation state judgment rule, the prediction range rule, and the life end criterion in that order, a set of lithium battery life prediction constraints is generated, providing unified constraints for operation record call, equivalent cycle conversion, degradation state judgment, and life end boundary determination.

[0118] In this embodiment, step two specifically involves:

[0119] Batch read the energy storage lithium battery operation records according to the prediction object range and operation record call range. Sort the records with the same battery cell code in ascending order of the collection timestamp. Delete duplicate records with the same battery cell code, collection timestamp and record content. Remove records that are missing collection timestamp or current value.

[0120] The sampling period for the operation record is set to 5s, the effective sampling interval is set to 10s, the recording chain is disconnected when the time interval between adjacent regular operation records exceeds 10s, the charge and discharge capacity is not calculated across the disconnection position, the absolute value of the current is less than 0.02C and it is judged as a static state. The current continuous charge and discharge interval ends when the current direction changes or the static state begins.

[0121] Each continuous charge and discharge interval uses trapezoidal integral to calculate the charge and discharge capacity of the interval. The average value of the absolute values ​​of adjacent currents is multiplied by the corresponding time interval, and the time unit is converted to hours and accumulated item by item. The rated capacity is the capacity value that is bound to the battery type and is effective during the task call period in the equivalent cyclic conversion rule. When the rated capacity is not greater than zero, the conversion configuration is recorded as abnormal and the current task is terminated.

[0122] The equivalent cycle increment of the interval is obtained by dividing the charge and discharge capacity of the interval by twice the rated capacity. The increment is accumulated sequentially according to the collection timestamp. Each accumulated equivalent cycle value is retained to six decimal places and written into the corresponding regularized operation record to form an energy storage lithium battery operation observation sequence, providing a unified cycle scale for the regularization of single cell state and the encoding of degradation characteristics.

[0123] In this embodiment, step three specifically includes:

[0124] The equivalent cycle window is divided according to the integer part of the cumulative equivalent cycle value. The integer boundary record is assigned to the next window. The operation record of each battery cell is arranged according to the collection timestamp.

[0125] When the current direction is consistent, the direction of change of state of charge is consistent, and the change of state of charge in a single segment is not less than 10 percentage points, the charge and discharge capacity of the segment is obtained by trapezoidal integration and divided by the proportion of change of state of charge. The median of the calculation results within the window is taken as the single-cell capacity state.

[0126] When the absolute value of the current difference between adjacent records is not less than 0.05C, calculate the ratio of the absolute value of the voltage difference to the absolute value of the current difference. When there are no less than 5 valid records, take the median of the ratio as the internal resistance state of the single cell.

[0127] Calculate the median voltage and median temperature of each battery cell separately, then subtract the corresponding median within the prediction range. Form a cell state deviation code according to the order of voltage deviation and temperature deviation. The balancing occupancy is calculated as the ratio of the number of effective records of balancing control to the number of effective records within the window.

[0128] When the individual cell capacity state, individual cell internal resistance state, and individual cell state deviation code are all generated, the state trust flag is set to one. When any state is not generated, the corresponding state is written to a zero value, and the state trust flag is set to zero.

[0129] Using only battery cells with a state confidence mark of one, the median absolute deviation of the cell capacity state and the cell internal resistance state is calculated to the corresponding median. The average of the two ratios is taken, and the average value is subtracted from one. The calculation result is limited to the range of zero to one to generate the battery cluster consistency state.

[0130] Using only battery cells with a state confidence label of one, the median and median absolute deviation of each feature dimension of cell capacity state, cell internal resistance state, voltage deviation, temperature deviation, and equalization occupancy are calculated. The corresponding feature value is subtracted from the median and then divided by the median absolute deviation. When the median absolute deviation is zero, the corresponding standardization result is set to zero, and the median and median absolute deviation of each feature dimension are retained. The state confidence label is kept to be zero or one and does not participate in the standardization.

[0131] The sequence of cell operation evidence is generated by combining the cell capacity state, cell internal resistance state, cell state deviation code, equalization occupancy and state credibility label in that order and arranging them according to the cumulative equivalent cycle value. This provides a same-scale input for battery physical correlation modeling and degradation relationship learning.

[0132] In this embodiment, step four specifically includes:

[0133] Read the relevant information of the energy storage lithium battery structure within the task call period according to the scope of the predicted object. When multiple structure records are valid at the same time, select the record with the latest effective time and establish a one-to-one correspondence according to the battery cell code and the arrangement position in the cell operation evidence sequence.

[0134] The electrical structure path, hot zone identifier, and equalization channel identifier are extracted from the structure association information of the energy storage lithium battery. The electrical structure path is split into the battery cluster, battery module, and electrical branch levels. Full character consistency verification is performed on the hot zone identifier and equalization channel identifier.

[0135] Establish a relation matrix whose number of rows and columns are consistent with the number of battery cells. The row and column positions of the matrix are consistent with the arrangement positions of battery cells in the battery cell operation evidence sequence. The intersection of each row and column stores four types of binary codes: self-association, electrical association, thermal association and control association.

[0136] The main diagonal position corresponding to the same battery cell will be written with a self-association code of 1, and the rest of the association codes will be written with zero. When different battery cells belong to the same module or the same electrical branch, the electrical association code will be written with a one. When the hot zone identifiers are consistent, the hot zone association code will be written with a one. When the equalization channel identifiers are consistent, the control association code will be written with a one. If the corresponding conditions are not met, zero will be written.

[0137] When the same pair of entities simultaneously satisfies multiple association conditions, each association code is retained, and no relationship merging or numerical accumulation is performed. The same code is written in the symmetric position of the matrix.

[0138] When any structural association item is missing, the corresponding association code is written to zero. When all three types of structural association information are missing, only the self-association code of the main diagonal position is retained. After all individual pairs are encoded, the battery physical association matrix is ​​generated, which provides categorized physical association input for the construction of prior association tensors and the learning of dynamic degradation relationships.

[0139] In this embodiment, the improved Graph WaveNet model includes a running evidence encoding layer, a state graph construction layer, a multi-scale diffusion layer, and a lifetime trajectory output layer. The state graph construction layer embeds an asynchronous weakening propagation mechanism, specifically:

[0140] The improved Graph WaveNet model receives a sequence of individual cell operation evidence and a battery physical correlation matrix. The sequence of individual cell operation evidence is arranged in order of task batch, individual cell, historical position and state features. 64 consecutive equivalent cyclic windows are selected for each individual cell. Each historical position contains 6-dimensional features, including individual cell capacity state, individual cell internal resistance state, voltage deviation, temperature deviation, equalization occupancy and state credibility mark. If there are fewer than 64 historical positions, a zero value is written at the beginning of the sequence and the corresponding state credibility mark is set to zero. If there are more than 64 historical positions, the 64 positions with the larger cumulative equivalent cyclic value are retained. The row and column order of the battery physical correlation matrix is ​​consistent with the order of the individual cells in the individual cell operation evidence sequence. Each individual cell position contains a 4-dimensional binary correlation code.

[0141] The evidence coding layer uses a six-input, thirty-two-output linear projection to process the 6-dimensional state features of each historical position, and generates a 32-dimensional individual state code through the ReLU function. Positions with a state credibility marker of zero are written into a zero vector after projection. The correlation neighborhood of each battery cell is extracted based on the battery physical correlation matrix. The median of the credible individual state codes in the neighborhood is taken according to the feature dimension to generate a 32-dimensional common state code. When only self-correlation exists, the current individual state code is directly used. The individual state code and the common state code are subjected to positional difference, and layer normalization is performed along the coding dimension to generate the individual degradation code.

[0142] The state diagram construction layer uses a four-input one-output linear mapping to process the 4-dimensional binary association code, generating a priori association tensor. The 32-dimensional individual cell degenerate codes corresponding to any two battery cells are concatenated into a 64-dimensional individual cell pair code according to the feature dimension. Then, a sixty-four-input one-output linear projection is used to generate the state association tensor. Off-diagonal positions in the battery physical association matrix that do not have associations do not participate in the state relationship calculation.

[0143] The asynchronous weakening propagation mechanism is set after the prior association tensor is generated and before the dynamic degenerate adjacency matrix is ​​generated. Based on the starting order, direction of change and continuous overlap of the deviation interval of the candidate propagation single entity pair in 64 historical positions, a directed propagation gate matrix with a value range of zero to one is generated. The directed propagation gate matrix is ​​multiplied with the state association tensor position by position. Positions with a directed propagation gate value of zero or a prior association value of zero are written into the normalization mask and do not participate in the SoftMax calculation. SoftMax normalization is performed on the remaining valid positions in each row. When there are no valid off-diagonal positions, only the main diagonal relationship is retained to generate the dynamic degenerate adjacency matrix.

[0144] The multi-scale diffusion layer sets up 6 groups of gated dilated causal convolutions with a kernel length of 2 and dilation coefficients of 1, 2, 4, 8, 16 and 32 respectively. The input and output channels of each group are set to 32. One convolution result is processed by the hyperbolic tangent function, and the other convolution result is processed by the sigmoid function. The two results are multiplied position by position to generate temporal convolutional features. The receptive range formed by the 6 groups of convolutions covers 64 historical positions.

[0145] The temporal convolutional features are respectively entered into the ordinary diffusion channel and the key single-unit channel. The ordinary diffusion channel performs two-order directed graph convolution according to the dynamic degenerate adjacency matrix, mapping the state of the current battery single unit, directly related single unit and second-order related single unit to 32-dimensional ordinary diffusion features.

[0146] The key cell channel uses the median and median absolute deviation retained during state regularization to transform the cell capacity state, cell internal resistance state, and cell state deviation coding judgment conditions in the degradation state judgment rules to the corresponding standardized scale. When the median absolute deviation is zero, it is determined whether the judgment conditions are met based on the comparison result between the corresponding median and the original judgment conditions. The convolutional features corresponding to the battery cells that meet the judgment condition combinations are retained, and the positions of the remaining battery cells are written into the zero vector to generate 32-dimensional key degradation features.

[0147] The ordinary diffusion features and key degradation features are concatenated into 64-dimensional channel features according to feature dimensions. A linear mapping of 64 inputs and 32 outputs is used to restore them to 32-dimensional features. The residuals at the same position are added to the input of the current convolutional group. The outputs of each convolutional group are simultaneously written into the skip connection. After completing 6 groups of processing, the skip connection results are added position by position and processed by the ReLU function to generate 32-dimensional multi-scale degradation propagation features.

[0148] The lifetime trajectory output layer extracts the 32-dimensional multi-scale degradation propagation features corresponding to the last historical position. The single-cell state branch uses a one-dimensional convolution with a kernel length of 1 to map the 32-dimensional features of each battery cell into a 256-dimensional output code, and rearranges them according to 128 consecutive predicted equivalent cyclic positions, single-cell capacity state and single-cell internal resistance state.

[0149] The battery cluster state branch takes the median of the 32-dimensional features of all individual cells according to the feature dimensions, and then maps it to a 128-dimensional output code to generate the battery cluster consistency state corresponding to each predicted equivalent cycle position. Based on the cumulative equivalent cycle value at the end of the energy storage lithium battery operation observation sequence, starting from the first equivalent cycle after that value, 128 consecutive predicted equivalent cycle positions are determined sequentially at the interval of one equivalent cycle.

[0150] The single-cell capacity state and single-cell internal resistance state are output linearly and inversely normalized according to the corresponding median and median absolute deviation retained during state normalization. The battery cluster consistency state is limited to the range of zero to one using the Sigmoid function. The battery degradation trajectory matrix is ​​generated by combining the predicted equivalent cycle position and the battery cell arrangement order.

[0151] Both the original Graph WaveNet model and the improved Graph WaveNet model retain the gated dilated causal convolution, graph diffusion processing, residual connections, skip connections, and graph spatiotemporal joint processing structure. They both first extract time-varying features, then propagate features along node relationships, and generate continuous state results through output mapping.

[0152] The original Graph WaveNet model uses a pre-defined adjacency matrix and an adaptive adjacency matrix generated by node embedding to perform graph convolution. The node relationships do not limit the electrical structure, thermal zone relationships, and equalization control relationships of the energy storage lithium battery, nor does it describe the transmission direction of degradation state between associated battery cells. The improved Graph WaveNet model introduces a battery physical correlation matrix at the construction position of the adaptive adjacency matrix in the original model, maps the categorical physical relationships to a priori correlation tensors, and embeds an asynchronous weakening transmission mechanism in the state graph construction layer. It generates a directed transmission gating matrix according to the order of degradation deviation of associated cells, forming a dynamic degradation adjacency matrix constrained by physical relationships.

[0153] The improved Graph WaveNet model also sets up ordinary diffusion channels and key cell channels at the original model's graph diffusion processing position, so that the overall degradation propagation between related cells and the preservation of key degradation cell features can be executed in parallel. The general time series output of the original model is changed to a joint trajectory output of cell capacity state, cell internal resistance state and battery cluster consistency state.

[0154] The above processing maps the individual cell operation evidence, battery physical correlation, and degradation state transmission relationship to a continuous predicted equivalent cycle position. The generated battery degradation trajectory matrix has clear battery cell positions, state types, and predicted equivalent cycle positions, providing continuous and consistent input basis for joint degradation segment cost calculation and degradation stage boundary identification.

[0155] In this embodiment, the asynchronous weakening propagation mechanism is specifically as follows:

[0156] Read the non-zero, non-diagonal positions in the prior association tensor, determine the candidate conduction cell pairs according to the row and column positions, map the corresponding prior association values ​​to the range of zero to one using the Sigmoid function, the main diagonal position does not participate in the conduction direction determination, and set the corresponding gate value to zero when the effective historical position of any cell in the candidate conduction cell pair is less than 16.

[0157] Calculate the cell degradation code and position difference between adjacent historical positions of each battery cell. Take the median of the differences of each coding dimension at the same historical position to generate the position change. When the state confidence mark of any adjacent position is zero, mark the corresponding position change as invalid.

[0158] The median of the absolute value of the effective position change and the sum of the median absolute deviation are used as the deviation threshold. Records whose absolute value of position change exceeds the deviation threshold and have no less than 3 consecutive historical positions are grouped into deviation intervals. When the cumulative result of position change within the interval is greater than zero, it is determined as a positive change; when it is less than zero, it is determined as a negative change; when it is equal to zero, the corresponding deviation interval is deleted.

[0159] Read the earliest deviation interval of the starting position of the two battery cells in the candidate conductive cell pair respectively. When the two deviation intervals change in the same direction and have no less than 2 overlapping positions, the battery cell with the earlier starting position is determined as the conductive end and the other battery cell is determined as the response end. When the starting positions are the same, the changing directions are inconsistent or no effective overlap is formed, the corresponding gate value is set to zero.

[0160] Divide the number of overlapping positions by the number of positions in the shorter deviation interval to generate a continuous matching value. Add one to the interval between the starting positions of the two deviation intervals and take the reciprocal to generate a hysteresis decay value.

[0161] Multiply the mapped prior association value, continuous matching value, and hysteresis decay value to generate a directed transmission gating value. Write the value into the corresponding position according to the row of the transmission end and the column of the response end, and write zero value into the remaining positions to generate a directed transmission gating matrix.

[0162] The directed transmission gating matrix constrains the state association tensor by transmission direction, continuous overlap degree and initial lag distance, so that the dynamic degenerate adjacency matrix retains the degenerate transmission path with physical association, continuous deviation basis and clear temporal sequence relationship.

[0163] In this embodiment, step six specifically includes:

[0164] The improved PELT algorithm receives the battery degradation trajectory matrix and reads the cell capacity state and cell internal resistance state of each battery cell in ascending order according to the predicted equivalent cycle position.

[0165] Read the battery cluster consistency status corresponding to each predicted equivalent cycle position, process 128 consecutive predicted equivalent cycle positions in a single run, and record the trajectory data as abnormal and terminate the current algorithm processing when any predicted equivalent cycle position lacks a corresponding status.

[0166] The decrease in the unit capacity state is generated by subtracting the unit capacity state at the current position from the unit capacity state at the previous predicted equivalent cycle position, and the increase in the unit internal resistance state is generated by subtracting the unit internal resistance state at the previous predicted equivalent cycle position from the unit internal resistance state at the current position.

[0167] The decrease in battery cluster consistency state is generated by subtracting the battery cluster consistency state at the current position from the battery cluster consistency state at the previous predicted equivalent cycle position. An increase in any of the three types of values ​​indicates aggravated degradation.

[0168] The decrease in single-cell capacity state, the increase in single-cell internal resistance state, and the decrease in battery cluster consistency state are retained as the same-direction degradation sequence.

[0169] For each individual cell, the median and median absolute deviation are calculated for the cell capacity state decline sequence and the cell internal resistance state rise sequence, respectively. The same processing is performed on the cell cluster consistency state decline sequence. Each state change is subtracted from the corresponding median and then divided by the median absolute deviation. When the median absolute deviation is zero, the corresponding standardization result is set to zero to generate a standardized degradation sequence.

[0170] For any candidate precursor position and the current candidate boundary, read the standardized degradation sequence from the next position after the candidate precursor position to the current candidate boundary, and calculate the interval mean of the two types of standardized state changes for each battery cell and the interval mean of the standardized state change for battery cluster consistency.

[0171] The difference between the state change at each position in the standardized degradation sequence and the mean of the corresponding interval is squared, and then accumulated sequentially according to the predicted equivalent cycle position, cell, and state type to generate the joint degradation segment cost.

[0172] The difference between the state change at each location and the mean of the corresponding interval is squared, and then accumulated sequentially according to the predicted equivalent cycle location, battery cell and state type to generate the joint degradation segment cost.

[0173] The continuous degradation candidate boundary rule is used to process the same-direction degradation sequence. The continuous degradation candidate boundary range is sorted in ascending order according to the predicted equivalent cycle position and duplicate positions are deleted. The first and last predicted equivalent cycle positions are added to the continuous degradation candidate boundary range. If the first and last positions already exist, they are not written again.

[0174] Set the cumulative cost of the first predicted equivalent cycle position to zero and write it into the candidate predecessor set. Read the remaining candidate boundaries in order of position. Set the segmented penalty value to the product of the number of degenerate state types and the natural logarithm of the number of predicted equivalent cycle positions.

[0175] For the current candidate boundary, calculate the sum of the cumulative cost, the corresponding joint degenerate segment cost, and the segment penalty value for each candidate predecessor position. Select the candidate predecessor position with the smallest sum. If the sums are the same, select the candidate predecessor position with the smaller predicted equivalent cycle position. Write the selected position and the corresponding sum into the predecessor index and the cumulative cost, respectively.

[0176] For candidate predecessor positions that are not selected, calculate the sum of the corresponding cumulative cost and the joint degenerate segment cost extending to the current candidate boundary. If the calculation result is not less than the cumulative cost of the current candidate boundary, delete the corresponding position from the candidate predecessor set and then write the current candidate boundary into the candidate predecessor set.

[0177] Starting from the last predicted equivalent cycle position, trace back to the first predicted equivalent cycle position in reverse order according to the predecessor index. Arrange the candidate boundaries obtained by tracing in ascending order according to the predicted equivalent cycle position. After deleting duplicate positions, generate the degenerate stage boundary set.

[0178] Both the original PELT algorithm and the improved PELT algorithm retain the structure of adding segmented cost, segmented penalty, dynamic programming, candidate predecessor pruning and predecessor index backtracking. Both determine the segmented boundary by comparing cumulative costs and delete candidate predecessor positions that no longer have the possibility of minimum cost.

[0179] The original PELT algorithm uses a general segmented cost for boundary search at all sequence positions, without limiting whether there is a basis for continuous degradation at the candidate positions, and without constructing a joint cost in the same direction for the individual cell capacity state, the individual cell internal resistance state, and the battery cluster consistency state.

[0180] The improved PELT algorithm adds the construction of the continuous degradation candidate boundary range before the dynamic programming process of the original algorithm. At the point where the segmented cost is calculated in the original algorithm, the general segmented cost is replaced with the joint degradation segmented cost. At the point where the candidate predecessor is updated and pruning is performed, the search range is limited to the continuous degradation candidate boundary range, so that the multi-state degradation change, candidate boundary screening and cumulative cost calculation form a continuous processing chain.

[0181] The above processing completes joint cost calculation, candidate precursor pruning, and boundary backtracking within candidate locations with continuous degradation basis. The generated degradation stage boundary set can exclude isolated state fluctuations and provide a continuous and traceable stage division basis for determining the target lifetime termination boundary.

[0182] In this embodiment, the specific rules for the candidate boundary of continuous degradation are as follows:

[0183] Read the same-direction degradation sequence according to the predicted equivalent cycle position, take the median of the decrease in cell capacity state and the increase in cell internal resistance state for each cell at the same position, and read the corresponding decrease in cell cluster consistency state. Form three types of position degradation in the order of decrease in cell capacity state, increase in cell internal resistance state and decrease in cell cluster consistency state.

[0184] When at least two of the three types of position degradation values ​​are greater than zero, the corresponding position is marked as a unidirectional degradation position. State changes with values ​​equal to zero are not included in the degradation category. When the unidirectional degradation position is empty, an empty continuous degradation candidate boundary range is generated.

[0185] Calculate the sum of the degradation amounts of the three types of positions corresponding to each degradation position in the same direction. Use the sum of the median and the median absolute deviation of all degradation sums as the threshold for continuous degradation. Retain the degradation positions in the same direction whose degradation sum is not less than the threshold for continuous degradation.

[0186] The continuity of the retained positions is checked according to the predicted equivalent cyclic positions. Adjacent retained positions are merged into persistent degradation intervals. Isolated positions that are not continuous with the previous or next position are deleted. The start and end positions of each persistent degradation interval are written into the candidate boundary.

[0187] For the middle position within the continuous degradation interval, calculate the first difference between adjacent degradation sums in turn, and then calculate the absolute value of the difference between two adjacent first differences.

[0188] The median of all absolute values ​​within the interval and the sum of the median absolute deviation are used as the turning point threshold. The intermediate predicted equivalent cyclic position corresponding to the absolute value not less than the turning point threshold is written into the candidate boundary.

[0189] When the continuous degradation interval contains only two positions, only the start and end positions of the interval are retained. All candidate boundaries are arranged in ascending order according to the predicted equivalent cyclic positions and duplicate positions are deleted to generate the range of candidate boundaries for continuous degradation. This limits the boundary search positions with continuous degradation basis and state transition characteristics for improving the PELT algorithm.

[0190] In this embodiment, step seven specifically includes:

[0191] The equivalent cycle position is limited according to the prediction range rule, and the boundary of the degradation stage within the limited range is retained. Boundaries outside the limited range are not included in the lifespan termination judgment.

[0192] The two types of cell termination conditions are defined as: the cell capacity not exceeding 80% of the rated capacity and the cell internal resistance not less than 1.50 times the reference internal resistance.

[0193] The cluster consistency state is defined as not being greater than 0.85 as the cluster termination condition. When the number of battery cells corresponding to the two types of individual cell termination conditions reaches 5% of the total number of battery cells, the quantity condition is determined to be met. The calculation result is rounded up and is at least one battery cell.

[0194] Starting from the boundary of each degradation stage, three types of states are read continuously. When the quantity condition and the cluster termination condition are satisfied for three consecutive predicted equivalent cycle positions, the corresponding degradation stage boundary is marked.

[0195] The boundary with the smallest position is determined as the target lifetime termination boundary. If no boundary is formed, the lifetime termination boundary is recorded as not reached and the current task is terminated.

[0196] Read the predicted equivalent cycle value corresponding to the target lifespan termination boundary, subtract the cumulative equivalent cycle value at the end of the energy storage lithium battery operation observation sequence, generate the predicted remaining equivalent cycle value, write the zero value when the calculation result is less than zero, combine the prediction object, target lifespan termination boundary, predicted equivalent cycle value, predicted remaining equivalent cycle value and battery degradation trajectory matrix to generate lithium battery lifespan prediction result;

[0197] When the actual operating state does not meet the lifespan termination criterion, the predicted equivalent cycle position closest to the actual cumulative equivalent cycle value is selected. If the distances are the same, the position with the smaller value is selected. The difference between the three types of actual states and the corresponding predicted states is calculated. When the actual operating state meets the lifespan termination criterion, the difference between the actual cumulative equivalent cycle value and the predicted equivalent cycle value is calculated.

[0198] By combining the predicted object, the corresponding difference, the actual cumulative equivalent cycle value, and the target life termination boundary, a life prediction feedback record is generated, providing a consistent data basis for life prediction deviation tracking and prediction result verification.

[0199] Example 1: To verify the feasibility of the present invention in practice, the present invention was applied to the prediction of the lifespan of lithium iron phosphate battery clusters in a new energy storage station. Historical operating data of 18 battery clusters of the same model were selected, of which 12 battery clusters were used to improve the training of the Graph WaveNet model and 6 battery clusters were used for verification. There was no overlap between the training and verification objects. Each battery cluster contained 240 battery cells with a rated capacity of 280Ah. The cumulative equivalent cycle value of the verified battery clusters was distributed between 2814.7 and 3162.3.

[0200] Data normalization, equivalent cycle transformation, and state feature encoding are performed on the operation records of energy storage lithium batteries to generate a sequence of single-cell operation evidence. A battery physical correlation matrix is ​​generated based on electrical structure path, hot zone identifier, and equalization channel identifier. Data corresponding to 64 consecutive equivalent cycle windows are input into an improved Graph WaveNet model to generate a battery degradation trajectory matrix corresponding to 128 predicted equivalent cycle positions. Then, an improved PELT algorithm is used to generate a degradation stage boundary set.

[0201] One of the verified battery clusters has a current cumulative equivalent cycle value of 2978.4, and the degradation stage boundaries are 3004.4, 3038.4, and 3061.4. At the position corresponding to 3061.4, there are 14 battery cells with a single-cell capacity state not greater than 224Ah and 16 battery cells with a single-cell internal resistance state not less than 1.50 times the reference internal resistance, both reaching 5% of the total number of battery cells. The battery cluster consistency state is 0.832. The three conditions are met for three consecutive predicted equivalent cycle positions. Therefore, 3061.4 is determined as the target lifetime termination boundary, and the predicted remaining equivalent cycle value is 83.0. The battery cluster actually meets the lifetime termination criterion when it runs for 3068.2 cumulative equivalent cycles. The prediction error of the target lifetime termination boundary is 6.8 equivalent cycles.

[0202] Six verification battery clusters were processed using the original Graph WaveNet model combined with the original PELT algorithm, the improved Graph WaveNet model combined with the original PELT algorithm, the original Graph WaveNet model combined with the improved PELT algorithm, and the method of this invention, respectively. The normalized average absolute error was used for the single cell capacity state, the average relative error was used for the single cell internal resistance state, and the average absolute error between the predicted equivalent cycle value and the actual equivalent cycle value was used for the degradation stage boundary and the remaining lifetime. The verification results are shown in Table 1.

[0203] Table 1. Validation results of different lifetime prediction methods

[0204] Combination of the original Graph WaveNet model and the original PELT algorithm 2.84% 8.17% 14.6 19.3 Improved Graph WaveNet model combined with the original PELT algorithm 1.72% 5.36% 10.9 14.2 Combination of the original Graph WaveNet model and the improved PELT algorithm 2.84% 8.17% 7.1 11.4 Method of the present invention 1.72% 5.36% 5.2 8.8

[0205] As shown in Table 1, when the PELT algorithm remains unchanged, after adopting the improved Graph WaveNet model, the normalized average absolute error of the single-cell capacity state changes from 2.84% to 1.72%, the average relative error of the single-cell internal resistance state changes from 8.17% to 5.36%, and the average absolute error of the degradation stage boundary changes from 14.6 equivalent cycles to 10.9 equivalent cycles. This indicates that the battery physical correlation matrix, asynchronous weakening conduction mechanism, and ordinary diffusion channels and key single-cell channels can form a battery degradation trajectory that is more consistent with the actual degradation process.

[0206] With the Graph WaveNet model unchanged, the prediction errors of the single-cell capacity state and the single-cell internal resistance state remain unchanged after adopting the improved PELT algorithm. The mean absolute error of the degradation stage boundary changes from 14.6 equivalent cycles to 7.1 equivalent cycles, and the mean absolute error of the remaining lifetime changes from 19.3 equivalent cycles to 11.4 equivalent cycles. This shows that the joint degradation segmentation cost and the range of continuous degradation candidate boundaries can exclude isolated fluctuation positions that lack continuous degradation basis.

[0207] This invention employs both an improved Graph WaveNet model and an improved PELT algorithm, resulting in a mean absolute error of 5.2 equivalent cycles at the boundary of the degradation stage and a mean absolute error of 8.8 equivalent cycles for the remaining lifetime. These represent reductions of 64.4% and 54.4% respectively compared to the original combination of the Graph WaveNet model and the original PELT algorithm.

[0208] Verification results show that the present invention can form a continuous battery degradation trajectory matrix under a unified equivalent cycle scale, and determine the target lifespan termination boundary based on the continuous changes of multiple states, generating lithium battery lifespan prediction results with clear prediction objects, prediction locations and deviation records.

[0209] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting the lifespan of new energy storage lithium batteries based on deep learning, characterized in that, Includes the following steps: Step 1: Receive lithium battery life prediction task information, parse the prediction object and prediction conditions, and generate a lithium battery life prediction constraint set; Step 2: Read the operation records of the energy storage lithium battery according to the lithium battery life prediction constraint set, perform data normalization and equivalent cyclic transformation on the operation records, and generate the energy storage lithium battery operation observation sequence; Step 3: Perform state regularization and feature encoding on the operational observation sequence of the energy storage lithium battery to generate a single-cell operational evidence sequence; Step 4: Obtain the structural association information of the energy storage lithium battery, establish association codes based on the physical association relationships between battery cells, and generate a battery physical association matrix; Step 5: Input the single-unit running evidence sequence and battery physical correlation matrix into the improved Graph WaveNet model. After running evidence encoding, dynamic degradation relationship construction, multi-scale degradation propagation and lifetime trajectory output, the battery degradation trajectory matrix is ​​generated. Step 6: Input the battery degradation trajectory matrix into the improved PELT algorithm, construct the joint degradation segmentation cost based on the multi-state degradation changes, perform boundary screening, cumulative cost calculation and boundary backtracking within the continuous degradation candidate boundary range, and generate the degradation stage boundary set; Step 7: Process the degradation stage boundary set according to the lithium battery life prediction constraint set, determine the target life termination boundary, generate the lithium battery life prediction result, and generate life prediction feedback record based on the deviation between the actual operating state and the lithium battery life prediction result.

2. The method for predicting the lifespan of new energy storage lithium batteries based on deep learning according to claim 1, characterized in that, Step one specifically involves: Read the prediction object description in the lithium battery life prediction task information, compare the object code in the prediction object description with the battery code in the energy storage lithium battery resource index item by item, and determine the resource record with the same battery code as the prediction object. When the prediction target is a single battery cell, the resource record of the corresponding battery cell is included in the prediction target range. When the prediction object is a battery module or battery cluster, the battery cells whose resource level paths belong to the corresponding paths of the prediction object are included in the prediction object scope. Read the running data index corresponding to the predicted object range, retain the running data index where the bound battery code belongs to the predicted object range and the recording time overlaps with the task call period, and determine the overlapping period as the running record call range; Read the prediction target from the lithium battery life prediction task information, and filter the configuration records in the life prediction configuration library that have the same target type and the same applicable battery type according to the prediction target and the battery type corresponding to the prediction object. Extract equivalent cyclic conversion rules, degradation state determination rules, prediction range rules, and lifetime termination criteria; The prediction object range, the running record call range, and the extracted rules are combined according to the preset field order to generate a lithium battery life prediction constraint set.

3. The method for predicting the lifespan of new energy storage lithium batteries based on deep learning according to claim 1, characterized in that, Step two specifically involves: Read the energy storage lithium battery operation records according to the prediction object range and operation record call range in the lithium battery life prediction constraint set; Sort the data in ascending order by the acquisition timestamp, delete duplicate records where the prediction object, acquisition timestamp, and record content are all the same, and remove records that are missing acquisition timestamps or current values ​​to generate a regular operation record. According to the effective sampling interval in the equivalent cycle conversion rule, check the adjacent regular operation records, disconnect the adjacent records whose time interval exceeds the effective sampling interval, and divide the continuous charging and discharging interval according to the change of current direction. For each continuous charge and discharge interval, calculate the product of the average absolute value of the current of adjacent regular operation records and the time interval converted to hours; The products within the interval are summed to generate the charge and discharge amount for that interval. Divide the interval charge / discharge capacity by twice the rated capacity bound in the equivalent cycle conversion rule to generate the interval equivalent cycle increment; Based on the equivalent cyclic increment of the accumulated interval according to the collection timestamp, the cumulative results are written into the corresponding regularized operation record, and the regularized operation records with the cumulative results are arranged according to the collection timestamp to generate the energy storage lithium battery operation observation sequence.

4. The method for predicting the lifespan of new energy storage lithium batteries based on deep learning according to claim 1, characterized in that, Step three specifically involves: The equivalent cycle window is divided according to the integer range of the cumulative equivalent cycle value in the energy storage lithium battery operation observation sequence, and the operation records of the same battery cell in each equivalent cycle window are arranged according to the collection timestamp. From the operation records where the current direction is consistent and the state of charge changes continuously, the product of the average absolute value of the current of adjacent records and the time interval is accumulated. The accumulated result is divided by the corresponding state of charge change ratio, and the median of the calculation results within the same equivalent cyclic window is taken to generate the single-cell capacity state. For adjacent operating records with non-zero current differences, calculate the ratio of the absolute value of the voltage difference to the absolute value of the current difference between the individual units, and take the median of each ratio to generate the internal resistance status of the individual unit. Calculate the median voltage and median temperature of each battery cell within the same equivalent cycle window, subtract the corresponding median within the prediction range from the median voltage and median temperature, and combine them to generate a cell state deviation code. Calculate the ratio of the number of valid records in the balance control to the number of valid records in the window, and generate the balance occupancy. When the individual unit capacity state, individual unit internal resistance state, and individual unit state deviation coding are all generated, the state trust flag is set to one; when any state is not generated, the state trust flag is set to zero. Based on the feature dimension, robust standardization is performed using the median and median absolute deviation of the cell capacity state, cell internal resistance state, cell state deviation code, and balanced occupancy of each battery cell within the prediction target range. When the median absolute deviation is zero, the corresponding standardization result is set to zero. The standardized unit capacity state, unit internal resistance state, unit state deviation code, and balanced occupancy are combined with the state credibility marker in a preset order and arranged according to the cumulative equivalent cyclic value to generate a unit operation evidence sequence.

5. The method for predicting the lifespan of new energy storage lithium batteries based on deep learning according to claim 1, characterized in that, Step four specifically involves: Read the energy storage lithium battery structure association information according to the prediction object range in the lithium battery life prediction constraint set, and extract the electrical structure path, hot zone identifier and equalization channel identifier corresponding to each battery cell from the structure association information. The battery cells are paired up according to the order of the battery cells in the single cell operation evidence sequence; when the electrical structure paths of two battery cells point to the same module or the same electrical branch, an electrical association code is written in the corresponding single cell pair; when the hot zone identifiers are consistent, a hot zone association code is written. When the equalization channel identifiers are consistent, the control association code is written; when the corresponding judgment condition is not met, the corresponding unassociated code is written, and a single-entity pair association code is generated. According to the arrangement position of each battery cell in the cell operation evidence sequence, the cell pair association code is written into the corresponding row and column position of the matrix, and the self-association code is written into the main diagonal position. When structural association information is missing, only the self-association code corresponding to the main diagonal position is retained to generate the battery physical association matrix.

6. The method for predicting the lifespan of new energy storage lithium batteries based on deep learning according to claim 1, characterized in that, The improved Graph WaveNet model includes a running evidence encoding layer, a state graph construction layer, a multi-scale diffusion layer, and a lifetime trajectory output layer. The state graph construction layer embeds an asynchronous weakening propagation mechanism. The operational evidence coding layer divides the individual operational evidence sequence according to a set historical window, performs linear projection on the individual operational evidence within the window, and generates individual state codes. Based on the battery physical correlation matrix, the states of the associated neighborhood are collected, and a common state code is generated by median merging. The difference between the co-position of the individual state code and the common state code is calculated, and layer normalization is performed to generate the individual degradation code. The state graph construction layer performs a linear mapping on the battery physical correlation matrix to generate a priori correlation tensor. The state-related tensor is generated by concatenating the degenerate codes of each pair of individuals and performing linear projection. The asynchronous weakening propagation mechanism processes prior correlation tensors and single-entity degenerate encodings to generate directed propagation gating matrices; The directed transmission gate matrix is ​​multiplied element-wise with the state association tensor, and row normalization is performed using the SoftMax function to generate a dynamically degenerate adjacency matrix. The multi-scale diffusion layer uses gated dilated causal convolution to process the monomer degradation encoding, and inputs the convolution results into the ordinary diffusion channel and the key monomer channel respectively. The ordinary diffusion channel performs directed graph convolution according to the dynamic degenerate adjacency matrix, while the key single-unit channel compares the single-unit capacity state, single-unit internal resistance state, and single-unit state deviation encoding with the corresponding judgment conditions in the degenerate state judgment rules item by item. Battery cells that meet the judgment criteria are marked as key degraded cells. The convolution features corresponding to the key degraded cells are retained, and zero values ​​are written at the positions of the remaining battery cells. The results from the two channels are concatenated and linearly mapped, and multi-scale degradation propagation features are generated through residual connections and skip connections. The lifetime trajectory output layer uses one-dimensional convolution to process multi-scale degradation propagation features, maps and generates the single-cell capacity state, single-cell internal resistance state and battery cluster consistency state corresponding to each predicted equivalent cycle position, and combines them according to the predicted equivalent cycle position and the battery single-cell arrangement order to generate the battery degradation trajectory matrix.

7. The method for predicting the lifespan of new energy storage lithium batteries based on deep learning according to claim 6, characterized in that, The asynchronous weakening propagation mechanism is specifically as follows: Read the non-zero correlation positions in the prior correlation tensor to determine candidate transmission singleton pairs; According to the positional order in the historical window, calculate the difference between adjacent positions of the individual cell degradation code of each battery cell, and take the median of the difference of each coding dimension at the same position to generate the positional change. The median of the absolute value of the position change and the sum of the median absolute deviation are used as the deviation threshold. Continuous positions whose absolute value of position change exceeds the deviation threshold are grouped into deviation intervals, and the direction of change is determined based on the sign of the cumulative result of the position change within the deviation interval. Compare the first deviation intervals of the two cells in the candidate conductive cell pair; When the two deviation intervals change in the same direction and overlap, the battery cell with the earlier starting position of the deviation interval is determined as the conduction end, and the other battery cell is determined as the response end. If two deviation intervals have the same starting position, different directions of change, or no overlap, the corresponding gate value is set to zero. Calculate the ratio of the overlap length of the two deviation intervals to the length of the shorter deviation interval to generate a continuous matching value; calculate the reciprocal of the interval between the starting positions of the two deviation intervals plus one to generate a hysteresis decay value. Multiply the corresponding prior correlation value, continuous matching value and hysteresis decay value to generate a directed conduction gating value, and write it into the matrix position pointing from the conduction end to the response end according to the battery cell arrangement order, and write zero value into the other positions to generate a directed conduction gating matrix.

8. The method for predicting the lifespan of new energy storage lithium batteries based on deep learning according to claim 1, characterized in that, Step six specifically involves: The battery degradation trajectory matrix is ​​read according to the predicted equivalent cycle position. The decrease in single cell capacity state, the increase in single cell internal resistance state, and the decrease in battery cluster consistency state are calculated respectively for adjacent predicted equivalent cycle positions. The three types of state changes are uniformly converted into a unidirectional degradation sequence in which the numerical value increases to indicate the aggravation of degradation. Robust standardization is performed using the median and median absolute deviation according to the state type. When the median absolute deviation is zero, the corresponding standardization result is set to zero. For any two candidate boundaries, calculate the squared difference between the decrease in the capacity state of each battery cell, the increase in the internal resistance state of each battery cell and the mean of the corresponding interval, and calculate the squared difference between the decrease in the uniformity state of the battery cluster and the mean of the corresponding interval. The squared differences corresponding to each individual battery cell and the squared differences corresponding to the consistency state of the battery cluster are summed to generate the joint degradation segmentation cost. The continuous degradation candidate boundary rule is invoked to process the same-direction degradation sequence, generate the continuous degradation candidate boundary range, and the first predicted equivalent cycle position and the last predicted equivalent cycle position are added to the continuous degradation candidate boundary range; Set the cumulative cost of the first predicted equivalent cycle position to zero, write the first predicted equivalent cycle position into the candidate predecessor set, and read the candidate boundaries within the continuously degenerate candidate boundary range in order of position. For the current candidate boundary, calculate the cumulative cost of each candidate predecessor position, the sum of the joint degradation segment cost from the candidate predecessor position to the current candidate boundary and the segment penalty value, where the segment penalty value is the product of the number of degradation state types and the natural logarithm of the number of predicted equivalent cycle positions. Select the candidate predecessor position with the smallest sum, and write the corresponding sum and candidate predecessor position into the cumulative cost and predecessor index of the current candidate boundary, respectively. For candidate predecessor positions that are not selected, calculate the sum of the corresponding cumulative cost and the joint degenerate segment cost extending to the current candidate boundary. Remove candidate predecessor positions whose calculation results are not less than the cumulative cost of the current candidate boundary from the candidate predecessor set, and write the current candidate boundary into the candidate predecessor set. Starting from the last predicted equivalent cycle position, candidate boundaries are traced in reverse order according to the predecessor index. The traced candidate boundaries are then arranged in ascending order according to the predicted equivalent cycle position to generate the degradation stage boundary set.

9. The method for predicting the lifespan of a new energy storage lithium battery based on deep learning according to claim 8, characterized in that, The specific rules for the candidate boundary of continuous degradation are as follows: Read the same-direction degradation sequence according to the predicted equivalent cycle position, take the median of the decrease in cell capacity state and the increase in cell internal resistance state for each cell at the same position, and read the corresponding decrease in cell cluster consistency state to generate three types of position degradation. When at least two of the three types of positional degradation are greater than zero, the corresponding position is marked as a position of degradation in the same direction. When the same-direction degradation position is empty, an empty candidate boundary range for continuous degradation is generated; when the same-direction degradation position is not empty, the sum of the degradation amount of the three types of positions at each same-direction degradation position is calculated, and the sum of the median and the median absolute deviation of the corresponding sum of all same-direction degradation positions is used as the threshold for continuous degradation judgment. Retain the same-direction degradation positions whose sum is not less than the continuous degradation judgment threshold, merge the retained positions adjacent to the predicted equivalent cycle positions into continuous degradation intervals, and delete isolated positions that are not continuous with the previous or next position. Write the start and end positions of each continuously degenerate interval into the candidate boundary. For each intermediate position within the continuously degenerate interval, calculate the first difference of the degenerate sum of adjacent positions, and then calculate the absolute value of the difference between adjacent first differences. The median and the sum of the absolute values ​​of each absolute value within the interval are used as the turning point threshold. Positions with absolute values ​​not less than the turning point threshold are written into candidate boundaries. The positions are sorted according to the predicted equivalent cyclic positions and duplicate positions are deleted to generate the range of candidate boundaries for continuous degradation.

10. The method for predicting the lifespan of a new energy storage lithium battery based on deep learning according to claim 1, characterized in that, Step seven specifically involves: The equivalent cycle position of the prediction is limited according to the prediction range rules in the lithium battery life prediction constraint set, and the degradation stage boundary set is read within the limited equivalent cycle position of the prediction. Extract the individual cell capacity state, individual cell internal resistance state, and battery cluster consistency state corresponding to the consecutive positions from the boundary of each degradation stage in the battery degradation trajectory matrix. The capacity status and internal resistance status of each cell are compared with the corresponding threshold in the lifespan termination criterion. The number of cell cells that meet the corresponding threshold is counted, and the consistency status of the battery cluster is compared with the corresponding threshold. Count the consecutive positions where the number of individual cells and the consistency of cell clusters in both types of batteries reach the life end criterion, and mark the boundary of the degradation stage where the number of consecutive positions reaches the life end criterion. The marked degradation stage boundaries are arranged in ascending order according to the predicted equivalent cycle positions, and the first marked degradation stage boundary is determined as the target lifetime termination boundary. Read the predicted equivalent cycle value corresponding to the target lifespan termination boundary, and read the cumulative equivalent cycle value corresponding to the last position of the energy storage lithium battery operation observation sequence. Calculate the difference between the predicted equivalent cycle value and the cumulative equivalent cycle value to generate the predicted remaining equivalent cycle value. The prediction object, target life end boundary, predicted equivalent cycle value, predicted remaining equivalent cycle value and battery degradation trajectory matrix are combined to generate lithium battery life prediction results. Read the actual running state of the predicted object, and extract the actual cumulative equivalent loop value and the corresponding three types of actual states from the actual running state; When the actual operating state does not meet the lifespan termination criterion, the three types of predicted states corresponding to the actual cumulative equivalent cycle value are extracted from the battery degradation trajectory matrix in the lithium battery lifespan prediction results, and the difference between the three types of actual states and the corresponding predicted states is calculated. When the actual operating condition meets the end-of-life criterion, calculate the difference between the actual cumulative equivalent cycle value and the predicted equivalent cycle value; The prediction object, corresponding difference, actual cumulative equivalent cyclic value, and target lifetime termination boundary are combined to generate a lifetime prediction feedback record.

Images