A lithium battery remaining capacity prediction method and system

Abstract

The application discloses a lithium battery residual capacity prediction method and system, and is applied to the lithium battery residual capacity prediction field. The method comprises the following steps: acquiring a running parameter set of a lithium battery under a charging and discharging working condition within a preset period of time, wherein the running parameter set at least comprises voltage data, capacity data and SOC data; performing interval division on the SOC data to obtain a plurality of SOC intervals; inputting the running parameter set into a self-adaptive block extraction network model to perform block extraction, obtaining block features, performing multi-scale coding on the block features to obtain a time sequence feature vector; respectively calculating statistical features and local dynamic features of the running parameter set in each SOC interval and performing fusion to obtain a first feature; and after weighting fusion of the time sequence feature vector and the first feature to obtain a second feature, inputting the second feature into a prediction network to perform prediction, obtaining a capacity prediction value, and obtaining the residual capacity of the lithium battery after inverse normalization processing. The application can effectively improve the precision and robustness of lithium battery residual capacity prediction.

Description

Technical Field

[0001] This invention relates to the field of lithium battery remaining capacity prediction, and in particular to a method and system for predicting lithium battery remaining capacity. Background Technology

[0002] Lithium-ion battery remaining capacity prediction is one of the core functions of a battery management system. It infers the current health status by analyzing battery operating data, providing a basis for battery maintenance, replacement, and secondary utilization. Currently, lithium-ion battery capacity prediction technology is in a stage of widespread application of deep learning and coordinated development with traditional algorithms. Mainstream solutions mainly include methods based on traditional machine learning, deep learning, and charging integration algorithms.

[0003] In data-driven model prediction, existing technologies typically input time-series data from a complete charge-discharge cycle into recurrent neural networks such as Long Short-Term Memory (LSTM) networks, capturing temporal dependencies through gating units. However, such methods have limited ability to capture long-distance dependencies, making it difficult to simultaneously address the modeling requirements of local capacity regeneration fluctuations and global long-term degradation trends. Furthermore, some existing technologies employ standard Transformer models for sequence modeling but fail to optimize for balancing local temporal correlations and global dependencies in battery data. Regarding mechanism-related integral algorithms, existing technologies divide fixed SOC intervals by obtaining the SOC (State of Charge) - OCV (Open Circuit Voltage) curve, calculate the integrated capacity after slicing the charging cycle data, filter the baseline slice and fill in missing values, and finally linearly expand to obtain the actual capacity. However, this approach fails to consider the decreased interval adaptability caused by the shift in inflection points of the charge-discharge curves during battery aging, making fixed interval divisions difficult to adapt to the dynamic changes in curve morphology throughout the entire lifespan.

[0004] In terms of feature fusion and network structure, existing technologies mostly use fixed weights or simple splicing methods to fuse multi-source features; the prediction networks are mostly shallow fully connected or basic residual structures, and have not been customized for nonlinear degradation of battery capacity, regeneration mutations and data noise. Deep networks are prone to gradient vanishing and have limited generalization ability in complex scenarios. In summary, current lithium battery remaining capacity prediction methods are all difficult to accurately predict the remaining capacity of current lithium batteries. Summary of the Invention

[0005] This invention provides a method and system for predicting the remaining capacity of lithium batteries, in order to solve the technical problem that existing lithium battery remaining capacity prediction methods do not fully mine the deep feature information in the operating data, resulting in low prediction accuracy.

[0006] To address the aforementioned technical problems, embodiments of the present invention provide a method for predicting the remaining capacity of a lithium battery, comprising: Obtain a set of operating parameters of a lithium battery within a preset time period under charging or discharging conditions. The set of operating parameters consists of preprocessed valid time-series data, including at least voltage data, capacity data, and SOC data. The SOC data is divided into intervals to obtain multiple SOC intervals; The set of running parameters is input into the adaptive block extraction network model for block extraction. Multiple running parameters are extracted in blocks to obtain block features. The block features are then encoded at multiple scales to obtain a temporal feature vector. Calculate the basic and higher-order statistical characteristics of the set of operating parameters within each SOC interval; Each local dynamic feature is fused with the corresponding basic statistical features and higher-order statistical features to obtain the first feature. Each local dynamic feature is obtained by extracting local temporal dependencies by inputting the set of running parameters into a gated recurrent neural network. Based on the set of operating parameters corresponding to each SOC interval, construct the corresponding state vector, process each state vector to obtain the first weight corresponding to the time series feature vector, and then perform weighted fusion of the time series feature vector and the first feature according to the first weight to obtain the second feature. The second feature is input into the prediction network for prediction to obtain the capacity prediction value. The capacity prediction value is then de-standardized to obtain the remaining capacity of the lithium battery.

[0007] As one preferred approach, the set of operating parameters consists of preprocessed valid time-series data, wherein the preprocessing steps are as follows: Obtain the set of original operating parameters of the lithium battery within a preset time period under charging or discharging conditions; Remove the data belonging to the first and last loops from the original set of running parameters to obtain the set of valid running parameters; Using the Laida criterion, the data sequences in the effective operating parameter set are denoised to obtain the corresponding effective data sequences; Each valid data sequence is unified into a sequence of a preset length to obtain the corresponding valid time series data.

[0008] As one preferred approach, the SOC data is divided into intervals to obtain multiple SOC intervals, including: The voltage curve is obtained by fitting and analyzing the voltage data and SOC data within a preset time period, and the capacity curve is obtained by fitting and analyzing the capacity data and SOC data within a preset time period. The voltage curve is subjected to first-order difference calculation to obtain the voltage difference sequence, and the capacity curve is subjected to first-order difference calculation to obtain the capacity difference sequence. Calculate the mean and standard deviation of the voltage differential sequence and the capacity differential sequence, respectively; Based on the mean and standard deviation of the voltage difference sequence, determine the starting point of the voltage plateau and the polarization inflection point corresponding to the voltage difference sequence; The inflection point of capacity regeneration is determined based on the mean and standard deviation of the capacity difference sequence; Based on the SOC values ​​corresponding to the voltage plateau start point, polarization inflection point, and capacity regeneration inflection point as boundary points, the SOC data is divided into multiple SOC intervals.

[0009] As one preferred approach, the SOC data is divided into multiple SOC intervals based on the SOC values ​​corresponding to the voltage plateau start point, polarization inflection point, and capacity regeneration inflection point, respectively, as boundary points. These intervals include: The SOC data within a preset time period is divided into equal-width sub-intervals to obtain multiple initial equal-width sub-intervals. Calculate the density difference between adjacent initial equal-width sub-intervals. If the density difference is less than a preset threshold, merge the two adjacent intervals. If the density difference is less than the preset threshold and there is an inflection point in any initial equal-width sub-interval, do not merge the intervals with the inflection point as the boundary point, so as to obtain multiple SOC intervals.

[0010] As one preferred approach, the set of operating parameters is input into an adaptive block extraction network model for block extraction, resulting in multiple operating parameters for block extraction, and thus block features, including: By using adaptive block extraction to process the set of running parameters in the fully connected layers of the network model, a learnable weight vector is obtained. The learnable weight vector is normalized to obtain the selection probability distribution of multiple candidate sizes; The set of running parameters is divided into sliding blocks according to each candidate size to obtain the block feature map of each candidate size; Using the selection probability distribution as weight, the feature maps of each block are weighted and summed to obtain the soft block mask; The soft block mask and the running parameters are multiplied element-wise before block mapping and feature embedding are performed to obtain the corresponding block features.

[0011] As one preferred approach, multi-scale encoding is performed on the block features to obtain a temporal feature vector, including: The block features are input into the encoder so that the encoder selects a preset number of expert branches according to the expert activation strategy; Within each expert branch, a local attention module is used to extract local features from a block of block features to obtain local features, and a sparse linear attention module is used to extract global dependencies between blocks of block features to obtain global features. Global and local features are fused to obtain a temporal feature vector.

[0012] As one preferred approach, each local dynamic feature is fused with the corresponding basic statistical features and higher-order statistical features to obtain the first feature, including: Each local dynamic feature is concatenated with the corresponding basic and higher-order statistical features to obtain the concatenation result. The first feature is obtained by nonlinearly transforming the splicing result using multiple fully connected layers.

[0013] As one preferred approach, the second feature is input into the prediction network for prediction to obtain the capacity prediction value, including: The second feature is input into the prediction network so that the prediction network performs global average pooling on the second feature to obtain a global representation value. The global representation value is transformed using a fully connected layer in the prediction network to obtain the transformation result; The transformation result is normalized by the Sigmoid activation function to generate feature weight coefficients. The feature weight coefficients are then multiplied element-wise with the second feature to obtain the feature recalibration result. The feature recalibration results are mapped to obtain the first output result. The second feature is input into the residual bypass and processed to obtain the second output result. The first output result is added element by element to the second output result, and then activated by the GELU activation function to obtain the third output result. The second feature is input into a parallel branch network for processing, resulting in the fourth and fifth output results. The third, fourth, and fifth output results are fused to obtain the capacity prediction value.

[0014] Another embodiment of the present invention provides a lithium battery remaining capacity prediction system, comprising: The acquisition module is used to acquire the set of operating parameters of the lithium battery within a preset time period under charging or discharging conditions. The set of operating parameters is preprocessed valid time-series data, including at least voltage data, capacity data and SOC data. The partitioning module is used to divide the SOC data into intervals, resulting in multiple SOC intervals. The block extraction module is used to input the set of running parameters into the adaptive block extraction network model for block extraction, obtain multiple running parameters for block extraction, obtain block features, and encode the block features at multiple scales to obtain a time-series feature vector. The statistical feature extraction module is used to calculate the basic and higher-order statistical features of the set of operating parameters within each SOC interval. The first fusion module is used to fuse each local dynamic feature with the corresponding basic statistical features and higher-order statistical features to obtain the first feature. The local dynamic features are obtained by inputting the set of running parameters into a gated recurrent neural network to extract local temporal dependencies. The second fusion module is used to construct the corresponding state vector based on the set of operating parameters corresponding to each SOC interval, process each state vector to obtain the first weight corresponding to the time series feature vector, and perform weighted fusion of the time series feature vector and the first feature based on the first weight to obtain the second feature. The prediction module is used to input the second feature into the prediction network for prediction, obtain the capacity prediction value, and perform inverse standardization on the capacity prediction value to obtain the remaining capacity of the lithium battery.

[0015] A third aspect of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the lithium battery remaining capacity prediction method of the first aspect.

[0016] This invention provides a method and system for predicting the remaining capacity of a lithium battery.

[0017] Compared with the prior art, the beneficial effects of the embodiments of the present invention are at least one of the following: A set of operating parameters for a lithium battery under charging and discharging conditions within a preset time period is obtained, including at least voltage data, capacity data, and State of Charge (SOC) data. The SOC data is divided into intervals to obtain multiple SOC intervals. The set of operating parameters is input into an adaptive block extraction network model for block extraction to obtain block features. The block features are then encoded at multiple scales to obtain a time-series feature vector. The basic and higher-order statistical features of the set of operating parameters within each SOC interval are calculated. Each local dynamic feature is fused with the corresponding basic and higher-order statistical features to obtain a first feature. The local dynamic features are obtained by extracting local time-series dependencies by inputting the set of operating parameters into a gated recurrent neural network. A state vector is constructed based on the set of operating parameters corresponding to each SOC interval. Each state vector is processed to obtain a first weight corresponding to the time-series feature vector. The time-series feature vector and the first feature are weighted and fused according to the first weight to obtain a second feature. The second feature is input into a prediction network for prediction to obtain a capacity prediction value. After de-standardization, the remaining capacity of the lithium battery is obtained, thereby effectively improving the accuracy and robustness of the remaining capacity prediction of the lithium battery. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the lithium battery remaining capacity prediction method provided by the present invention; Figure 2This is a schematic diagram of the capacity curve of the lithium battery remaining capacity prediction method provided by the present invention; Figure 3 This is a schematic diagram of the prediction results of lithium batteries in different SOC ranges using the lithium battery remaining capacity prediction method provided by this invention. Figure 4 This is a comparison of the prediction errors of the lithium battery remaining capacity prediction method provided by the present invention in different SOC ranges of lithium batteries. Figure 5 This is a schematic diagram of the lithium battery remaining capacity prediction system provided by the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0020] In the description of this application, the terms "first," "second," "third," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined with "first," "second," "third," etc., may explicitly or implicitly include one or more of that feature. In the description of this application, unless otherwise stated, "a plurality of" means two or more.

[0021] In the description of this application, it should be noted that, unless otherwise defined, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this specification is for the purpose of describing specific embodiments only and is not intended to limit the invention. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0022] See Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the lithium battery remaining capacity prediction method provided by the present invention, as shown below. Figure 1 As shown, this embodiment includes steps S101 to S107, and the specific steps are as follows: S101. Obtain the set of operating parameters of the lithium battery within a preset time period under charging or discharging conditions. The set of operating parameters is preprocessed valid time-series data, including at least voltage data, capacity data and SOC data. This embodiment does not limit the specific content of the set of operating parameters of the lithium battery under charging or discharging conditions. For example, the set of operating parameters includes at least current data, voltage data, SOC data, temperature data, internal resistance data, and capacity data. In addition to the real-time operating parameter set of the lithium battery, it also includes basic information parameters of the lithium battery, such as the lithium battery serial number, cycle count, aging stage label, and other metadata. This embodiment does not limit the specific duration of the preset time period. For example, the preset time period can be a complete charge-discharge cycle; or it can be multiple charge-discharge cycles; or it can be a specific continuous sampling period. This embodiment does not limit the data sampling frequency. For example, the sampling frequency can be 0.5Hz, that is, 1 data point is collected every 2 seconds; or it can be 1Hz, that is, 1 data point is collected every second; or it can be 2Hz, that is, 2 data points are collected every second.

[0023] It should be noted that the set of operating parameters within the preset time period constitutes multi-source sequence data, specifically including current sequence data, voltage sequence data, SOC sequence data, temperature sequence data, internal resistance sequence data, and capacity sequence data. Each type of sequence has the same length, and the six types of data points at the same sampling time are interconnected, together forming a multi-source synchronous time-series dataset with complete time correspondence.

[0024] In one embodiment, the set of operating parameters consists of preprocessed valid time-series data, wherein the preprocessing steps are as follows: Obtain the set of original operating parameters of the lithium battery within a preset time period under charging or discharging conditions; Remove the data belonging to the first and last loops from the original set of running parameters to obtain the set of valid running parameters; Using the Laida criterion, the data sequences in the effective operating parameter set are denoised to obtain the corresponding effective data sequences; Each valid data sequence is unified into a sequence of a preset length to obtain the corresponding valid time series data.

[0025] In this embodiment, the operating parameter set is valid time-series data that has been preprocessed from the original operating parameter set. For example, after obtaining the original operating parameter set of the battery, the data is preprocessed and cyclically filtered. Specifically, the original operating parameter set includes unprocessed current, voltage, SOC, temperature, internal resistance, and capacity sequences. The data in each sequence are arranged in order of sampling time, and each data point corresponds to a cycle number identifier. Since the battery has not yet reached a stable electrochemical operating state during the first cycle and is nearing the end of its lifespan during the last cycle, neither can represent the normal aging behavior of the battery. Therefore, all data belonging to the first and last cycles in each sequence are removed, and the remaining cycle number data is taken as valid operating data to obtain the valid operating parameter set.

[0026] Then, the Raida criterion is used to remove noise points from the current, voltage, SOC, temperature, internal resistance, and capacity series data. Specifically, the mean μ and standard deviation σ of each parameter series are calculated, and a reasonable data interval is set as [μ-3σ, μ+3σ]. Sampling points outside this interval are identified as abnormal noise points and removed, while data points within this interval are retained as valid data points, thus obtaining the corresponding valid data series. It should be noted that the mean and standard deviation of each parameter series are calculated independently, and the filtering is performed independently.

[0027] After the above denoising process, the length of each valid data sequence may vary due to differences in the total number of battery cycles or the number of noise points removed. Therefore, the valid data sequences corresponding to each operating parameter are uniformly standardized to a preset length. This preset length is determined based on the actual data volume and model requirements. It should be noted that the specific method for uniformly standardizing the valid data sequences is not limited; for example, linear interpolation can be used.

[0028] S102. Divide the SOC data into intervals to obtain multiple SOC intervals.

[0029] This embodiment employs a SOC adaptive segmentation method driven by inflection point and density to divide SOC sequence data into intervals. For example, under charging conditions, it is divided into 5 adaptive intervals: the first interval is the low SOC activation segment, the second interval is the voltage rise transition segment, the third interval is the main voltage plateau segment, the fourth interval is the polarization intensification segment, and the fifth interval is the high SOC termination segment. Under discharging conditions, the 5 adaptive intervals for discharging include: the first interval is the high SOC discharge start segment, the second interval is the main discharge plateau segment, the third interval is the discharge transition segment, the fourth interval is the low SOC polarization segment, and the fifth interval is the deep discharge termination segment.

[0030] It should be noted that the aging stage is used to characterize the health status level of a lithium battery during cycle use. This embodiment divides the aging stage into three levels: S1 initial healthy stage, S2 stable aging stage, and S3 severe aging stage. Specifically, S1 initial healthy stage refers to the stage where the number of cycles is less than or equal to the first threshold N1, and the current capacity is not less than 95% of the rated capacity. In this stage, the battery has sufficient internal electrochemical active materials, and the charge / discharge curve is regular. S2 stable aging stage refers to the stage where the number of cycles is greater than N1 and less than or equal to the second threshold N2, and the current capacity is between 80% and 95% of the rated capacity. In this stage, battery polarization gradually intensifies, and the curve inflection point begins to shift. S3 severe aging stage refers to the stage where the number of cycles is greater than N2, and the current capacity is less than 80% of the rated capacity. In this stage, battery capacity decay accelerates, the curve distortion is obvious, and capacity regeneration occurs frequently.

[0031] Different aging stages correspond to different SOC interval division results and feature weight allocation strategies, thereby achieving dynamic adaptation throughout the entire life cycle. For example, taking charging conditions as an example, the SOC interval division results of the same lithium battery will be significantly different at the 50th cycle (S1 early health stage) and the 450th cycle (S3 severe aging stage).

[0032] In some embodiments, the SOC data is divided into intervals to obtain multiple SOC intervals, including: The voltage curve is obtained by fitting and analyzing the voltage data and SOC data within a preset time period, and the capacity curve is obtained by fitting and analyzing the capacity data and SOC data within a preset time period. The voltage curve is subjected to first-order difference calculation to obtain the voltage difference sequence, and the capacity curve is subjected to first-order difference calculation to obtain the capacity difference sequence. Calculate the mean and standard deviation of the voltage differential sequence and the capacity differential sequence, respectively; Based on the mean and standard deviation of the voltage difference sequence, determine the starting point of the voltage plateau and the polarization inflection point corresponding to the voltage difference sequence; The inflection point of capacity regeneration is determined based on the mean and standard deviation of the capacity difference sequence; Based on the SOC values ​​corresponding to the voltage plateau start point, polarization inflection point, and capacity regeneration inflection point as boundary points, the SOC data is divided into multiple SOC intervals.

[0033] The specific implementation process of dividing SOC data into intervals in this embodiment is as follows: First, a fitting analysis is performed on the voltage data and SOC data within a preset time period to obtain a voltage curve. Then, a fitting analysis is performed on the volume data and SOC data within the preset time period to obtain a capacity curve. This embodiment does not limit the specific method of curve fitting. For example, the least squares method can be used for polynomial fitting; or spline interpolation can be used; or the original sampling points can be directly connected as the curve. It should be noted that when used to train the prediction network, historical operating data of multiple different sample batteries can be obtained, and each sample battery can be preprocessed according to the aforementioned steps to form a multi-sample time-series dataset and the capacity curve of each sample battery, such as... Figure 2 As shown, it is used for the training and validation of the subsequent prediction network.

[0034] Then, first-order difference calculations are performed on the voltage curve to obtain the voltage difference sequence, and first-order difference calculations are performed on the capacity curve to obtain the capacity difference sequence. Before calculating the differences, the voltage and capacity curves are smoothed and denoised using Savitzky-Golay filters to eliminate the interference of random noise in the original data on inflection point identification. The Savitzky-Golay filter is a smoothing method based on local polynomial least squares fitting, which can filter out high-frequency noise while preserving the signal shape characteristics. This embodiment does not limit the window size and polynomial order of the Savitzky-Golay filter. For example, the window size can be 5, 7, or 11; the polynomial order can be 2, 3, or 4.

[0035] For each smoothed curve, the first-order difference is calculated point by point to approximate the first derivative of the curve, resulting in voltage difference sequences and capacity difference sequences. The mean and standard deviation of the voltage difference sequences and capacity difference sequences are then statistically analyzed, and a sudden change threshold is set as the mean of the difference sequences ± 2 times the standard deviation. This embodiment does not limit the specific value of the standard deviation multiple. For example, the multiple can be 1.5; or 2; or 2.5. Next, the voltage difference sequences and capacity difference sequences are traversed, and three types of inflection points are distinguished based on the difference amplitude, trend, and local interval characteristics, according to the following specific rules: The voltage plateau inflection point is determined based on the voltage difference sequence. During charging, the voltage difference at multiple consecutive sampling points rapidly decreases from a large amplitude to a stable small amplitude, and the abrupt change in the difference value falls within the dynamic threshold range; this is identified as the voltage plateau inflection point. During discharging, the voltage difference rapidly increases from a small amplitude, and the corresponding abrupt change is identified as the voltage plateau inflection point. This inflection point characterizes the starting position of the main charging / discharging voltage plateau. The polarization inflection point is also determined based on the voltage difference sequence. Within the voltage plateau range, if the voltage difference continuously deviates from the stable average value, exceeds the dynamic threshold positively or negatively, and this abrupt change occurs in the middle or later part of the plateau without obvious plateau start-stop characteristics, it is identified as a polarization inflection point, corresponding to the location where battery polarization intensifies. The capacity regeneration inflection point is determined based on the capacity difference sequence. Alternating positive and negative capacity differences with amplitudes significantly exceeding the dynamic threshold, exhibiting localized sharp peak abrupt changes, are identified as capacity regeneration inflection points, specifically identifying the abrupt change location corresponding to battery capacity regeneration phenomena.

[0036] After identifying all inflection points, the SOC values ​​corresponding to the detected voltage plateau start point, polarization inflection point, and capacity regeneration inflection point are summarized, deduplicated, and sorted in ascending order of SOC value. All types of inflection points are treated as equivalent forced anchor point boundaries without priority distinction. Based on the sorted unified boundary set, a coarse interval is initially divided into SOC value ranges, generating only one set of coarse interval results, rather than dividing intervals separately for each type of inflection point.

[0037] In some embodiments, the SOC data is divided into multiple SOC intervals based on the SOC values ​​corresponding to the voltage plateau start point, polarization inflection point, and capacity regeneration inflection point as boundary points, including: The SOC data within a preset time period is divided into equal-width sub-intervals to obtain multiple initial equal-width sub-intervals. Calculate the density difference between adjacent initial equal-width sub-intervals. If the density difference is less than a preset threshold, merge the two adjacent sub-intervals. If the density difference is less than the preset threshold and there is an inflection point in any initial equal-width sub-interval, do not merge the sub-intervals with the inflection point as the boundary point, and obtain the second division result. The second division result includes multiple SOC intervals.

[0038] In this embodiment, the SOC value range is initially divided into coarse intervals based on the sorted unified boundary set, generating only one set of coarse interval results, namely the first division result. Subsequently, adjacent initial sub-intervals are merged according to their density differences. Specifically, the SOC value range is divided into multiple initial equal-width sub-intervals, and the distribution density of data points within each interval is calculated. Intervals with a density difference less than a preset threshold and no critical inflection points are merged; if a critical inflection point is present, the inflection point SOC is used as a mandatory boundary for non-merging. This embodiment is not limited to the calculation method of the density difference; for example, it can count the number of time-series data points within each initial equal-width sub-interval and use the ratio of the smaller to the larger of the ratios of the number of data points in two adjacent sub-intervals as the density difference. Regarding the preset threshold for merging density differences, this embodiment does not limit its specific value. For example, the density difference threshold can be 0.05; or it can be 0.08.

[0039] Then, interval length constraints are applied to filter out excessively short intervals with fewer than 5 data points and excessively long intervals with more than 200 data points, further splitting them according to local density minima to ensure smooth SOC changes and high feature homogeneity within the intervals. This embodiment does not limit the specific values ​​of the minimum and maximum number of data points. For example, the minimum number of data points can be 3, 5, or 8; the maximum number of data points can be 150, 200, or 250.

[0040] This embodiment does not limit the specific numerical values ​​used to divide the stages into early healthy phase, stable aging phase, and severe aging phase. For example, early healthy phase could correspond to a stage where the number of cycles is less than or equal to 50 and the capacity is greater than or equal to 95% of the rated capacity; or it could correspond to a stage where the number of cycles is less than or equal to 80 and the capacity is greater than or equal to 95% of the rated capacity; or it could correspond to a stage where the number of cycles is less than or equal to 100 and the capacity is greater than or equal to 95% of the rated capacity. Stable aging phase could correspond to a stage where the number of cycles is between 50 and 300 and the capacity is between 80% and 95% of the rated capacity; or it could correspond to a stage where the number of cycles is between 80 and 500 and the capacity is between 80% and 95% of the rated capacity; or it could correspond to a stage where the number of cycles is between 100 and 800 and the capacity is between 80% and 95% of the rated capacity. Severe aging phase could correspond to a stage where the number of cycles is greater than 300 and the capacity is less than 80% of the rated capacity; or it could correspond to a stage where the number of cycles is greater than 500 and the capacity is less than 80% of the rated capacity; or it could correspond to a stage where the number of cycles is greater than 800 and the capacity is less than 80% of the rated capacity. The adaptive segmentation method driven by inflection point and density can adapt to the dynamic changes of the charge and discharge curves throughout the battery's entire life cycle, solving the problem of sharp decline in accuracy of fixed SOC interval segmentation in the late aging stage and strong regeneration scenarios.

[0041] S103. Input the set of running parameters into the adaptive block extraction network model for block extraction, obtain multiple running parameters for block extraction, obtain block features, and encode the block features at multiple scales to obtain the temporal feature vector.

[0042] In this embodiment, various types of sequence data are uniformly truncated or padded to a fixed length. This embodiment does not limit the specific value of the fixed length. For example, the fixed length can be 1024, 2048, or 4096. The specific method of sequence length normalization is not limited in this embodiment. For example, for sequences shorter than the fixed length, zero-padding can be used; or repeated padding at the end of the sequence can be used; or linear interpolation can be used to padded to the fixed length. For sequences longer than the fixed length, truncation can be used, retaining the first fixed length of data points; or random sampling can be used, uniformly sampling a fixed length of data points from the sequence; or a sliding window method can be used, selecting a window segment containing the most valid information. Furthermore, a learnable dynamic patch mechanism is introduced into the adaptive block extraction network model. By combining dynamic patch embedding, learnable block mask and Gumbel Softmax soft selection, patch size adaptive learning is achieved, thereby extracting block features from the input set of running parameters. Then, the block features are encoded using adaptive Patch-SparseTransformer-MoE multi-scale encoding to obtain temporal feature vectors.

[0043] In some embodiments, the set of running parameters is input into an adaptive block extraction network model for block extraction to obtain block features, including: By using adaptive block extraction to process the set of running parameters in the fully connected layers of the network model, a learnable weight vector is obtained. The learnable weight vector is normalized to obtain the selection probability distribution of multiple candidate sizes; The set of running parameters is divided into sliding blocks according to each candidate size to obtain the block feature map of each candidate size; Using the selection probability distribution as weight, the feature maps of each block are weighted and summed to obtain the soft block mask; The soft block mask and the running parameters are multiplied element-wise before block mapping and feature embedding are performed to obtain the corresponding block features.

[0044] In this embodiment, after the length-normalized multi-source time series is input into the adaptive block extraction network model, it is first calculated through a fully connected layer with fixed parameters to output a weight vector with a dimension of 4. Then, the weight vector is normalized using the Gumbel Softmax function to output the selection probability distribution corresponding to the four candidate sizes (32, 64, 128, 256), and the sum of the four probability values ​​is 1.

[0045] For the normalized multi-source time series, sliding block partitioning is performed according to four candidate sizes: 32, 64, 128, and 256, resulting in four sets of block feature maps at different scales. Each partitioning method corresponds to a set of basic binary block masks, used to mark the block boundaries and effective regions in the temporal dimension. This embodiment does not limit the sliding step size. For example, the sliding step size can be the same as the block size, i.e., non-overlapping blocks; or the sliding step size can be half the block size, i.e., 50% overlapping blocks; or the sliding step size can be one-third of the block size. Then, using the four probability values ​​output by Gumbel Softmax as weights, the four sets of basic binary masks are summed element-wise to generate a unified soft block mask. The value range of each element in the mask is [0,1]. The higher the probability of a size, the greater the weight of the corresponding mask region, thus achieving a soft expression of size preference. Finally, the soft-block mask is multiplied element-wise with the multi-source temporal sequence to complete temporal dimension feature modulation. Then, a dynamic patch embedding layer performs block mapping and feature embedding on the modulated sequence to obtain block features incorporating multi-scale information, which serve as input to the subsequent multi-scale encoder. Throughout the inference process, all network parameters remain fixed, performing only forward computation without gradient calculation or parameter updates, thus ensuring the efficiency and stability of the prediction process. During the training phase, the learnable weight vector parameters are updated synchronously with model training.

[0046] In the dynamic patch embedding layer, block mapping and feature embedding are performed on the modulated sequence to obtain features that fuse multi-size information. Specifically, the modulated sequence is uniformly divided into blocks according to a preset baseline size, and each block is mapped to a fixed-dimensional embedding vector through a learnable linear projection matrix. This embodiment does not limit the specific value of the baseline size. For example, the baseline size can be 16; or 32; or 64.

[0047] In some embodiments, the block features are encoded using multi-scale methods to obtain a temporal feature vector, including: The block features are input into the encoder so that the encoder selects a preset number of expert branches according to the expert activation strategy; Within each expert branch, a local attention module is used to extract local features from a block of block features to obtain local features, and a sparse linear attention module is used to extract global dependencies between blocks of block features to obtain global features. Global and local features are fused to obtain a temporal feature vector.

[0048] In this embodiment, the block features are input into the encoder, which employs a heterogeneous multi-scale coding structure, including the aforementioned learnable dynamic patch module, local attention module, sparse linear attention module, and heterogeneous MoE expert module. The learnable dynamic patch module performs adaptive block segmentation and feature embedding of the multi-source temporal sequence before encoding. The encoder internally has four heterogeneous expert branches, each with the same network structure but independent parameters. Each branch integrates a local attention module and a sparse linear attention module. An adaptive gating network is set at the encoder front end. This gating network calculates the activation weights of each expert branch based on the input block features and executes a Top-2 expert activation strategy, retaining only the two expert branches with the highest weights for subsequent calculations, while masking the remaining branches, thereby reducing computational redundancy during inference. During inference, all network parameters remain fixed, and the weights output by the gating network are uniquely determined by the trained parameters, eliminating the need for gradient updates.

[0049] For each activated expert branch, the internal processing is as follows: First, the local attention module uses a single block as the basic processing unit, calculating attention weights only between time points within the block to extract local temporal dependencies, thereby accurately capturing short-term local fluctuation features such as capacity regeneration spikes and voltage drops. Simultaneously, the sparse linear attention module uses the Performer kernel function to perform a low-rank approximation of global temporal dependencies, establishing attention associations between blocks and extracting long-distance temporal dependencies across blocks, thus effectively modeling the long-term degradation trend of battery capacity. Each expert branch concatenates and fuses its extracted local features with global features along the feature dimension to obtain a representative feature output for that branch. Finally, the encoder uses the Top-2 weights output by the adaptive gating network to weighted aggregate the features output by the two activated expert branches, outputting a temporal feature vector with dimension 32.

[0050] It should be noted that the segmentation here refers to the segmentation along the time dimension within the same sequence.

[0051] S104. Calculate the basic and higher-order statistical characteristics of the set of operating parameters in each SOC interval.

[0052] In this embodiment, four basic statistical features—mean, maximum, minimum, and standard deviation—are extracted from each type of sequence data within each SOC interval to characterize the central tendency and dispersion of the data within that interval. Taking voltage sequences as an example, the mean reflects the average voltage level within the SOC interval, the maximum and minimum values ​​reflect the voltage fluctuation range, and the standard deviation reflects the dispersion of the voltage data. Similarly, the above four basic statistical features are extracted for current, SOC, temperature, internal resistance, and capacity sequences. Based on the above basic statistical features, seven higher-order statistical features are further extracted, specifically including coefficient of variation, kurtosis, skewness, autocorrelation coefficient, capacity regeneration amplitude, temperature fluctuation coefficient, and internal resistance change rate. The calculation methods for the basic statistical features can refer to the existing calculation methods for mean, maximum, minimum, and standard deviation, which will not be elaborated in this embodiment. The coefficient of variation characterizes the dispersion of the sequence and is obtained by dividing the standard deviation of the sequence data by the mean. The formula for calculating kurtosis is: In the formula, The mean of the voltage sequence. For the i-th data point in the voltage sequence, The mean of the sequence. The standard deviation of the sequence. The length of the sequence data.

[0053] Kurtosis is used to characterize the steepness of a data distribution. The larger the value, the steeper the distribution and the thicker the tail, which can effectively highlight the existence of extreme events such as capacity regeneration spikes.

[0054] The formula for calculating skewness is: In the formula, For skewness, For the first in the sequence data One data point, The mean of the sequence. The standard deviation of the sequence. The length of the sequence data.

[0055] The formula for calculating the autocorrelation coefficient is: In the formula, For the first in the sequence data One data point, For the first in the sequence data One data point, is the standard deviation of the sequence.

[0056] The calculation method for capacity regeneration amplitude is as follows: In the formula, Given the current interval capacity sequence, the regeneration intensity is quantified by the difference in extreme values ​​of the interval capacity.

[0057] The temperature fluctuation coefficient is calculated as follows: In the formula, T is a temperature sequence. This represents the average temperature.

[0058] The method for calculating the rate of change of internal resistance is as follows: In the formula, , These are the first and last values ​​of the interval internal resistance sequence, respectively, representing the aging rate of the internal resistance.

[0059] S105. The local dynamic features are fused with the basic statistical features and higher-order statistical features corresponding to the local dynamic features to obtain the first feature. The local dynamic features are obtained by inputting the set of running parameters into the gated recurrent neural network to extract local temporal dependencies.

[0060] In this embodiment, the preprocessed multi-source time series (including current series, voltage series, SOC series, temperature series, internal resistance series and capacity series) are first input into a single-layer gated recurrent unit network to perform local time-series dependency modeling and output local dynamic features.

[0061] In some embodiments, each local dynamic feature is fused with the corresponding basic statistical features and higher-order statistical features to obtain a first feature, including: Each local dynamic feature is concatenated with the corresponding basic and higher-order statistical features to obtain the concatenation result. The first feature is obtained by nonlinearly transforming the splicing result using multiple fully connected layers.

[0062] In this embodiment, while acquiring local dynamic features, the aforementioned basic statistical features (mean, maximum, minimum, standard deviation) and higher-order statistical features (coefficient of variation, kurtosis, skewness, autocorrelation coefficient, capacity regeneration amplitude, temperature fluctuation coefficient, internal resistance change rate) are extracted from each sequence within each SOC interval, resulting in a total of 11-dimensional statistical feature vectors. The local dynamic features output by the gated recurrent neural network are then concatenated with these 11-dimensional statistical feature vectors along the feature dimension to form a concatenated feature that integrates local temporal information and global statistical information.

[0063] Subsequently, the concatenated features are sequentially input into a four-layer fully connected network for nonlinear feature transformation. The dimensionality transformation of each layer is as follows: input dimension → 64 → 32 → 16 → output dimension. The first three fully connected layers are followed by ReLU activation functions and batch normalization layers to introduce nonlinearity, accelerate training convergence, and prevent gradient vanishing. The output dimension of the last fully connected layer is consistent with the dimension of the input concatenated features to facilitate subsequent residual connections or attention-weighted operations. After the four fully connected layers, an attention mechanism is introduced to dynamically weight each feature dimension: the intermediate features output by the fully connected network are mapped to attention score vectors through a fully connected layer, then normalized using a Softmax function to generate attention weights for each feature dimension. Finally, the attention weights are multiplied element-wise with the intermediate features to obtain the weighted feature representation. Ultimately, the attention-weighted features are used as the first feature output for subsequent fusion prediction with multi-scale temporal features. It should be noted that this embodiment does not limit the number of layers or hidden units in the gated recurrent neural network. For example, a gated recurrent neural network can employ a two-layer stacked structure to extract higher-level temporal abstract features; or a bidirectional GRU structure can be used to utilize both forward and backward temporal information. This embodiment does not limit the number of layers or the dimensions of each layer in the fully connected network; for example, a five-layer fully connected network can be used, with the dimensions transformed to match the input dimensions.

[0064] S106. Construct the corresponding state vector based on the set of operating parameters corresponding to each SOC interval, process each state vector to obtain the first weight corresponding to the time series feature vector, and perform weighted fusion of the time series feature vector and the first feature based on the first weight to obtain the second feature.

[0065] In this embodiment, a multi-source sensing dynamic fusion algorithm is used to adaptively weight and fuse the aforementioned multi-scale temporal feature vector and the first feature. The specific process is as follows: First, based on the current SOC range division, aging stage determination, and relevant data from the operating parameter set, a 5-dimensional state-aware feature vector is constructed. This vector includes: SOC range encoding (using one-hot encoding or sequence encoding to distinguish which of the five adaptive ranges the battery is currently in for charging or discharging), aging stage encoding (identifying whether it is currently in S1 (early healthy stage), S2 (stable aging stage), or S3 (severe aging stage), capacity regeneration intensity (i.e., the aforementioned capacity regeneration amplitude, quantifying the extreme difference in capacity within the current range), temperature fluctuation coefficient (i.e., the aforementioned, characterizing the relative temperature fluctuation within the current range), and internal resistance change rate (i.e., the aforementioned, characterizing the rate of change of internal resistance within the current range). This state vector comprehensively describes the battery's operating condition and aging state within the current SOC range. The 5-dimensional state vector is then input into a two-layer lightweight fully connected mapping network. This network consists of two fully connected layers, with a ReLU activation function in the middle layer to introduce nonlinear mapping capability. After performing feature transformation on the state vector, the mapping network outputs two scalar values: the original weights corresponding to the time-series features and the original weights corresponding to the statistical features. Subsequently, the two original weight values ​​were normalized to obtain the final fusion weight coefficients. and ,in The fusion weights for time-series features, These are the fusion weights for statistical features. To ensure the rationality and stability of the fusion weights, [the following is omitted as it is not relevant to the context]. The range of values ​​is constrained to be between 0.2 and 0.8.

[0066] In the early healthy stage (S1 stage), the battery charge-discharge curves are regular in shape and exhibit strong temporal regularity. Therefore, the reference value of temporal characteristics is relatively high at this stage, resulting in a relatively high α value (approximately 0.6 to 0.7) within the main SOC range. In the stable aging stage (S2 stage), battery polarization gradually intensifies, and the curves begin to distort. Both temporal regularity and statistical characteristics remain valuable, with α values ​​relatively balanced between the transition and polarization ranges (approximately 0.4 to 0.6). In the severe aging stage (S3 stage), the battery charge-discharge curves are severely distorted, and the temporal regularity is disordered. The reference value of statistical characteristics (such as kurtosis, skewness, and autocorrelation coefficient) increases relatively, leading to a relatively low α value (approximately 0.2 to 0.4) in the activation and termination stages. Furthermore, when the capacity regeneration intensity is high, temperature fluctuations are drastic, or the internal resistance change rate is high, it indicates strong disturbances under the current operating conditions, reducing the reliability of temporal characteristics. In this case, the fusion algorithm automatically reduces the temporal weight and correspondingly increases the statistical weight to enhance the model's adaptability to abnormal operating conditions. Based on the above adaptively generated fusion weights and The second feature is obtained by weighting and fusing the temporal feature vector with the first feature according to the following formula. This second feature integrates information from both temporal modeling and statistical modeling, and the fusion weights are dynamically adjusted according to the current operating conditions. This effectively adapts to the differences in feature contribution under different aging stages and different SOC intervals, providing a more comprehensive and accurate feature representation for subsequent capacity prediction.

[0067] S107. Input the second feature into the prediction network for prediction to obtain the capacity prediction value. Perform inverse standardization on the capacity prediction value to obtain the remaining capacity of the lithium battery.

[0068] In this embodiment, a prediction network is used to process the second feature to output a capacity prediction value. The overall dimensionality of the prediction network is transformed to 48→128→64→32→1, and it is composed of multiple layers of fully connected residual blocks connected in series. At the end of the network, a capacity regeneration sensing branch and a temperature-internal resistance correction branch are set in parallel. Finally, multi-branch feature fusion is performed through a hybrid expert output head.

[0069] In some embodiments, the second feature is input into the prediction network for prediction to obtain a capacity prediction value, including: The second feature is input into the prediction network so that the prediction network performs global average pooling on the second feature to obtain a global representation value. The global representation value is transformed using a fully connected layer in the prediction network to obtain the transformation result; The transformation result is normalized by the Sigmoid activation function to generate feature weight coefficients. The feature weight coefficients are then multiplied element-wise with the second feature to obtain the feature recalibration result. The feature recalibration results are mapped to obtain the first output result. The second feature is input into the residual bypass and processed to obtain the second output result. The first output result is added element by element to the second output result, and then activated by the GELU activation function to obtain the third output result. The second feature is input into a parallel branch network for processing, resulting in the fourth and fifth output results. The third, fourth, and fifth output results are fused to obtain the capacity prediction value.

[0070] In this embodiment, the main path sequentially includes two fully connected layers and a feature recalibration module. The hidden dimensions of the two fully connected layers are 128 and 64, respectively, and the activation function for both is GELU. The specific calculation process of the feature recalibration module is as follows: First, global average pooling is performed on the one-dimensional feature vector of the input main path, that is, the sum of all elements in the vector is taken as the average value to obtain a global representation value, which is used to represent the overall response level of the entire feature vector. Then, this global representation value is input into a fully connected layer for feature transformation, mapping the scalar value to a weight vector with the same dimension as the input feature vector. Then, it is normalized by the Sigmoid activation function to compress the values ​​of each element in the weight vector to the range of 0 to 1, generating feature weight coefficients. Finally, the feature weight coefficients are multiplied element-wise with the original input one-dimensional feature vector to complete the feature recalibration, thereby strengthening the effective feature dimensions and suppressing the invalid noise feature dimensions.

[0071] The residual bypass is configured based on the dimensional relationship between the input and output of the current residual block: when the input and output dimensions are the same, an identity mapping is used as the bypass; when the input and output dimensions are different, a fully connected layer without an activation function is used for dimension adaptation. The main path output and the residual bypass output are added element-wise, then activated by the GELU activation function, becoming the final output of the residual block. These three residual blocks are sequentially connected, with the output of each layer serving as the input to the next, progressively completing feature compression and nonlinear transformation from 48 dimensions to 1 dimension. At the end of the network, the features output from the last residual block are input into two parallel functional branches. The capacity regeneration sensing branch consists of two fully connected layers with a dimensional transformation of 48→32→16, used to extract patterns strongly correlated with capacity regeneration from the features; the temperature-internal resistance correction branch also consists of two fully connected layers with a dimensional transformation of 48→32→16, used to capture the impact of temperature fluctuations and internal resistance changes on capacity.

[0072] The output features of the third-layer residual block of the main path, the output features of the capacity regeneration sensing branch, and the output features of the temperature-internal resistance correction branch are all input into the hybrid expert output head. This output head dynamically calculates the contribution weight of each branch based on the input multi-source features and performs weighted fusion of the three features to obtain the final capacity prediction value, such as... Figure 3 As shown, Figure 3 These are the prediction results for different SOC ranges of lithium batteries. It should be noted that the third output result refers to the output characteristics of the third layer residual block of the main path, the fourth output result refers to the output characteristics of the capacity regeneration sensing branch, and the fifth output result refers to the output characteristics of the temperature-internal resistance correction branch.

[0073] To suppress overfitting, a dropout mechanism is introduced during network training, with a dropout probability set to 0.3, meaning 30% of neurons are randomly discarded during each forward propagation. Simultaneously, L2 regularization is introduced with a regularization coefficient set to 1e-4 to penalize the network weights and enhance the model's generalization ability. Finally, the prediction network outputs a standardized capacity prediction value, which requires subsequent destandardization to obtain the actual remaining capacity of the lithium battery.

[0074] The predicted values ​​are de-standardized using a normalizer to obtain the actual remaining capacity of the lithium battery. A prediction confidence level is generated by combining MoE expert weights, feature contribution, regeneration fluctuation intensity, and temperature / internal resistance influence factors. Meta-information such as fragment type, SOC range, aging stage, activated expert information, and adaptive patch size are output. This meta-information can be directly used for BMS health status display, capacity regeneration early warning, temperature / internal resistance anomaly alarm, and triggering of proactive maintenance strategies. Figure 4To assess the prediction error of remaining capacity under different SOC ranges and charge / discharge conditions of lithium batteries, the root mean square error (RMSE) was used as the evaluation metric. From the perspective of charging conditions, the RMSE was 0.0275 Ah in the low SOC activation stage (0-20%), decreasing to 0.0245 Ah in the voltage rise transition stage (20-40%), further decreasing to 0.0225 Ah in the main voltage plateau stage (40-60%), reaching its lowest value of 0.0205 Ah in the polarization intensification stage (60-80%), and rising back to 0.0300 Ah in the high SOC tail end stage (80-100%). This error distribution is closely related to the voltage change characteristics within each range: the voltage change is gradual and the data stability is high in the main voltage plateau stage, resulting in the best model prediction accuracy; while the voltage changes drastically or polarization intensifies in the beginning and end stages, increasing data volatility and consequently raising the prediction error. From the discharge conditions, the RMSE is 0.0245 Ah in the high SOC discharge initiation stage (100-80%), maintained at 0.0245 Ah in the main discharge plateau stage (80-40%), drops to 0.0195 Ah in the discharge transition stage (40-20%), reaches its lowest value of 0.0185 Ah in the low SOC polarization stage (20-40%, likely the capacity regeneration sensitive region), and is 0.0205 Ah in the deep discharge termination stage (40-60%, likely 60-80%). The overall error under discharge conditions is lower than under charging conditions, and also exhibits a distribution characteristic of low error in the middle range and relatively high error at both ends. The RMSE of the method in this invention is controlled within 0.0300 Ah in each SOC range, verifying that the method has good prediction accuracy and robustness across the entire SOC range and different aging stages. Furthermore, the table indirectly verifies the effectiveness of the SOC adaptive segmentation strategy, namely, dividing the SOC value range into 5 dynamic intervals by jointly driving the inflection point and density, so that the data in each interval have high feature homogeneity, thereby obtaining a low prediction error in each interval.

[0075] The lithium battery remaining capacity prediction system provided by the present invention is described below. The lithium battery remaining capacity prediction system described below can be referred to in correspondence with the lithium battery remaining capacity prediction method described above.

[0076] Please refer to Figure 5 The present invention also provides a lithium battery remaining capacity prediction system 200, comprising: The acquisition module 201 is used to acquire a set of operating parameters of the lithium battery within a preset time period under charging or discharging conditions. The set of operating parameters is preprocessed valid time-series data, including at least voltage data, capacity data and SOC data. The partitioning module 202 is used to partition the SOC data into intervals to obtain multiple SOC intervals; The block extraction module 203 is used to input the set of running parameters into the adaptive block extraction network model for block extraction, obtain multiple running parameters for block extraction, obtain block features, and encode the block features at multiple scales to obtain a time-series feature vector. The statistical feature extraction module 204 is used to calculate the basic and higher-order statistical features of the set of operating parameters in each SOC interval. The first fusion module 205 is used to fuse each local dynamic feature with the basic statistical features and higher-order statistical features corresponding to the local dynamic features to obtain the first feature. The local dynamic features are obtained by inputting the set of running parameters into a gated recurrent neural network to extract local temporal dependencies. The second fusion module 206 is used to construct a corresponding state vector based on the set of operating parameters corresponding to each SOC interval, process each state vector to obtain the first weight corresponding to the time series feature vector, and perform weighted fusion of the time series feature vector and the first feature based on the first weight to obtain the second feature. The prediction module 207 is used to input the second feature into the prediction network for prediction, obtain the capacity prediction value, and perform inverse standardization on the capacity prediction value to obtain the remaining capacity of the lithium battery.

[0077] The specific implementation of this lithium battery remaining capacity prediction system is basically the same as the specific implementation of the lithium battery remaining capacity prediction method described above, and will not be repeated here.

[0078] In one embodiment, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method.

[0079] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for predicting the remaining capacity of a lithium battery, characterized in that, include: Obtain a set of operating parameters of a lithium battery within a preset time period under charging or discharging conditions, wherein the set of operating parameters is preprocessed valid time-series data, including at least voltage data, capacity data, and SOC data; The SOC data is divided into intervals to obtain multiple SOC intervals; The set of operating parameters is input into the adaptive block extraction network model for block extraction to obtain block features. The block features are then encoded at multiple scales to obtain a temporal feature vector. Calculate the basic and higher-order statistical characteristics of the set of operating parameters within each SOC interval; Each local dynamic feature is fused with the corresponding basic statistical feature and higher-order statistical feature to obtain a first feature. Each local dynamic feature is obtained by inputting the set of running parameters into a gated recurrent neural network to extract local temporal dependencies. Based on the set of operating parameters corresponding to each SOC interval, a corresponding state vector is constructed. Each state vector is processed to obtain a first weight corresponding to the time series feature vector. Based on the first weight, the time series feature vector and the first feature are weighted and fused to obtain a second feature. The second feature is input into the prediction network for prediction to obtain the capacity prediction value. The capacity prediction value is then de-standardized to obtain the remaining capacity of the lithium battery.

2. The lithium battery remaining capacity prediction method as described in claim 1, characterized in that, The set of operating parameters consists of preprocessed valid time-series data, wherein the preprocessing steps are as follows: Obtain the set of original operating parameters of the lithium battery within a preset time period under charging or discharging conditions; Remove the data belonging to the first and last loops from the original set of operating parameters to obtain the set of valid operating parameters; Using the Laida criterion, each of the data sequences in the set of effective operating parameters is denoised to obtain the corresponding effective data sequence; The valid data sequences are unified into a sequence of a preset length to obtain the corresponding valid time series data.

3. The lithium battery remaining capacity prediction method as described in claim 1, characterized in that, The SOC data is divided into intervals to obtain multiple SOC intervals, including: A voltage curve is obtained by fitting and analyzing the voltage data and SOC data within a preset time period; a capacity curve is obtained by fitting and analyzing the capacity data and SOC data within the preset time period. The voltage curve is subjected to first-order difference calculation to obtain a voltage difference sequence, and the capacity curve is subjected to first-order difference calculation to obtain a capacity difference sequence. Calculate the mean and standard deviation of the voltage differential sequence and the capacity differential sequence, respectively; Based on the mean and standard deviation of the voltage difference sequence, determine the voltage plateau start point and polarization inflection point corresponding to the voltage difference sequence; The capacity regeneration inflection point is determined based on the mean and standard deviation of the capacity difference sequence. The SOC data is divided into multiple SOC intervals based on the SOC values ​​corresponding to the voltage platform start point, the polarization inflection point, and the capacity regeneration inflection point as boundary points.

4. The lithium battery remaining capacity prediction method as described in claim 3, characterized in that, The SOC data is divided into multiple SOC intervals based on the SOC values ​​corresponding to the voltage plateau start point, the polarization inflection point, and the capacity regeneration inflection point as boundary points, including: The SOC data within a preset time period is divided into equal-width sub-intervals to obtain multiple initial equal-width sub-intervals. Calculate the density difference between adjacent initial equal-width sub-intervals. If the density difference is less than a preset threshold, merge the two adjacent intervals. If the density difference is less than the preset threshold and there is an inflection point in any of the initial equal-width sub-intervals, do not merge them using the inflection point as the boundary point, so as to obtain multiple SOC intervals.

5. The lithium battery remaining capacity prediction method as described in claim 1, characterized in that, The step of inputting the set of operating parameters into the adaptive block extraction network model for block extraction to obtain block features includes: The set of running parameters is processed using the fully connected layer in the adaptive block extraction network model to obtain a learnable weight vector; The learnable weight vector is normalized to obtain the selection probability distribution of multiple candidate sizes; The set of operating parameters is divided into sliding blocks according to each of the candidate sizes to obtain a block feature map of each candidate size; Using the selection probability distribution as weight, the feature maps of each block are weighted and summed to obtain the soft block mask; The soft block mask and the running parameters are multiplied element-wise, and then block mapping and feature embedding are performed to obtain the corresponding block features.

6. The lithium battery remaining capacity prediction method as described in claim 1, characterized in that, The step of multi-scale encoding the segmented features to obtain a temporal feature vector includes: The block features are input into the encoder so that the encoder selects a preset number of expert branches according to the expert activation strategy; Within each of the expert branches, a local attention module is used to extract local features from a block of the block features to obtain local features, and a sparse linear attention module is used to extract the global dependencies between the blocks of the block features to obtain global features. The global features and the local features are fused to obtain a temporal feature vector.

7. The lithium battery remaining capacity prediction method as described in claim 1, characterized in that, The step of fusing each local dynamic feature with the corresponding basic statistical feature and higher-order statistical feature to obtain the first feature includes: Each local dynamic feature is concatenated with the corresponding basic statistical feature and higher-order statistical feature to obtain the concatenation result. The splicing result is nonlinearly transformed using multiple fully connected layers to obtain the first feature.

8. The lithium battery remaining capacity prediction method as described in claim 1, characterized in that, The step of inputting the second feature into the prediction network for prediction to obtain the capacity prediction value includes: The second feature is input into the prediction network, so that the prediction network performs global average pooling on the second feature to obtain a global representation value; The global representation value is transformed using a fully connected layer in the prediction network to obtain the transformation result. The transformation result is normalized by the Sigmoid activation function to generate feature weight coefficients. The feature weight coefficients are then multiplied element-wise by the second feature to obtain the feature recalibration result. The feature recalibration result is mapped to obtain a first output result, and the second feature input residual is bypassed and processed to obtain a second output result; The first output result is added element by element to the second output result, and then activated by the GELU activation function to obtain the third output result. The second feature is input into a parallel branch network for processing to obtain the fourth and fifth output results. The third, fourth, and fifth output results are fused to obtain the capacity prediction value.

9. A lithium battery remaining capacity prediction system, characterized in that, include: The acquisition module is used to acquire a set of operating parameters of the lithium battery within a preset time period under charging or discharging conditions. The set of operating parameters is preprocessed valid time-series data, including at least voltage data, capacity data, and SOC data. The partitioning module is used to divide the SOC data into intervals to obtain multiple SOC intervals; The block extraction module is used to input the set of running parameters into the adaptive block extraction network model for block extraction, obtain multiple running parameters for block extraction, obtain block features, and encode the block features at multiple scales to obtain a time-series feature vector. The statistical feature extraction module is used to calculate the basic and higher-order statistical features of the set of operating parameters within each SOC interval; The first fusion module is used to fuse each local dynamic feature with the basic statistical feature and the higher-order statistical feature corresponding to the local dynamic feature to obtain a first feature. Each local dynamic feature is obtained by inputting each set of running parameters into a gated recurrent neural network to extract local temporal dependencies. The second fusion module is used to construct a corresponding state vector based on the set of operating parameters corresponding to each SOC interval, process each state vector to obtain a first weight corresponding to the time series feature vector, and perform weighted fusion of the time series feature vector and the first feature based on the first weight to obtain a second feature. The prediction module is used to input the second feature into the prediction network for prediction to obtain a capacity prediction value, and to perform inverse normalization on the capacity prediction value to obtain the remaining capacity of the lithium battery.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the lithium battery remaining capacity prediction method as described in any one of claims 1 to 8.

Images