A method for predicting and controlling state of health of energy storage battery considering calendar aging

Abstract

The application discloses a kind of energy storage battery SOH prediction and control method considering calendar aging.The method comprises the following steps: collecting the continuous uncharged time length, SOC, terminal voltage, average voltage in cluster, internal resistance, cycle number and charge-discharge sequence of single battery;According to the continuous uncharged time length, SOC, under-voltage, consistency and cycle activity, the hit number is generated and the A-type battery and B-type battery are divided;Discharge lockout, preferential power supply, constant voltage saturation, active balancing and whole cluster unlocking are performed on the A-type battery, and pre-power supply is performed on the activated battery;Extract IC peak value, IC peak voltage and charge-discharge time length health characteristics, input the prediction model of super parameter determined by dream optimization algorithm combining calendar aging risk characteristics and working condition variables, and output SOH prediction value.The application links calendar aging risk control and SOH prediction, and is suitable for online health management of energy storage power station.

Description

Technical Field

[0001] This invention relates to the fields of energy storage battery state monitoring, battery energy storage system control, and battery health status estimation, and particularly to a method for predicting and managing the state of health (SOH) of energy storage batteries considering calendar aging. More specifically, this invention relates to a comprehensive management method for individual battery cells during the operation of an energy storage power station, including calendar aging risk identification, discharge lockout, charging priority ranking, active balancing, pre-charging, and SOH prediction. Background Technology

[0002] As a crucial component of the new power system, the energy storage industry undertakes tasks such as photovoltaic power consumption, peak shaving and valley filling, frequency regulation, and emergency backup. The core equipment in an energy storage power station is the battery energy storage system, which typically comprises multiple battery clusters, each containing multiple individual cells. The battery energy storage system needs to continuously collect operating parameters such as battery voltage, current, temperature, state of charge, internal resistance, and cycle count, and perform protection, balancing, and dispatch control based on these parameters.

[0003] The State of Health (SOH) of a battery characterizes the degree of performance degradation of the battery relative to its initial performance, and is typically expressed as the ratio of current usable capacity to initial capacity. Accurately estimating the SOH of energy storage batteries helps energy storage power stations identify abnormal batteries and reduce the risk of thermal runaway. Simultaneously, SOH is also a crucial basis for aging assessment, equalization control, single-cell replacement, cascade utilization, and decommissioning recycling. Failure to identify degraded cells in a timely manner can negatively impact the entire battery cluster, leading to slower power response, insufficient usable capacity, and increased operation and maintenance costs.

[0004] Existing SOH estimation methods can be broadly categorized into experiment-based prediction methods, model-based prediction methods, data-driven prediction methods, and fusion-based prediction methods. Experiment-based prediction methods typically require measuring parameters that directly reflect the battery's aging state, such as capacity and internal resistance, but they demand high sensor accuracy and stringent testing conditions, making them more suitable for offline laboratory testing. Model-based prediction methods can describe the battery degradation process using functions, circuit components, or electrochemical models, but model granularity, computational complexity, and the difficulty of parameter identification can affect their engineering deployment. Data-driven methods can extract health features and establish a mapping between health features and SOH, but they are susceptible to the coupling of health features, complex operating conditions, hyperparameter settings, and the quality of training data. Fusion-based prediction methods can combine the advantages of multiple methods, but they often suffer from high implementation costs, complex deployment, and reliance on training data.

[0005] The aging of energy storage batteries includes not only cycle aging during use but also calendar aging during storage. Calendar aging refers to the natural degradation of a battery over time, even when used infrequently or left unused. Calendar aging is related to irreversible, weak side reactions that occur continuously within the battery and are difficult to completely avoid even under optimal storage conditions. For energy storage power stations, due to the varying frequencies of operation of different battery clusters and individual cells, some batteries may be used frequently, while others may be used infrequently or left unused for extended periods. Batteries left idle for long periods at low energy levels are prone to capacity decay, increased internal resistance, and deterioration in voltage consistency.

[0006] In actual energy storage power station operation, prolonged periods without charging, low State of Charge (SOC), undervoltage terminal voltage, excessive voltage difference from the average voltage of the same cluster, increased internal resistance compared to the initial value, and a recent low cycle count all indicate that a single battery cell is in a low-energy, long-term stagnant state at risk. If such batteries continue to participate in discharge or grid-connected scheduling, it may lead to over-discharge, accelerated calendar aging, and irreversible consequences such as active lithium loss and abnormal changes in the SEI film. On the other hand, if only SOH estimation is performed without locking, charging, and balancing the risky batteries, the prediction results are difficult to directly translate into management actions for the energy storage power station.

[0007] Therefore, there is an urgent need for a technical solution suitable for the operation of BESS (Battery Emergency Storage System) in energy storage power stations. This solution can automatically identify low-energy, long-term stagnant batteries in the power station's operation data, classify and manage risky batteries, lock out discharge, prioritize recharging, maintain constant voltage saturation, actively balance, and prevent recurrence through pre-recharging. Based on this, it can extract health characteristics and calendar aging risk characteristics to complete online SOH (State of Health) prediction, thereby forming an integrated process of aging identification, active repair, health prediction, and scheduling linkage. Summary of the Invention

[0008] The purpose of this invention is to provide a method for predicting and managing the state of energy (SOH) of energy storage batteries that takes into account calendar aging, in order to solve the problems of difficulty in timely identification of low-energy batteries that have been sitting for a long time in energy storage power stations, lack of closed-loop protection and recharge management for risky batteries, and failure to fully utilize calendar aging risk information and exogenous variables of operating conditions in SOH prediction.

[0009] To achieve the above objectives, a method for predicting and managing State of Harm (SOH) of energy storage batteries considering calendar aging is proposed. This method is executed by a battery energy storage system on battery clusters and individual cells in an energy storage power station. The battery energy storage system can be a Battery Energy Storage System (BESS), a Battery Management System (BMS), or a data processing device communicating with a BESS. The method includes data acquisition, risk assessment, classification and labeling, lockout protection, recharging sequencing, constant current charging, constant voltage saturation, active balancing, cluster unlocking, pre-recharging, health feature extraction, construction of endogenous and exogenous variables, predictive model processing, and DOA hyperparameter optimization.

[0010] In terms of data acquisition, the battery energy storage system collects the continuous uncharged time t, state of charge (SOC), terminal voltage U, and average voltage U of each individual battery cell. avg The data includes: current internal resistance R, initial internal resistance R0, number of effective charge / discharge cycles N in the past 30 days, charge / discharge voltage and current time series, temperature, and charge / discharge rate. The continuous uncharged time t characterizes the battery's resting time, SOC characterizes the remaining capacity, and terminal voltage U and cluster average voltage Uc are also included. avg The internal resistance is used to characterize the consistency of a single cell relative to cells in the same cluster. The current internal resistance R and the initial internal resistance R0 are used to characterize the growth of internal resistance. The number of effective charge-discharge cycles N in the past 30 days is used to characterize the cycle activity. The charge-discharge voltage and current time series are used for subsequent health feature extraction.

[0011] In terms of risk assessment, the battery energy storage system performs five assessments on each individual cell. The first assessment is whether the continuous uncharged time t reaches 7 days; the second assessment is whether the SOC is less than 20%; the third assessment is whether the terminal voltage U is lower than the undervoltage threshold, where 3.0V can be used as the undervoltage threshold for lithium iron phosphate batteries, and the undervoltage threshold of the corresponding chemical system is used for other battery systems; the fourth assessment is whether the voltage difference between the individual cell and the average voltage of the same cluster is greater than 50mV, or whether the growth rate of the current internal resistance R relative to the initial internal resistance R0 reaches 15%; and the fifth assessment is whether the number of effective charge-discharge cycles N in the past 30 days is less than 2.

[0012] When a single battery cell meets two or more of the above five criteria, the battery storage system classifies it as a Class A battery; if a single battery cell meets fewer than two criteria, it is classified as a Class B battery. Class A batteries are low-energy batteries at risk of long-term stagnation, while Class B batteries are normal batteries. This classification result can be stored in the BESS backend database, BMS status register, or battery cluster scheduling table, serving as the basis for subsequent lockout, recharging, and balancing operations.

[0013] For Class A batteries, a discharge lockout state is programmed into the battery energy storage system, preventing them from participating in discharge and grid dispatch until recharging and activation are complete; they are only permitted to charge. This discharge lockout state can be implemented through discharge circuit control commands, discharge permission flags, grid-connected dispatch interface lockout flags, or BESS software permission bits. This lockout prevents low-energy batteries from continuing to discharge after long periods of idling, thus avoiding over-discharge and accelerated calendar aging.

[0014] For multiple Class A batteries, the battery energy storage system prioritizes them for replenishment based on a replenishment priority score P. Preferably, the replenishment priority score P is calculated according to the following formula:

[0015] (1)

[0016] In formula (1), w1, w2, w3, and w4 are non-negative weights and w 1+ w 2+ w 3+ w4=1; when the current resistance R is not collected, set w4=0 and renormalize w1, w2, and w3. The higher the charging priority score P, the higher the priority of the corresponding Class A battery in the charging queue. The above scoring comprehensively considers the resting time, charge level, cluster consistency, and internal resistance growth, giving priority to individual cells that have been resting for a long time, have low charge, large voltage difference, or significant internal resistance growth.

[0017] In terms of charging and balancing, the battery energy storage system performs constant current charging on Class A batteries according to the descending order of charging priority score P. When the power station dispatch pressure is lower than the preset pressure threshold, a charging rate of no more than 0.2C is used; when the power station dispatch pressure is not lower than the preset pressure threshold, a charging rate of 0.3C to 0.5C is used. After charging to the full charge cutoff condition, a short-term constant voltage saturation stage is entered. During the constant voltage saturation stage, active balancing is simultaneously activated to ensure that the difference between the maximum and minimum terminal voltages of individual cells in the same cluster is no greater than the balancing cutoff voltage difference.

[0018] The preset pressure threshold can be calculated based on at least one of the current grid dispatch command power, the available power of the energy storage station, the number of Class A batteries to be recharged, and the expected recharge duration, or determined by the BESS preset strategy.

[0019] Regarding activation and deactivation, after a single Class A battery completes recharging, constant voltage saturation, and active balancing, the battery storage system marks it as an activated battery. Only after all Class A batteries within the same battery cluster are marked as activated are the discharge lockout state of that battery cluster released, allowing the cluster to participate in discharge, grid peak shaving, frequency regulation, or grid-connected scheduling. This cluster-wide deactivation logic prevents the entire cluster from prematurely participating in scheduling before individual cells have recovered.

[0020] In terms of preventing recurrence and pre-charging, the battery storage system continues to monitor the quiescent state of activated batteries. When an activated battery experiences a continuous period of non-charging exceeding 5 days or a State of Charge (SOC) not exceeding 30%, a small-amount pre-charging is automatically triggered to restore the individual battery to an SOC of 50%. Through this pre-charging, the battery is prevented from re-entering a long-term low-energy quiescent state.

[0021] In terms of health feature extraction, the battery energy storage system or its communication-connected data processing equipment extracts the IC peak value, IC peak voltage, constant current charging duration, constant voltage charging duration, constant voltage rise charging duration, and constant voltage drop discharging duration from the charging and discharging voltage-current time series. The IC peak value and IC peak voltage are obtained by smoothing the capacity Q-voltage V curve of the constant current charging segment and then calculating the dQ / dV curve. The constant current charging duration is the time difference from the start of the rated charging current stabilization to the end of the transition to constant voltage mode. The constant voltage charging duration is the time difference from reaching the full charge cutoff voltage to the moment the charging current drops to the minimum cutoff current. The constant voltage rise charging duration is the time difference between the start and end of the stable voltage rise interval during charging. The constant voltage drop discharging duration is the time difference between the start and end of the voltage plateau interval during discharging.

[0022] In constructing endogenous and exogenous variables, the aforementioned health characteristics are arranged according to sampling period, cycle period, or time window to obtain an endogenous variable sequence; calendar aging risk characteristics, temperature, charge / discharge rate, and cycle count are arranged in the same time order to obtain an exogenous variable sequence. The calendar aging risk characteristics may include at least one of the following: continuous uncharged duration t, SOC, voltage difference within the same cluster, internal resistance growth rate, number of hits, or charging priority score P.

[0023] Regarding the prediction model, the endogenous and exogenous variable sequences are input into the prediction model whose hyperparameters are determined by the dream optimization algorithm. This prediction model sequentially performs adaptive spectral processing, interactive convolution processing, endogenous and exogenous variable cross-embedding, and linear projection, outputting the SOH prediction value at the current time or in the next S steps.

[0024] Furthermore, adaptive spectral processing includes performing a discrete Fourier transform on the input time series sequence x[n] to obtain the expression for the frequency domain sequence X[k], as follows:

[0025] (2)

[0026] In formula (2), x[n] is a time sequence and X[k] is a frequency domain sequence.

[0027] The power sequence P[k] = |X[k]|² is calculated based on the frequency domain sequence X[k], and a frequency domain mask is generated based on the learnable threshold θ. The frequency domain features F are extracted from the frequency domain sequence X[k]. The frequency domain features F are multiplied element-wise with the frequency domain mask to obtain the denoised frequency domain sequence F. a ;Denoising frequency domain sequence F a Respectively with global filter W G and local filter W L Multiply and sum to obtain the fused frequency domain sequence G;

[0028] Specifically:

[0029] (6)

[0030] (7)

[0031] In formula (6), P is the power spectrum. The learnable threshold is represented by ⊙, which indicates element-wise multiplication.

[0032] Formula (7) is the fused frequency domain sequence G, where, To capture the frequency domain features after long-range dependence, To capture the frequency domain characteristics after short-range fluctuations, W G For a global filter, W L For local filters;

[0033] The fused frequency domain features are then transformed back into the time domain to complete adaptive spectral enhancement, as follows:

[0034] (8)

[0035] In formula (8), Let S represent IFFT, where S is a spectral enhancement time-domain sequence.

[0036] Furthermore, the interactive convolutional processing includes: convolving the spectral enhancement temporal sequence S with a first one-dimensional convolutional kernel to obtain a first convolutional feature; convolving the spectral enhancement temporal sequence S with a second one-dimensional convolutional kernel to obtain a second convolutional feature; multiplying the first convolutional feature element-wise with the second convolutional feature processed by the activation function to obtain a first interactive feature; multiplying the second convolutional feature element-wise with the first convolutional feature processed by the activation function to obtain a second interactive feature; concatenating or adding the first and second interactive features and inputting them into a fusion convolutional kernel to obtain the enhanced temporal representation O. ICB ;

[0037] Specifically:

[0038] (9)

[0039] (10)

[0040] In formula (9), The first one-dimensional convolution kernel, For the second one-dimensional convolution kernel, Activate the function;

[0041] In formula (10), 3 represents the fusion convolution kernel, O ICB To enhance the temporal representation.

[0042] Furthermore, the cross-embedding and linear projection of endogenous and exogenous variables include: performing instance normalization on the endogenous variable sequence to obtain... Parametric instance normalization of the exogenous variable sequence yields Stacked along the time dimension and Cross-correlation features are obtained through one-dimensional convolution. ;according to Generate embedded sequences ,in For learnable parameters; embedding sequences The sequence is divided into a patch sequence of length p. The patch projection features and the position embedding PE are fused together by weighting with β, where β is a learnable parameter. The fused patch features are then concatenated and passed through a linear projection layer to generate SOH prediction values.

[0043] Specifically, the enhanced time-series representation is input into the CrossLinear algorithm for SOH prediction, and the endogenous variables are instance-normalized and the normalized results are output:

[0044] (3)

[0045] In formula (3), To output the normalized result, μ is the mean and σ is the variance. Input for endogenous variables;

[0046] Parametric instance normalization of the exogenous variable sequence yields:

[0047] (4)

[0048] In formula (4), For the output results, Input for exogenous variables;

[0049] The normalized data is input into the backbone network, which outputs the predicted values ​​in the normalized space. The predicted values ​​are then restored using the original mean μ and variance σ to obtain the final results in the real space, as follows:

[0050] (11)

[0051] In formula (11), This represents the computation of the algorithm's backbone network. This refers to the inverse normalization operation. To predict intermediate values, These are the predicted values ​​of the restored endogenous variables.

[0052] Endogenous and exogenous variables are stacked along the time dimension, and the cross-correlation features between the variables are extracted through one-dimensional convolution, as shown below:

[0053] (12)

[0054] In formula (12), As a cross-correlation feature, Represents one-dimensional convolution. Indicates folding along the time dimension;

[0055] The learnable parameter α is weighted and fused with the original endogenous variables and cross-correlation features to obtain the final embedding, as shown below:

[0056] (5)

[0057] Obtain the embedded sequence , in formula (5), These are learnable parameters;

[0058] The embedded sequence is divided into patches of length p, resulting in a total of k = ⌈T / p⌉ patches, as shown below:

[0059] (13)

[0060] Where T is the embedding sequence length, p represents the patch length, and k is the number of patches; in formula (13), Patchify represents the equal division operation;

[0061] The learnable parameter β is weighted and fused with the patch projection features and the location embedding, as shown below:

[0062] (14)

[0063] In formula (14), PE represents the position embedding operation, and Projection1 represents the weighted fusion;

[0064] All patch features are concatenated and then passed through a linear projection layer to generate the future S-step prediction result, as shown below:

[0065] (15)

[0066] In formula (15), Projection2 refers to the linear projection layer, and Concat represents the stitching patch. This is the final output prediction result.

[0067] Furthermore, the dream optimization algorithm uses the SOH prediction error of the prediction model on the validation set as the fitness value, and uses the learning rate, kernel size, patch length p, batch size, spectral threshold θ, and learnable parameters as the fitness values. At least two of the learnable parameters β are selected as hyperparameters to be optimized. The dream optimization algorithm includes: randomly initializing the population within the hyperparameter search boundary; dividing the population into five subpopulations in the first 90% of iterations, retaining the best individual in each subpopulation and updating the global best individual; performing random forgetting and boundary-based supplementation on some dimensions of the individuals, and sharing position information among the individuals; canceling the population division in the last 10% of iterations and performing a local search around the global best individual; when the iteration termination condition is met or the validation set error no longer decreases, outputting the hyperparameter combination corresponding to the global best individual.

[0068] Specifically:

[0069] Initialize the sample space and divide the population as follows:

[0070] (16)

[0071] In formula (16), rand represents the initialization space. Represents an individual vector. Let be a scalar, representing the decision variable (the value of a single dimension) for the i-th individual.

[0072] Each population is divided into five groups for global exploration, retaining the globally optimal individual position, as shown below:

[0073] (17)

[0074] In formula (17), This indicates the initial position after grouping. This indicates the optimal individual position obtained;

[0075] Instead of grouping the population, we explore local optima, as shown below:

[0076] (18)

[0077] In formula (18), This represents the value of the i-th individual and the j-th dimension of the decision variable after this round of iterations. Similarly, This represents the value of the j-th dimension after the current update of a random individual m. This indicates that the random individual m has not been updated in this round and still retains the original value of the j-th dimension from the previous iteration;

[0078] For iterations that do not meet the conditions, the process terminates at the boundary; once the conditions are met, the output is terminated immediately to obtain the optimal hyperparameters, as shown below:

[0079] (19)

[0080] In formula (19), These represent decision variables, i.e., the hyperparameters of the model to be optimized. This indicates the core objective, namely, the prediction error; This refers to inequality constraints and their corresponding penalty weights. Then it refers to the equality constraint and the corresponding penalty weight; That is the ultimate goal of optimization.

[0081] Furthermore, after obtaining the SOH prediction value, the calendar aging risk status of the corresponding individual battery cell is updated based on the SOH prediction value, and the charging priority of the individual battery cell is adjusted according to the updated calendar aging risk status. Thus, the SOH prediction value is not used as isolated health assessment data, but rather as a feedback parameter for calendar aging risk identification and charging scheduling, participating in battery health management and creating a closed loop between risk identification, charging repair, and SOH prediction.

[0082] Furthermore, when the predicted SOH value is lower than a preset health threshold, the corresponding individual battery cell is maintained as a Class A battery, or its charging priority is increased; when the predicted SOH value is not lower than the preset health threshold, and the individual battery cell no longer meets the criteria for a Class A battery, the individual battery cell is maintained as a Class B battery or upgraded from a Class A battery to a Class B battery. The preset health threshold can be determined based on the operation and maintenance rules of the energy storage power station, battery rated life criteria, or BESS preset strategy.

[0083] Compared with the prior art, the present invention has at least the following beneficial effects.

[0084] First, this invention combines factors such as long-term non-charging, low SOC, undervoltage, intra-cluster voltage difference, internal resistance growth, and insufficient cycle count into a calendar aging risk assessment process, which is conducive to forming an executable classification of Class A and Class B batteries at the operation end of energy storage power stations.

[0085] Second, this invention sets up discharge lockout, priority charging, constant voltage saturation, active balancing, cluster unlocking and pre-charging for Class A batteries, forming a closed-loop process from risk identification to control execution.

[0086] Third, this invention uses both health characteristics and calendar aging risk characteristics for SOH prediction, enabling the prediction process to simultaneously consider charge and discharge behavior, resting state, and external operating conditions.

[0087] Fourth, this invention establishes an SOH prediction process through adaptive spectral processing, interactive convolution processing, cross-embedding of endogenous and exogenous variables, and dream optimization algorithm, which facilitates offline training and online deployment at the BESS terminal of the energy storage power station.

[0088] Fifth, this invention feeds back the SOH prediction value to the calendar aging risk status update and recharge priority adjustment process, so that the SOH prediction value is no longer just a health assessment result, but participates in the dynamic update of Class A and Class B batteries, as well as the dynamic adjustment of the recharge repair order, thereby forming a closed-loop management of risk identification, recharge repair, SOH prediction and health control. Attached Figure Description

[0089] Figure 1 This is a flowchart illustrating the overall process of predicting and controlling the state of hazardous waste (SOH) of energy storage batteries considering calendar aging, as described in this invention.

[0090] Figure 2 This is a flowchart of the calendar aging risk assessment and Class A / B battery classification for the present invention;

[0091] Figure 3 This is a flowchart of the Class A battery locking, priority charging, active balancing, and cluster unlocking process of the present invention;

[0092] Figure 4 This is a flowchart of the health feature extraction process of the present invention;

[0093] Figure 5 This is a flowchart of the overall processing of the SOH prediction model of the present invention;

[0094] Figure 6 This is a flowchart illustrating the joint processing of the ASB adaptive spectral processing module and the ICB interactive convolution processing module of the present invention.

[0095] Figure 7 This is a diagram illustrating the CrossLinear prediction process and system application framework of the present invention. Detailed Implementation

[0096] These embodiments are provided to make the invention thorough and complete, and to fully express the scope of the invention to those skilled in the art. It should be noted that, unless otherwise specifically stated, the relative arrangement of components and steps, material composition, numerical expressions, and values ​​set forth in these embodiments should be interpreted as merely exemplary and not as limiting.

[0097] The embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the following embodiments are used to further explain the technical solution of the present invention and are not intended to limit the scope of protection of the present invention. Without departing from the technical concept of the present invention, those skilled in the art can make adaptive adjustments to the threshold, weight, charging rate and model parameters according to the scale of the energy storage power station, the battery chemistry system, the BESS sampling period, the contactor configuration and the active balancing method.

[0098] Example 1 like Figures 1 to 7As shown, this embodiment uses lithium iron phosphate energy storage batteries in an energy storage power station as an example for illustration. The energy storage power station includes multiple battery clusters, and each battery cluster includes multiple individual cells connected in series or in series-parallel combination. BESS automatically starts a full-station battery scan according to a fixed scan cycle, preferably once per hour. During each scan, BESS traverses all battery clusters and the individual cells within each battery cluster, and reads the continuous uncharged time t, SOC, terminal voltage U, and average voltage U of the same cluster for each individual cell. avg Current internal resistance R, initial internal resistance R0, effective charge / discharge cycle count N in the past 30 days, charge / discharge voltage and current time series, temperature, and charge / discharge rate.

[0099] The continuous uncharged duration t can be calculated by BESS based on the end time of the most recent charge of the individual cell and the current scan time. SOC can be read by BESS based on current integration, open-circuit voltage correction, or existing BMS estimation results. Terminal voltage U is the voltage across the individual cell at the current scan time. Cluster average voltage Uc avg This is the average terminal voltage of all individual cells within the same battery cluster. The current internal resistance R can be obtained by the BESS through pulse current response, online impedance estimation, or BMS internal resistance estimation. The initial internal resistance R0 can be the factory-calibrated internal resistance, the initial commissioning internal resistance, or the maintenance calibration internal resistance.

[0100] The number of effective charge-discharge cycles N in the past 30 days can be obtained by reviewing the battery operation records. Preferably, a complete charge-discharge process with a SOC fluctuation of not less than 20% is counted as one effective charge-discharge cycle. If the energy storage power station adopts other cycle counting rules, the number of cycles obtained under that rule can also be used as the number of effective charge-discharge cycles N in the past 30 days.

[0101] This invention standardizes the definition of a low-energy, long-term static risk state using five criteria. First, a single battery cell has not been charged for 7 days or more. Second, the current state of charge (SOC) of the single battery cell is less than 20%. Third, the terminal voltage of the single battery cell is below the undervoltage threshold; for lithium iron phosphate batteries, the undervoltage threshold can be 3.0V; for other chemical systems, the undervoltage threshold is the corresponding undervoltage range threshold for that system. Fourth, the voltage difference between the single battery cell and the average voltage of the cluster is greater than 50mV, or the current internal resistance R increases by 15% or more relative to the initial internal resistance R0. Fifth, the single battery cell has had less than 2 effective charge-discharge cycles in the past 30 days.

[0102] BESS assesses each individual cell against the five conditions mentioned above and counts the number of conditions met. If two or more conditions are met, the cell is classified as a Class A cell; if fewer than two conditions are met, it is classified as a Class B cell. Class A cells represent cells with low energy density and high risk of long-term static storage, while Class B cells maintain normal operating conditions. This method of determining whether two or more conditions are met avoids misjudgments caused by short-term fluctuations in a single parameter and also identifies calendar aging risks caused by multiple factors.

[0103] Example 2 like Figure 3 As shown, after classifying Class A and Class B batteries, BESS immediately implements discharge circuit lockout protection for all Class A batteries. Discharge lockout protection can manifest in one or more ways, such as disconnecting the discharge contactor, disabling discharge control authority, writing a disabling status word, or prohibiting the individual cell or its battery cluster from participating in grid-connected scheduling. For Class A batteries, before completing recharging, constant voltage saturation, and activation, only the charging circuit is allowed to conduct; further discharge, peak shaving, frequency regulation, or grid-connected output is not permitted.

[0104] When multiple Class A batteries exist within the same energy storage power station, BESS prioritizes them based on a charging priority score P. The charging priority score P considers factors such as continuous uncharged time, charge level, voltage difference within the same cluster, and internal resistance growth rate. The longer the resting time, the lower the SOC, the greater the voltage difference within the same cluster, and the higher the internal resistance growth rate, the higher the charging priority score P, and the higher the corresponding Class A battery will enter the charging queue first.

[0105] The priority score P for power replenishment is calculated according to the following formula:

[0106] (1)

[0107] In formula (1), w1, w2, w3, and w4 are non-negative weights and w 1+ w 2+ w 3+ w4=1; when the current resistance R is not sampled, set w4=0 and renormalize w1, w2 and w3.

[0108] BESS generates a charging queue based on the charging priority score P of all Class A batteries in the station, and charges Class A batteries according to the charging queue. Charging can be performed using low-rate gentle charging or high-efficiency constant-current charging. Preferably, when the station's dispatch pressure is lower than a preset pressure threshold, a constant-current charging rate of no more than 0.2C is used to reduce the impact of the charging process on the internal electrochemical system of the battery; when the station's dispatch pressure is not lower than the preset pressure threshold and a faster restoration of the battery's dispatchable state is required, a constant-current charging rate of 0.3C to 0.5C is used to balance repair efficiency and station operation requirements.

[0109] Once a single Class A battery cell is charged to 100% SOC or reaches the full charge cutoff voltage, the BESS stops constant current charging and switches to a short-term constant voltage saturation phase. During the constant voltage saturation phase, the BESS simultaneously initiates active balancing, gradually reducing the voltage difference between individual cells within the same cluster. Active balancing can be achieved using energy transfer balancing, bypass balancing, or the BESS's existing active balancing circuitry. Active balancing continues until the difference between the maximum and minimum terminal voltages of the individual cells within the same cluster is no greater than the balancing cutoff voltage difference, or until the balancing duration set by the BESS ends.

[0110] Once a single Class A battery completes full charging, constant voltage saturation, and active balancing, BESS marks it as an activated battery. For the same battery cluster, if there are still Class A batteries that have not been activated, the cluster remains in a discharge-locked state. Only after all Class A batteries in the cluster have been marked as activated will BESS release the cluster's discharge and grid dispatch permissions. This prevents the entire cluster from prematurely participating in discharge before a single high-risk cell has recovered.

[0111] To prevent activated batteries from re-entering a prolonged low-energy quiescent state, BESS continues to monitor the continuous uncharged time and SOC of activated batteries. When the continuous uncharged time of an activated battery exceeds 5 days, or the SOC drops to 30% or below, BESS automatically triggers a small pre-charge, charging the individual battery to SOC=50% and then stopping. This pre-charge process maintains the battery within a suitable SOC range for quiescent storage, reducing the likelihood of prolonged low-energy quiescent scenarios.

[0112] Example 3 like Figure 4 As shown, the BESS or a data processing device connected to the BESS extracts health features based on the collected charging and discharging voltage and current time series. Health feature extraction can be performed according to a cycle period, daily cycle, charging process, or a preset time window. For each data window, if there are effective constant current and constant voltage charging segments and effective discharging segments, the IC peak value, IC peak voltage, constant current charging duration, constant voltage charging duration, constant voltage rise charging duration, and constant voltage drop discharging duration are extracted.

[0113] The steps for extracting the IC peak value and IC peak voltage are as follows: First, extract the capacity-voltage curve Q(V) of the constant current charging segment. Then, perform smoothing and noise reduction on the capacity-voltage curve. The smoothing and noise reduction can be achieved through moving average, local polynomial smoothing, or other filtering methods that do not alter the voltage monotonicity. Next, calculate the dQ / dV curve based on the smoothed capacity-voltage curve. The maximum value of the dQ / dV curve is taken as the IC peak value, and the voltage value corresponding to this maximum value is taken as the IC peak voltage.

[0114] The steps for extracting the constant current charging time are as follows: Determine the starting point when the current stabilizes at the rated charging current and the ending point when the charging process transitions to constant voltage mode; use the time difference between these two points as the constant current charging time. The steps for extracting the constant voltage charging time are as follows: Take the moment when the voltage reaches the full charge cutoff voltage as the starting point and the moment when the charging current drops to the minimum charging cutoff current as the ending point; use the time difference between these two points as the constant voltage charging time.

[0115] The steps for extracting the constant voltage rise charging time are as follows: During the charging process, select a voltage range where the voltage rises steadily and uniformly, determine the start and end times of this range, and use the time difference between the two as the constant voltage rise charging time. The steps for extracting the constant voltage drop discharging time are as follows: During the discharging process, select a voltage plateau range where the voltage drops steadily and uniformly, determine the start and end times of this range, and use the time difference between the two as the constant voltage drop discharging time.

[0116] The above six health characteristics reflect the changes in charging curves, discharging curves, and battery polarization state with aging, and can be arranged in chronological order to form endogenous variable inputs. Endogenous variable sequences can be established individually for each cell or separately for each cell within a cell cluster. If a certain health feature cannot be extracted within a certain time window, the previous valid window value, interpolated value, or missing value marker can be used for processing.

[0117] Exogenous variable input It consists of calendar aging risk characteristics and operating condition variables. Calendar aging risk characteristics may include continuous uncharged time t, SOC, terminal voltage U, and average voltage U of the same cluster. avg Voltage difference | UU avg The parameters are: internal resistance growth rate (R-R0) / R0, effective charge / discharge cycle count N in the past 30 days, hit count H, and recharge priority score P. Operating condition variables may include temperature, charge / discharge rate, and cycle count. Endogenous and exogenous variable sequences are aligned in the same time order before being input into the prediction model.

[0118] Example 4 like Figure 5 As shown, the prediction model first performs adaptive spectral processing on the input sequence. The input time series sequence x[n] can be an endogenous variable input. The sequence of a single health feature can also be a multi-channel sequence concatenated from multiple health features. For an input time series sequence x[n] of length N, performing a discrete Fourier transform yields a frequency domain sequence X[k], as follows:

[0119] (2)

[0120] In formula (2), x[n] is a time series sequence and X[k] is a frequency domain sequence;

[0121] By converting from the time domain to the frequency domain, long-term low-frequency trends and short-term high-frequency fluctuations can be expressed separately.

[0122] The power sequence P[k] = |X[k]|² is calculated based on the frequency domain sequence X[k]. The power spectrum P is extracted from the power sequence P[k], and a frequency domain mask is generated by combining it with a learnable threshold θ. This mask is used to suppress high-frequency noise components while preserving the effective frequency domain signal. Subsequently, the frequency domain features F extracted from the frequency domain sequence X[k] are multiplied element-wise with the frequency domain mask to obtain the denoised frequency domain sequence F. a The above process suppresses short-term disturbances such as charging and discharging fluctuations and sampling noise before they enter the subsequent prediction network. Specifically:

[0123] (6)

[0124] In formula (6), P is the power spectrum. The learnable threshold is represented by ⊙, which indicates element-wise multiplication.

[0125] The noise reduction frequency domain sequence F a Respectively with global filter W G and local filter W L Multiplying and summing yields the fused frequency domain sequence G[k]. Here, the global filter W... G Local filter W is used to express aging trends over longer time scales. L Used to express local fluctuations on a shorter time scale. Specifically:

[0126] (7)

[0127] Then, an inverse Fourier transform is performed on the fused frequency domain sequence G to obtain the spectral-enhanced time domain sequence S, specifically:

[0128] (8)

[0129] In formula (8), Let S represent IFFT, where S is a spectral enhancement time-domain sequence.

[0130] The spectral-enhanced temporal sequence S is then processed using interactive convolution. This interactive convolution employs one-dimensional convolution with different kernel sizes to simultaneously extract both fine-grained and coarse-grained features. A first one-dimensional convolution kernel is used to convolve the spectral-enhanced temporal sequence S to obtain the first convolutional feature, and a second one-dimensional convolution kernel is used to convolve the spectral-enhanced temporal sequence S to obtain the second convolutional feature. The two convolutional kernels can have different lengths to respectively cover local details and larger time windows.

[0131] The first convolutional feature is element-wise multiplied with the second convolutional feature processed by the activation function to obtain the first interactive feature; the second convolutional feature is then element-wise multiplied with the first convolutional feature processed by the activation function to obtain the second interactive feature. The activation function can be GELU, ReLU, or other non-linear activation functions. The first and second interactive features can be concatenated or added, and then input into a third one-dimensional convolutional kernel to obtain an enhanced temporal representation. This enhanced temporal representation serves as the input to the subsequent CrossLinear prediction part, specifically:

[0132] (9)

[0133] (10)

[0134] In formula (9), The first one-dimensional convolution kernel, For the second one-dimensional convolution kernel, Activate the function;

[0135] In formula (10), 3 represents the fusion convolution kernel, O ICB To enhance the temporal representation.

[0136] Example 5 After enhancing the temporal representation input to the CrossLinear prediction part, the endogenous variable sequence is first processed. Instance normalization is performed to obtain and exogenous variable sequence Perform parameterless instance normalization to obtain Instance normalization can reduce the impact of differences in dimensions and scales between different individual cells and different runtime windows on model training and online inference.

[0137] Then, and Stacked along the time dimension and then subjected to one-dimensional convolution to obtain cross-correlation features. Cross-correlation characteristics This is used to express a stable association between health characteristics and calendar aging risk characteristics, temperature, scalar rate, and cycle number. Then, according to... Generate embedded sequences ,in These are learnable parameters.

[0138] Embedded sequence The sequence is divided into patch sequences of length p, with the number of patches being k = ⌈T / p⌉, where T is the length of the embedding sequence. The patch sequences are projected to obtain patch projection features, and position embeddings (PEs) are introduced. The patch projection features and position embeddings (PEs) are then weighted and fused according to a learnable parameter β. Finally, the fused patch features are concatenated and passed through a linear projection layer to generate the S-step S-time SOH prediction value.

[0139] As an example, the enhanced time-series representation is input into the CrossLinear algorithm for SOH prediction, the endogenous variables are instance-normalized, and the normalized results are output:

[0140] (3)

[0141] In formula (3), To output the normalized result, μ is the mean and σ is the variance. Input for endogenous variables;

[0142] Parametric instance normalization of the exogenous variable sequence yields:

[0143] (4)

[0144] In formula (4), For the output results, Input for exogenous variables;

[0145] The normalized data is input into the backbone network, which outputs the predicted values ​​in the normalized space. The predicted values ​​are then restored using the original mean μ and variance σ to obtain the final results in the real space, as follows:

[0146] (11)

[0147] In formula (11), This represents the computation of the algorithm's backbone network. This refers to the inverse normalization operation. To predict intermediate values, These are the predicted values ​​of the restored endogenous variables.

[0148] Endogenous and exogenous variables are stacked along the time dimension, and the cross-correlation features between the variables are extracted through one-dimensional convolution, as shown below:

[0149] (12)

[0150] In formula (12), As a cross-correlation feature, Represents one-dimensional convolution. Indicates folding along the time dimension;

[0151] The learnable parameter α is weighted and fused with the original endogenous variables and cross-correlation features to obtain the final embedding, as shown below:

[0152] (5)

[0153] Obtain the embedded sequence , in formula (5), These are learnable parameters;

[0154] The embedded sequence is divided into patches of length p, resulting in a total of k = ⌈T / p⌉ patches, as shown below:

[0155] (13)

[0156] Where T is the embedding sequence length, p represents the patch length, and k is the number of patches; in formula (13), Patchify represents the equal division operation;

[0157] The learnable parameter β is weighted and fused with the patch projection features and the location embedding, as shown below:

[0158] (14)

[0159] In formula (14), PE represents the position embedding operation, and Projection1 represents the weighted fusion;

[0160] All patch features are concatenated and then passed through a linear projection layer to generate the future S-step prediction result, as shown below:

[0161] (15)

[0162] In formula (15), Projection2 refers to the linear projection layer, and Concat represents the stitching patch. This is the final output prediction result.

[0163] SOH predictions can be used to display the health status of individual cells, correct the risk ranking of Class A cells, adjust the charging strategy, or make cell cluster scheduling decisions. When running online, BESS can update the Class A / Class B classification results after each scan, and output SOH predictions based on the latest health features and exogenous variables, thus forming a closed-loop process of "calendar aging risk identification - locked charging control - health feature update - SOH prediction - scheduling linkage".

[0164] Example 6 like Figure 6 As shown, this embodiment uses the dream optimization algorithm to determine the hyperparameters of the prediction model. The hyperparameters to be optimized include the learning rate, kernel size, patch length p, batch size, spectral threshold θ, and learnable parameters. The fitness value is determined by at least two of the learnable parameters β. The SOH prediction error of the prediction model on the validation set is used as the fitness value, which can be the mean squared error, mean absolute error, mean absolute percentage error, or a combination of the above errors.

[0165] First, a population is randomly initialized within the hyperparameter search boundary, with each individual in the population representing a set of candidate hyperparameters. For each candidate hyperparameter, a prediction model is trained or invoked, and the SOH prediction error on the validation set is calculated. Then, within the first 90% of the iterations, the population is divided into five subpopulations, each retaining its best individual, and the global best individual is updated accordingly.

[0166] During the global exploration phase, random forgetting and boundary-based supplementation are performed on some dimensions of individuals, causing them to redistribute within the search space. Simultaneously, location information sharing is implemented among different individuals to increase search diversity and reduce the probability of getting trapped in local optima. In the last 10% of iterations, subpopulation partitioning is canceled, and local searches are performed around the globally optimal individual, thereby refining the optimal hyperparameter combinations.

[0167] When the maximum number of iterations is reached, or the fitness value meets the preset termination condition, the dream optimization algorithm outputs the hyperparameter combination corresponding to the globally optimal individual. The prediction model loads this hyperparameter combination to form a deployable model. The deployable model can be stored in the BESS backend server, edge computing device, or battery management system's memory, and can be called upon during subsequent online operation.

[0168] As an example, the sample space is initialized and the population is divided, as shown below:

[0169] (16)

[0170] In formula (16), rand represents the initialization space. Represents an individual vector. Let be a scalar, representing the decision variable (the value of a single dimension) for the i-th individual.

[0171] Each population is divided into five groups for global exploration, retaining the globally optimal individual position, as shown below:

[0172] (17)

[0173] In formula (17), This indicates the initial position after grouping. This indicates the optimal individual position obtained;

[0174] Instead of grouping the population, we explore local optima, as shown below:

[0175] (18)

[0176] In formula (18), This represents the value of the i-th individual and the j-th dimension of the decision variable after this round of iterations. Similarly, This represents the value of the j-th dimension after the current update of a random individual m. This indicates that the random individual m has not been updated in this round and still retains the original value of the j-th dimension from the previous iteration;

[0177] For iterations that do not meet the conditions, the process terminates at the boundary; once the conditions are met, the output is terminated immediately to obtain the optimal hyperparameters, as shown below:

[0178] (19)

[0179] In formula (19), These represent decision variables, i.e., the hyperparameters of the model to be optimized. This indicates the core objective, namely, the prediction error; This refers to inequality constraints and their corresponding penalty weights. Then it refers to the equality constraint and the corresponding penalty weight; That is the ultimate goal of optimization.

[0180] Example 7 After completing the SOH prediction for individual cells and battery clusters, BESS associates and stores the predicted SOH values ​​with the corresponding individual cell's classification label, uncharged duration t, state of charge (SOC), terminal voltage U, voltage difference within the same cluster, internal resistance change rate, and the number of charge-discharge cycles N in the past 30 days. Based on the SOH prediction results, BESS updates the calendar aging risk status of the corresponding individual cell.

[0181] Specifically, for a single cell already marked as a Class A battery, if its predicted State of Health (SOH) value is lower than a preset health threshold, BESS maintains the Class A battery designation for that single cell and increases its charging priority in the charging queue, allowing it to enter the lockout protection, priority charging, constant voltage saturation, and active balancing processes first. If, after charging and active balancing, the predicted SOH value of the single cell is not lower than the preset health threshold, and its continuous uncharged time, State of Charge (SOC), terminal voltage, voltage difference within the same cluster, internal resistance change rate, and charge-discharge cycle count in the past 30 days no longer meet the criteria for a Class A battery, then BESS updates the single cell from Class A to Class B, or maintains it as a Class B battery.

[0182] For a single cell already marked as a Class B battery, if its predicted State of Health (SOH) value is lower than a preset health threshold, or if its subsequent operating data again meets the criteria for a Class A battery, then BESS will update the single cell to a Class A battery and execute discharge circuit blocking, priority sorting, recharging repair, and active balancing according to the Class A battery management process. If its predicted SOH value is not lower than the preset health threshold and does not meet the criteria for a Class A battery, then the single cell's Class B battery status will be maintained, allowing it to continue participating in normal operation or grid-connected scheduling.

[0183] The preset health threshold can be pre-configured by BESS based on battery type, power station operation and maintenance strategy, historical operating data, or rated lifespan threshold. For example, the preset health threshold can be set to 80%, 85%, 90%, or other SOH thresholds suitable for the operation and maintenance management of energy storage power stations. Different power stations can adjust the preset health threshold according to battery chemistry, operating years, dispatch frequency, and safety redundancy requirements.

[0184] Through the above methods, the SOH prediction results are fed back to the calendar aging risk identification and charging priority ranking stages. This enables the method to not only identify low-energy batteries that have been sitting for a long time and charge them for repair, but also to dynamically adjust subsequent management and control strategies based on the battery health status prediction results, thus forming a closed-loop battery health management process of "risk identification - lockout protection - charging balance - SOH prediction - risk status update".

[0185] The various embodiments of the present invention have now been described in detail. To avoid obscuring the concept of the invention, some details known in the art have not been described. Those skilled in the art can fully understand how to implement the technical solutions of this invention based on the above description.

[0186] It should be noted that, in the description of this invention, unless otherwise stated, "a plurality of" means two or more; the terms "upper," "lower," "left," "right," "inner," and "outer," etc., indicating orientation or positional relationships, are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this invention. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0187] Furthermore, the terms "first," "second," and similar terms used in this invention do not indicate any order, quantity, or importance, but are merely used to distinguish different parts. "Vertical" is not strictly vertical, but within the permissible range of error. "Parallel" is not strictly parallel, but within the permissible range of error. Terms such as "including" or "comprising" mean that the element preceding the word encompasses the element listed after the word, and do not exclude the possibility of encompassing other elements as well.

[0188] It should also be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "joining" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a direct connection or an indirect connection through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms in this invention depending on the specific circumstances. When a specific device is described as being located between a first device and a second device, an intermediary device may or may not be present between the specific device and the first or second device.

[0189] All terms used in this invention have the same meaning as understood by one of ordinary skill in the art to which this invention pertains, unless otherwise specifically defined. It should also be understood that terms defined in general dictionaries should be interpreted as having meanings consistent with their meanings in the context of the relevant art, and not as idealized or highly formalized, unless expressly defined herein.

[0190] Techniques, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, they should be considered part of the specification.

[0191] While specific embodiments of the present invention have been described in detail by way of examples, those skilled in the art should understand that the above examples are for illustrative purposes only and are not intended to limit the scope of the invention. Those skilled in the art should understand that modifications can be made to the above embodiments or equivalent substitutions can be made to some technical features without departing from the scope and spirit of the invention. In particular, as long as there is no structural conflict, the various technical features mentioned in the embodiments can be combined in any manner.

Claims

1. A method for predicting and controlling the state of harm (SOH) of energy storage batteries considering calendar aging, characterized in that, The method, performed by a battery energy storage system on battery clusters and individual cells in an energy storage power station, includes: S1: Collect the continuous uncharged time t, state of charge (SOC), terminal voltage U, and average voltage U of each individual cell. avg Current internal resistance R, initial internal resistance R0, effective charge / discharge cycle count N in the past 30 days, charge / discharge voltage and current time series, temperature and charge / discharge rate; S2: Based on the continuous uncharged duration t, state of charge (SOC), terminal voltage U, and average voltage of the same cluster U... avg The current internal resistance R, the initial internal resistance R0, and the effective charge-discharge cycle count N in the past 30 days are used to generate the hit count H. Specifically, a first hit is generated when the continuous uncharged duration t is not less than 7 days; a second hit is generated when the SOC is less than 20%; a third hit is generated when the terminal voltage U is less than the undervoltage threshold of the corresponding battery system; and a fourth hit is generated when |UU avg A fourth hit value is generated when the value is greater than 50mV or (R-R0) / R0 is not less than 15%. A fifth hit value is generated when the effective charge-discharge cycle count N in the past 30 days is less than 2. Each hit value is 1 when the corresponding judgment condition is met and 0 when the corresponding judgment condition is not met. The number of hits H is the sum of the first hit value to the fifth hit value. Individual cells with a hit count H of not less than 2 are marked as Class A cells, and individual cells with a hit count H of less than 2 are marked as Class B cells. S3: Write the discharge lockout state to the Class A battery, calculate the charging priority score of the Class A battery, and charge the Class A battery with constant current in descending order of the charging priority score. After charging to 100% SOC or reaching the full charge cutoff voltage, switch to the constant voltage saturation stage, and perform active balancing on the individual cells in the same cluster during the constant voltage saturation stage. Once all Class A batteries in the same battery cluster have completed constant voltage saturation and been marked as activated batteries, the discharge lockout state of that battery cluster is released. When the continuous uncharged time of the activated battery exceeds 5 days or the SOC is not greater than 30%, the activated battery is recharged to a SOC of 50%. S4: Extract health features from the charge / discharge voltage and current time series, form an endogenous variable sequence according to the time order, and form an exogenous variable sequence with calendar aging risk features, temperature, charge / discharge rate and cycle number according to the same time order. The endogenous and exogenous variable sequences are input into a prediction model whose hyperparameters are determined by the dream optimization algorithm. Adaptive spectral processing, interactive convolution processing, cross-embedding of endogenous and exogenous variables, and linear projection are performed sequentially to output the SOH prediction value at the current time or in the future S steps.

2. The method according to claim 1, characterized in that, The priority score P for power replenishment is calculated according to the following formula: （1） In formula (1), w1, w2, w3, and w4 are non-negative weights and w 1+ w 2+ w 3+ w4=1; when the current resistance R is not sampled, set w4=0 and renormalize w1, w2 and w3.

3. The method according to claim 1, characterized in that, The constant current charging includes: when the power station scheduling pressure is lower than a preset pressure threshold, a charging rate of no more than 0.2C is adopted; when the power station scheduling pressure is not lower than the preset pressure threshold, a charging rate of 0.3C to 0.5C is adopted; the active balancing continues until the difference between the maximum and minimum terminal voltages of the individual cells in the same cluster is no greater than the balancing cutoff voltage difference.

4. The method according to claim 1, characterized in that, The health characteristics include IC peak value, IC peak voltage, constant current charging duration, constant voltage charging duration, constant voltage rise charging duration, and constant voltage drop discharging duration. Specifically, the IC peak value and IC peak voltage are obtained by smoothing the capacity Q-voltage V curve of the constant current charging segment and calculating the dQ / dV curve. The constant current charging duration is the time difference from the start of stable maintenance of the rated charging current to the end of transitioning to constant voltage mode. The constant voltage charging duration is the time difference from reaching the full charge cutoff voltage to the moment the charging current drops to the minimum cutoff current. The constant voltage rise charging duration is the start and end time difference of the voltage steady rise interval during charging. The constant voltage drop discharging duration is the start and end time difference of the voltage plateau interval during discharging.

5. The method according to claim 1, characterized in that, The adaptive spectral processing includes: performing a discrete Fourier transform on the input time series sequence x[n] to obtain the expression for the frequency domain sequence X[k], as follows: (2) The power sequence P[k] = |X[k]|² is calculated based on the frequency domain sequence X[k], and a frequency domain mask is generated based on a learnable threshold θ. The frequency domain features F extracted from the frequency domain sequence X[k] are multiplied element-wise with the frequency domain mask to obtain the denoised frequency domain sequence F. a ;Denoising frequency domain sequence F a respectively with global filter W G and local filter W L Multiply and sum to obtain the fused frequency domain sequence G; perform an inverse Fourier transform on the fused frequency domain sequence G to obtain the spectral enhancement time domain sequence S.

6. The method according to claim 5, characterized in that, The interactive convolutional processing includes: convolving the spectral enhancement temporal sequence S with a first one-dimensional convolutional kernel to obtain a first convolutional feature; convolving the spectral enhancement temporal sequence S with a second one-dimensional convolutional kernel to obtain a second convolutional feature; multiplying the first convolutional feature element-wise with the second convolutional feature processed by the activation function to obtain a first interactive feature; multiplying the second convolutional feature element-wise with the first convolutional feature processed by the activation function to obtain a second interactive feature; concatenating or adding the first and second interactive features and inputting them into a fusion convolutional kernel to obtain the enhanced temporal representation O. ICB .

7. The method according to claim 6, characterized in that, The cross-embedding and linear projection of endogenous and exogenous variables include: performing instance normalization on the endogenous variable sequence to obtain: （3） In formula (3), To output the normalized result, μ is the mean and σ is the variance. Input for endogenous variables; Parametric instance normalization of the exogenous variable sequence yields: （4） In formula (4), For the output results, Input for exogenous variables; Stacked along the time dimension and Cross-correlation features are obtained through one-dimensional convolution. ;according to: （5） Obtain the embedded sequence , in formula (5), These are learnable parameters; Embedded sequence The sequence is divided into patch sequences of length p. The patch projection features and position embeddings are fused using a learnable parameter β. The fused patch features are then concatenated and passed through a linear projection layer to generate normalized spatial prediction values. The normalized spatial prediction values ​​are then denormalized using the mean and variance of the endogenous variable sequence to obtain the SOH prediction values.

8. The method according to claim 7, characterized in that, The dream optimization algorithm uses the SOH prediction error of the prediction model on the validation set as the fitness value, and the learning rate, kernel size, patch length p, batch size, spectral threshold θ, and learnable parameters as the fitness values. At least two of the learnable parameters β are selected as hyperparameters to be optimized. The dream optimization algorithm includes: randomly initializing the population within the hyperparameter search boundary; dividing the population into five subpopulations in the first 90% of iterations, retaining the best individual in each subpopulation and updating the globally optimal individual; performing random forgetting and boundary-based supplementation on some dimensions of the individuals, and sharing position information among the individuals; canceling the population division in the last 10% of iterations and performing a local search around the globally optimal individual; and outputting the hyperparameter combination corresponding to the globally optimal individual when the iteration termination condition is met or the validation set error no longer decreases.

9. The method according to claim 1, characterized in that, After obtaining the SOH prediction value, the calendar aging risk status of the corresponding individual battery is updated according to the SOH prediction value, and the charging priority of the individual battery is adjusted according to the updated calendar aging risk status.

10. The method according to claim 9, characterized in that, When the predicted SOH value is lower than the preset health threshold, the corresponding single battery cell is maintained as a Class A battery, or the charging priority of the single battery cell is increased; when the predicted SOH value is not lower than the preset health threshold, and the single battery cell no longer meets the determination conditions for a Class A battery, the single battery cell is maintained as a Class B battery or is updated from a Class A battery to a Class B battery.

Images