A method for adaptive equalization control of energy storage battery pack capacity

Abstract

The application belongs to the technical field of energy storage batteries, and particularly relates to a capacity self-adaptive equalization regulation method for an energy storage battery pack, which comprises the following steps: step one: during the operation of the energy storage system, the cumulative throughput of each battery monomer, temperature fluctuation records and terminal voltage behavior parameters are collected after each charge-discharge cycle, the monomer aging characteristic fingerprint is constructed, the remaining effective charge-discharge cycle number of each battery monomer expected to reach the retirement health state threshold is deduced, and the retirement time interval difference ordering vector of each monomer is output; step two: the retirement time interval difference ordering vector output in step one is taken as a state space, the output proportion coefficient adjustment amount of each battery monomer in the next charge-discharge cycle is taken as an action space, a multi-agent deep deterministic policy gradient framework is used to train a life equalization strategy model, and after the training is completed, the strategy model is operated online, and the output adjustment vector of the next stage is output according to the current state.

Description

Technical Field

[0001] This invention belongs to the field of energy storage battery technology, specifically a method for adaptive equalization control of energy storage battery pack capacity. Background Technology

[0002] As the core energy storage unit of energy storage power stations, photovoltaic energy storage systems, new energy backup power sources and large-scale industrial and commercial energy storage equipment, energy storage battery packs are usually integrated by connecting multiple individual lithium batteries in series and in parallel, and have the characteristics of large capacity, high power and long-term continuous charging and discharging.

[0003] With the large-scale grid connection of renewable energy power generation and the continuous growth of grid frequency regulation and peak shaving needs, energy storage systems, especially battery energy storage systems represented by lithium-ion batteries, have become a key link in supporting the transformation of energy structure. In practical applications, in order to meet voltage and power level requirements, dozens or even hundreds of individual batteries are usually connected in series or parallel to form a battery pack.

[0004] However, due to differences in manufacturing processes, initial capacity, internal resistance, and uneven temperature distribution and inconsistent depth of charge and discharge during operation, the state of each individual cell in the battery pack will inevitably exhibit nonlinear and time-varying differences. The aggravation of these differences may also lead to local overcharging or over-discharging, accelerate battery aging, and even induce thermal runaway safety issues, resulting in a gradual decline in the balancing effect under long-term operation, ultimately limiting the lifespan and economic efficiency of the energy storage system. To address this, the present invention provides an adaptive balancing control method for the capacity of energy storage battery packs. Summary of the Invention

[0005] In order to overcome the shortcomings of the prior art, at least one technical problem raised in the background art is solved.

[0006] The technical solution adopted by this invention to solve its technical problem is: an adaptive equalization control method for the capacity of an energy storage battery pack, comprising the following steps: Step 1: During the operation of the energy storage system, the cumulative throughput, temperature fluctuation records and terminal voltage behavior parameters of each battery cell are collected after each charge and discharge cycle, with the charge and discharge cycle as the time unit. The aging characteristic fingerprint of each cell is constructed, the remaining effective charge and discharge cycles of each battery cell are estimated to reach the retirement health state threshold, and the retirement time difference sorting vector of each cell is output. Step 2: Using the retirement time difference sorting vector output in Step 1 as the state space, and the adjustment amount of the output ratio coefficient of each battery cell in the next charge-discharge cycle as the action space, the lifetime balancing strategy model is trained using a multi-subject deep deterministic strategy gradient framework. After training, the strategy model runs online and outputs the output adjustment vector for the next stage based on the current state. Step 3: Use the output adjustment vector output in Step 2 as the macro trend command. In each charge and discharge cycle, use the model predictive control framework to perform rolling correction on the macro trend command. During the correction process, the safety threshold of the battery cell is used as the constraint condition, and the final output coefficient of each cell in this cycle is output. Step 4: When the final output coefficient causes energy deviation when each individual battery performs charging and discharging tasks, during the rest period after the end of the charging and discharging cycle, cross-cycle energy calculation is performed using a distributed floating energy storage unit cluster relay method to achieve local energy transfer. Step 5: Repeat steps 1 to 4, with the goal of minimizing the standard deviation of the retirement time difference of each battery cell as the long-term optimization objective. Dynamically adjust the output coefficient of each cell in each cycle so that the capacity decay trajectory of each battery cell tends to be synchronized throughout the entire life cycle.

[0007] Preferably, in step one, the individual aging feature fingerprint is constructed by recording the cumulative throughput, temperature fluctuation, and terminal voltage behavior parameters at the end of each charge-discharge cycle. The parameters are aggregated to form a feature vector, and the differential decay factor is obtained by normalizing the ratio of the statistical rate of change of the feature vector to the standard rate of change within the sliding time window.

[0008] Preferably, the method for extrapolating the remaining effective charge-discharge cycles based on aging feature fingerprints in step one is as follows: using the aging feature fingerprint sequence of a battery cell in the most recent few cycles as the original data, estimating the predicted value of future cycles through cumulative generation and exponential trend extrapolation models, comparing the predicted value with the retirement health status threshold to obtain the remaining cycle number, updating the data and re-estimating the model parameters after each cycle, and calculating the difference between each cell and the median of the remaining cycle number of all cells as the benchmark to form a sorting vector.

[0009] Preferably, the adjustment amount of the output ratio coefficient in step two satisfies: The sum of the output ratio coefficients of all individual units remains 1 before and after the adjustment; When the health status of a single cell is below the first threshold, its adjustment amount is limited to a non-positive value; When the health status of a single entity is above the second threshold, its adjustment amount is allowed to be positive and the upper limit is relaxed.

[0010] Preferably, in step two, the multi-agent deep deterministic policy gradient framework adopts a centralized training and decentralized execution architecture: during the training phase, each agent acts as an independent agent and shares a centralized commentator network that can access global information; during the execution phase, each agent outputs actions based only on local observation information; and the policy model is trained offline using historical data during the initial deployment phase.

[0011] Preferably, the model predictive control framework in step three is implemented as follows: the prediction time domain width is set to several charge and discharge cycles. At the beginning of each decision cycle, a constrained quadratic programming problem is constructed based on the current state of charge and health. The optimization variable is the output coefficient fine-tuning amount for multiple future cycles. The objective function is to minimize the weighted sum of the macroscopic command tracking error and the output fluctuation penalty term. The constraints include the discharge depth safety threshold and the voltage safety boundary. The fine-tuning amount obtained in the first cycle is superimposed with the macroscopic command as the final output coefficient for this cycle. The solution is rolled over again in the next cycle.

[0012] Preferably, the distributed floating energy storage unit in step four satisfies the following: the capacity of each energy anchor point is a percentage of the capacity of a single cell in the battery pack; it uses supercapacitors, lithium-ion capacitors, or small lithium batteries; the operating voltage window matches the nominal voltage of the battery cell; each anchor point is electrically isolated; it is not grounded or directly connected to the main power circuit; and each anchor point is equipped with an independent management system to maintain its state of charge within a preset operating window.

[0013] Preferably, the cluster relay method in step four includes: converting the final output coefficient into the planned charging and discharging energy of each individual unit before the start of each cycle; each individual unit performs charging and discharging according to the plan during the cycle; allowing temporary deviations without real-time intervention; calculating the total deviation between the actual and planned energy of the entire group during the rest period; and triggering relay equalization if the deviation exceeds the threshold. The relay balancing process includes: identifying sets of surplus and deficit cells, controlling nearby anchor points to draw energy from surplus cells up to the upper limit, disconnecting the anchor points and transmitting the energy to another anchor point via an isolation bus, and then releasing the energy to the deficit cells from that anchor point, repeating this process until the total deviation is less than the threshold or the rest period ends.

[0014] Preferably, the collaborative operation mode of energy transfer during the rest period in step four is as follows: when multiple groups of surplus and deficit cells need to be processed simultaneously, as long as the battery local clusters connected to each anchor point are electrically isolated and have no shared path, multiple anchor points are allowed to perform energy absorption or release in parallel. The multiple relay transmissions on the isolation bus adopt time-division multiplexing, allocate non-overlapping time windows for each source anchor point and target anchor point and insert protection intervals. The management system of each anchor point monitors the state of charge and temperature in real time. When the safety threshold is exceeded, participation in parallel exchange is automatically suspended and rejoined after recovery.

[0015] Preferably, the synchronous retirement determination condition for the battery pack in step five is as follows: when the ratio of the maximum to the minimum of the remaining operational charge-discharge cycles calculated by the system for all battery cells is lower than a set ratio, and the health status assessment value of each cell is lower than the retirement threshold, the synchronous retirement window is determined. When the condition is met for the first time, a retirement takeover report is automatically generated, which includes a tiered list of the remaining available capacity of each cell and a recommendation of tiered utilization scenarios.

[0016] The beneficial effects of this invention are as follows: 1. The energy storage battery pack capacity adaptive equalization control method of the present invention abandons the traditional voltage / SOC surface equalization logic, and optimizes the synchronicity of single cell decay from the perspective of the whole life cycle based on the remaining life prediction mechanism of aging characteristic fingerprint, so as to solve the problem of long-term equalization effect decay. It adopts a two-layer optimization architecture of MADDPG macro-control and MPC micro-correction, which takes into account both the global life-cycle optimization and single-cycle safe operation, and breaks through the rigidity defect of traditional fixed threshold equalization.

[0017] 2. The energy storage battery pack capacity adaptive equalization control method described in this invention avoids the problems of high energy consumption and large disturbances in real-time equalization by innovatively using distributed floating anchor point clusters for cross-cycle relay equalization. The energy transfer efficiency is greatly improved, and the battery pack cells are eventually aged and retired synchronously. This significantly improves the utilization rate and operating economy of the energy storage system throughout its entire life cycle, reduces the risk of overcharging, over-discharging, and thermal runaway of individual cells from the root, and is suitable for long-term stable operation under various complex energy storage conditions. Attached Figure Description

[0018] The invention will now be further described with reference to the accompanying drawings.

[0019] Figure 1 This is a schematic diagram of the overall closed-loop process for adaptive equalization control of energy storage battery pack capacity in this invention. Figure 2 This is a schematic diagram of the process for aging characteristic deduction and dual-layer output optimization regulation in this invention; Figure 3 This is a schematic diagram of the process for determining relay balancing and synchronous decommissioning in a distributed cluster in this invention. Detailed Implementation

[0020] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0021] like Figure 1-3 As shown in the embodiment of the present invention, an adaptive equalization control method for the capacity of an energy storage battery pack includes the following steps: Step 1: During the operation of the energy storage system, the cumulative throughput, temperature fluctuation records and terminal voltage behavior parameters of each battery cell are collected after each charge and discharge cycle, with the charge and discharge cycle as the time unit. The aging characteristic fingerprint of each cell is constructed, the remaining effective charge and discharge cycles of each battery cell are estimated to reach the retirement health state threshold, and the retirement time difference sorting vector of each cell is output. Aging Feature Acquisition and Remaining Life Prediction: Taking a single complete charge-discharge cycle as the smallest time unit, i.e., 1 hour of charging + 0.5 hours of rest + 1 hour of discharging, after each charge-discharge cycle, three core operating parameters are collected simultaneously for each of the 16 battery cells: cumulative throughput, extreme temperature fluctuation within the cycle, and real-time terminal voltage change curve. Based on the multi-dimensional operating parameters, an aging feature fingerprint for each cell is constructed to accurately quantify the differentiated aging degree of each cell. Through a time-series prediction model, the remaining effective charge-discharge cycles for each cell to reach the retirement health state threshold are predicted. The retirement health state threshold is SOH=80%. The cells in the group are sorted according to the number of remaining cycles, and the retirement time difference sorting vector is output to accurately represent the differentiated distribution of the lifespan of the cells in the group. Step 2: Using the retirement time difference sorting vector output in Step 1 as the state space, and the adjustment amount of the output ratio coefficient of each battery cell in the next charge-discharge cycle as the action space, a lifetime balancing strategy model is trained using a multi-subject deep deterministic strategy gradient framework. After training, the strategy model runs online, reads the current individual cell lifetime differentiation status in real time, and outputs the output adjustment vector for the next stage based on the current status, thereby realizing the macroscopic output differentiation ratio based on the remaining lifetime. Step 3: The output adjustment vector output in Step 2 is used as the macro trend command. In order to eliminate the control deviation caused by the fluctuation of single-cycle operating conditions and safety boundary constraints, the macro trend command is rolled over and corrected in each charge and discharge cycle using a model predictive control framework. During the correction process, the safety threshold of the battery cell is used as the constraint condition. The safety threshold is strictly bound to the overcharge, over-discharge, over-temperature, and over-voltage of the cell. A constrained optimization solution model is constructed to output the final output coefficient of each cell in this cycle, taking into account both life balance and operational safety. Step 4: When the final output coefficient causes energy deviation when each individual battery performs charging and discharging tasks, temporary energy deviation will occur due to real-time operating condition disturbances when each individual battery performs charging and discharging tasks according to the final output coefficient. In order to avoid real-time balance interference with main power operation and reduce energy consumption, the deviation is not corrected in real time during the charging and discharging process. During the rest period after the end of the charging and discharging cycle, cross-cycle energy calculation is performed using a distributed floating energy storage unit cluster relay method to realize local energy transfer and eliminate the energy imbalance problem caused by single-cycle output regulation. Step 5: Repeat steps 1 to 4, with the long-term optimization goal of minimizing the standard deviation of the retirement time difference of each battery cell, dynamically adjust the output coefficient of each cell in each cycle, gradually reduce the difference in capacity decay rate of each cell, so that the capacity decay trajectory of each battery cell tends to be synchronized throughout the entire life cycle, and finally achieve synchronous retirement of the battery pack. The core formula for the long-term optimization objective is: In the formula: The standard deviation of the time interval between individual decommissioning events; This represents the total number of individual cells in the battery pack. For the first The number of remaining effective charge / discharge cycles for a single unit; This represents the average number of remaining cycles for all individual cells in the group.

[0022] like Figure 1-3 As shown, in step one, the single-cell aging feature fingerprint is constructed by recording the cumulative throughput, temperature fluctuation and terminal voltage behavior parameters at the end of each charge-discharge cycle. The parameters are aggregated to form a feature vector, and the differential aging factor is obtained by normalizing the ratio of the statistical rate of change of the feature vector to the standard rate of change within the sliding time window. The larger the value, the more severe the accelerated aging. By aggregating multi-dimensional operating parameters, a unique aging feature fingerprint for each individual cell is constructed, accurately quantifying the degree of nonlinear accelerated aging of the individual cell. The specific implementation method is as follows: Feature vector construction: After each charge / discharge cycle, the first feature vector is collected. The three core parameters of a single block: cumulative throughput per cycle. Temperature fluctuation amplitude within the period Maximum deviation of periodic terminal voltage The aggregated features form a single-cycle aging feature vector: ; Sliding window feature statistics: Set a fixed sliding time window, take 10 consecutive charge-discharge cycles, and count the real-time change rate of feature vectors within the window. At the same time, pre-store the standard feature change rate of brand-new battery cells in the same batch. ; Differential decay factor calculation: The differential decay factor is obtained by combining the ratio of the real-time characteristic change rate to the standard change rate with normalization processing. Core calculation formula: In the formula: This represents the number of sliding window cycles. For the first Cycle number Changes in the eigenvectors of a single entity; This represents the change in the periodic eigenvector corresponding to a standard single entity; ; The larger the attenuation factor value, the greater the rate of degradation of the cell's temperature fluctuation, power throughput, and voltage deviation compared to the standard state, and the more severe the accelerated aging. In subsequent regulation, it is necessary to reduce its output load to suppress the acceleration of aging. Compared with the single SOH characterization of aging state, this fingerprint feature can accurately capture the subtle differences in nonlinear and time-varying aging of the cell and predict the aging trend in advance.

[0023] like Figure 1-3 As shown, the method for extrapolating the remaining effective charge-discharge cycles based on aging feature fingerprints in step one is as follows: using the aging feature fingerprint sequence of a battery cell in the most recent few cycles as the original data, the predicted value of future cycles is estimated through cumulative generation and exponential trend extrapolation model, the predicted value is compared with the retirement health status threshold to obtain the remaining cycle number, the data is updated and the model parameters are re-estimated after each cycle, and the difference between each cell and the median of the remaining cycle number of all cells is calculated based on the median to form a sorting vector; Based on aging feature fingerprint time series data, the remaining effective charge-discharge cycles are predicted, and a retirement time difference sorting vector is accurately constructed. The specific implementation method is as follows: Raw data acquisition: Extract the differential decay factor sequence of the most recent 15 cycles for each individual unit as the raw time-series data for aging trend prediction. ; Accumulation generation preprocessing: A first-order accumulation generation operator is used to weaken the random fluctuations in time series data, resulting in a smoothed aging trend sequence. The calculation formula is as follows: In the formula: For the first Before single-unit The cumulative value of the decay factor over each cycle; Exponential trend extrapolation prediction: Based on the preprocessed smooth sequence, an exponential trend extrapolation model is constructed, and the formula for predicting the aging trend of individual cells is obtained by fitting. The predicted value of the decay factor for future cycles is iteratively predicted until the predicted decay factor reaches the critical value corresponding to the retirement threshold. Remaining cycle calculation: Calculate the total predicted number of cycles from the current cycle until the decommissioning threshold is reached; this is the remaining effective charge / discharge cycle count for a single cell. ; Sort vector construction: Calculate the median of the remaining periods for all individual units in the group. Using the difference between the remaining number of cycles of each individual unit and the median as the distance parameter, a sorting vector of retirement time distance is constructed. The positive and negative values ​​in the vector represent that the lifetime of a single unit is better or worse than the average level, respectively, providing a precise state basis for subsequent intelligent regulation.

[0024] like Figure 1-3 As shown, the adjustment amount of the output ratio coefficient in step two satisfies: The sum of the output ratio coefficients of all individual units remains 1 before and after the adjustment; When the health status of a single cell is below the first threshold, its adjustment amount is limited to a non-positive value; When the health status of a single entity is higher than the second threshold, the adjustment amount is allowed to be positive and the upper limit is relaxed; To ensure compliant, safe, and balanced power output control, rigid constraints are set on the adjustment of the power output ratio coefficient for each individual unit, ensuring the stability of the overall group's power output and the rationality of differentiated control. The specific constraint rules and formulas are as follows: Let the first one before adjustment be... The single-unit output ratio coefficient is The adjustment amount is The adjusted output coefficient is It satisfies three core constraints: Total output conservation constraint: The sum of the output ratio coefficients of all individual units in the group remains constant, ensuring that the total power of the system remains unchanged. Formula: From this, the constraint of conservation of output adjustment can be derived: To prevent fluctuations in the total output of the system during the regulation process; Degraded monomer constraint: Set the first health threshold SOH1=85%. When monomer SOH<85%, aging accelerates and lifespan is shortened, limiting the output adjustment amount. It can only maintain or reduce output to avoid overloading of the weaker units and accelerating aging. Superior monomer constraint: Set a second health threshold SOH2=92%. When monomer SOH>92%, the condition is excellent and the lifetime is sufficient, and the output adjustment amount is adjusted accordingly. Positive values ​​can be taken, and the upper limit can be increased to 0.05 to fully utilize the capacity of high-quality cells and balance the aging rate of the entire group.

[0025] like Figure 1-3 As shown, the multi-agent deep deterministic policy gradient framework in step two adopts a centralized training and decentralized execution architecture: during the training phase, each agent acts as an independent agent and shares a centralized commentator network that can access global information; during the execution phase, each agent outputs actions based only on local observation information; the policy model is trained offline using historical data during the initial deployment phase, and after solidification, it only performs forward inference online. Step two employs the MADDPG architecture, which features centralized training and distributed execution. This architecture is adapted to multi-agent collaborative balance control, addressing the shortcomings of traditional single-agent systems that cannot simultaneously address the global optimization of multiple agents. The specific implementation method is as follows: Architecture Deployment: Each battery cell corresponds to an independent intelligent agent, with a total of 16 intelligent agents in the group. The system is configured with one centralized commentator network and 16 distributed executor networks. Offline training phase: Offline training is carried out using 1,000 sets of complete historical data from the energy storage system. The centralized commentator network can globally read all the aging characteristics, remaining cycles, output status, and safety parameters of all agents, uniformly evaluate the benefits of each agent's actions, and iteratively optimize the network parameters of each agent. The training objective is to minimize the standard deviation of the retirement time of the entire group and maximize the system's safety redundancy. After training, all network parameters are solidified to prevent online parameter fluctuations. In the online execution phase: the trained and solidified model only performs forward inference. The distributed agents corresponding to each unit complete the action output based only on the aging fingerprint, remaining cycle, and real-time state of the local unit. It does not need to rely on global data, which greatly reduces the computing power consumption of BMS. Each cycle adaptively outputs the optimal output adjustment amount, realizing decentralized precise control and taking into account both the real-time control and global optimization.

[0026] like Figure 1-3 As shown, the implementation of the model prediction control framework in step three is as follows: the prediction time domain width is set to several charge and discharge cycles. At the beginning of each decision cycle, a constrained quadratic programming problem is constructed based on the current state of charge and health. The optimization variable is the output coefficient fine-tuning amount for multiple future cycles. The objective function is to minimize the weighted sum of the macroscopic command tracking error and the output fluctuation penalty term. The constraints include the discharge depth safety threshold and the voltage safety boundary. Only the fine-tuning amount obtained in the first cycle is superimposed with the macroscopic command as the final output coefficient for this cycle. The solution is rolled over again in the next cycle. By using the MPC framework to perform micro-rolling corrections on macro-output commands, single-cycle operating condition disturbance deviations are eliminated, and a quadratic programming optimization model with safety constraints is constructed. The specific implementation method is as follows: Parameter configuration: Set the prediction time domain width P = 5 charge and discharge cycles, and the control time domain M = 1 cycle. That is, each time the optimal output fine-tuning amount for the next 5 cycles is optimized, only the correction instruction for the first cycle is executed, and the solution is rolled over again in the next cycle. Optimization variables: Define the fine-tuning amount of the unit output coefficient for each cycle in the future prediction time domain. Optimize variables as the core; Objective function construction: The dual optimization objectives are minimizing macroscopic command tracking error and minimizing output fluctuation. The formula is as follows: In the formula: For the macroscopic output command of MADDPG; , The weighting coefficient is determined by taking... =0.7、 =0.3, prioritizing the achievement of macroeconomic life control targets while suppressing frequent fluctuations in power output; Safety constraints: Discharge depth constraint: The standard setting is a maximum discharge depth of 85%. Voltage constraint: The voltage constraint for lithium iron phosphate cells is 2.5V~3.65V; Output coefficient constraint: superimposed adjustment constraint rules; Rolling execution: At the beginning of each cycle, a quadratic programming problem is solved, and the fine-tuning amount of the first cycle is superimposed with the macro output command to obtain the final output coefficient of the current cycle. The cycle is rolled and iterated to achieve a precise match between macro life optimization and micro safe operation.

[0027] like Figure 1-3 As shown, the distributed floating energy storage unit in step four satisfies the following requirements: the capacity of each energy anchor point is a small percentage of the capacity of a single cell in the battery pack; it uses supercapacitors, lithium-ion capacitors, or small lithium batteries; the operating voltage window matches the nominal voltage of the battery cell; each anchor point is electrically isolated; it is not grounded or directly connected to the main power circuit; each anchor point is equipped with an independent management system to maintain its state of charge within a preset operating window; when it deviates, it injects or absorbs energy into the isolation bus through a bidirectional isolation unit to restore its state of charge. The specific configuration parameters and control methods are as follows: Capacity configuration: The capacity of each floating energy storage anchor point is 6% of the rated capacity of a single cell, with a single cell capacity of 20Ah and a single anchor point capacity of 1.2Ah. High-response-speed lithium-ion capacitors are selected to adapt to short-time energy relay scenarios. Electrical parameters: The anchor point operating voltage window is 2.8V~3.5V, which is fully matched with the nominal voltage of the energy storage cell. All anchor points are electrically isolated from each other, not grounded, and not directly connected to the main power circuit of the battery to prevent interference with the main charging and discharging operation of the system during the equalization process. SOC steady-state control: Each anchor point is equipped with an independent embedded management system. The preset optimal SOC working window for the anchor point is 40%~60%. When the anchor point SOC is greater than 60%, excess energy is released to the isolation bus through the bidirectional isolated DC-DC unit. When the anchor point SOC is less than 40%, a small amount of energy is drawn from the isolation bus to replenish power, always keeping the anchor point in the optimal standby state and ensuring that it can participate in energy relay at any time.

[0028] like Figure 1-3 As shown, the cluster relay method in step four includes: converting the final output coefficient into the planned charging and discharging energy of each individual unit before the start of each cycle; each individual unit performs charging and discharging according to the plan during the cycle, allowing temporary deviations without real-time intervention; calculating the total deviation between the actual and planned energy of the entire group during the rest period; and triggering relay equalization if the deviation exceeds the threshold. The relay balancing execution includes: identifying sets of surplus and deficit cells, controlling nearby anchor points to draw energy from surplus cells to the upper limit, disconnecting and transmitting it to another anchor point through an isolation bus, and then releasing it to the deficit cells from that anchor point, repeating until the total deviation is less than the threshold or the rest period ends; By adopting a cluster relay balancing mode that uses periodic planned output and concentrated correction during rest periods, the problems of high energy consumption and large disturbances in real-time balancing are avoided. The specific implementation process is as follows: Planned Energy Conversion: Before the start of each cycle, the final output coefficient of each unit is converted into the corresponding planned charge and discharge energy. The system performs charging and discharging according to the planned energy, without real-time energy correction during operation, and allows temporary energy deviations; Deviation determination: After the cycle ends and the rest period begins, the actual charge and discharge energy of each cell is collected. Calculate the energy deviation of a single unit Calculate the total absolute deviation of the entire group. Set the deviation threshold to 0.5Ah. If the total deviation exceeds the threshold, relay equalization will be triggered. Energy cluster partitioning: Individual clusters are divided based on the sign of the deviation. It is a collection of surplus energy units, with actual output exceeding the planned output, resulting in energy redundancy. As a collection of energy-deficient units, the actual output is less than planned, resulting in insufficient electricity. Relay transfer execution: Prioritize calling the floating anchor point near the surplus cell, draw excess energy from the surplus cell to the anchor point's SOC limit, then disconnect from the surplus cell, and transfer the energy to the nearby anchor point corresponding to the deficit cell through the isolation bus. The anchor point releases energy to make up for the deficit in the cell. The process of drawing, transferring, and releasing is executed iteratively until the total energy deviation of the entire group is lower than the threshold or the rest period ends, effectively eliminating cross-cycle energy imbalance.

[0029] like Figure 1-3 As shown, the collaborative operation mode of energy transfer during the rest period in step four is as follows: When multiple groups of surplus and deficit cells need to be processed simultaneously for energy transfer, as long as the battery local clusters connected to each anchor point are electrically isolated and have no shared path, multiple anchor points are allowed to perform energy absorption or release in parallel. The multiple relay transmissions on the isolation bus adopt time-division multiplexing, allocating non-overlapping time windows and inserting protection intervals for each source anchor point and target anchor point. The management system of each anchor point monitors the state of charge and temperature in real time. When the safety threshold is exceeded, participation in parallel exchange is automatically suspended, and the anchor point rejoins after recovery. The specific collaborative operation mode is as follows: Parallel admission conditions: If multiple groups of surplus and deficit cells are electrically completely isolated from their corresponding local clusters and have no shared transmission paths, multiple anchor points are allowed to perform energy harvesting and release operations simultaneously, which greatly improves the equalization efficiency; Bus time-division multiplexing mechanism: For multiple relay tasks sharing a common isolation bus, a time-division multiplexing allocation strategy is adopted to allocate an independent and non-overlapping time window for each transmission task. The duration of a single window is 20ms, and a 5ms protection interval is set between adjacent windows to eliminate bus energy crosstalk and signal interference problems. Safety self-healing mechanism: Each anchor point management system monitors its own SOC, operating temperature, and output current parameters in real time. If the temperature exceeds 50℃, the SOC exceeds the operating window, or an overcurrent abnormality occurs, the current transmission task will be automatically paused immediately. After the parameters return to the normal range, the task will be re-entered into the parallel balancing queue to continue execution, ensuring the safety and stability of the balancing process.

[0030] like Figure 1-3 As shown, the synchronous retirement judgment condition of the battery pack in step five is as follows: when the ratio of the maximum to the minimum of the remaining operational charge-discharge cycles calculated by the system for all battery cells is lower than a set ratio, and the health status assessment value of each cell is lower than the retirement threshold, the synchronous retirement window is entered. When the condition is met for the first time, a retirement takeover report is automatically generated, which includes a tiered list of the remaining available capacity of each cell and a recommendation of tiered utilization scenarios. This judgment is performed every few charge-discharge cycles for users to refer to and select the retirement time, without forcing immediate termination of operation. With the ultimate optimization goal of simultaneous retirement of individual cells, a precise judgment mechanism is set up to periodically detect the aging synchronization of the battery pack. The specific implementation method is as follows: Judgment cycle: Synchronous decommissioning status detection is performed once every 30 charge-discharge cycles, balancing detection accuracy and system computing power consumption; Simultaneous retirement core determination formula: In the formula: The maximum remaining effective number of cycles for all individual cells in the group; The minimum remaining effective cycle count for all cells in the group is 1.15, which is the preset synchronization deviation ratio threshold, and 80% is the common SOH threshold for battery retirement in the battery industry. Retirement Report Generation: When the above dual judgment conditions are met for the first time, the battery pack is judged to enter the synchronous retirement window, and the system automatically generates a retirement takeover report. The report includes a graded list of the remaining available capacity of all cells in the group, a health status rating of each cell, and recommended content for tiered utilization scenarios (low-speed energy storage, backup power). Flexible operation mechanism: This judgment result is only for reference in operation and maintenance decision-making and does not force the system to be shut down and decommissioned immediately. Users can flexibly choose the timing of decommissioning according to the grid dispatching needs and operation and maintenance plans, taking into account both system operation stability and operation and maintenance flexibility.

[0031] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for adaptive equalization control of energy storage battery pack capacity, characterized in that: The method includes the following steps: Step 1: During the operation of the energy storage system, the cumulative throughput, temperature fluctuation records and terminal voltage behavior parameters of each battery cell are collected after each charge and discharge cycle, with the charge and discharge cycle as the time unit. The aging characteristic fingerprint of each cell is constructed, the remaining effective charge and discharge cycles of each battery cell are estimated to reach the retirement health state threshold, and the retirement time difference sorting vector of each cell is output. Step 2: Using the retirement time difference sorting vector output in Step 1 as the state space, and the adjustment amount of the output ratio coefficient of each battery cell in the next charge-discharge cycle as the action space, the lifetime balancing strategy model is trained using a multi-subject deep deterministic strategy gradient framework. After training, the strategy model runs online and outputs the output adjustment vector for the next stage based on the current state. Step 3: Use the output adjustment vector output in Step 2 as the macro trend command. In each charge and discharge cycle, use the model predictive control framework to perform rolling correction on the macro trend command. During the correction process, the safety threshold of the battery cell is used as the constraint condition, and the final output coefficient of each cell in this cycle is output. Step 4: When the final output coefficient causes energy deviation when each individual battery performs charging and discharging tasks, during the rest period after the end of the charging and discharging cycle, cross-cycle energy calculation is performed using a distributed floating energy storage unit cluster relay method to achieve local energy transfer. Step 5: Repeat steps 1 to 4, with the goal of minimizing the standard deviation of the retirement time difference of each battery cell as the long-term optimization objective. Dynamically adjust the output coefficient of each cell in each cycle so that the capacity decay trajectory of each battery cell tends to be synchronized throughout the entire life cycle.

2. The adaptive equalization control method for the capacity of an energy storage battery pack according to claim 1, characterized in that: In step one, the single-cell aging feature fingerprint is constructed by recording the cumulative throughput, temperature fluctuation, and terminal voltage behavior parameters at the end of each charge-discharge cycle. The parameters are aggregated to form a feature vector, and the differential aging factor is obtained by normalizing the ratio of the statistical rate of change of the feature vector to the standard rate of change within the sliding time window.

3. The adaptive equalization control method for the capacity of an energy storage battery pack according to claim 1, characterized in that: The method for extrapolating the remaining effective charge-discharge cycles based on aging feature fingerprints in step one is as follows: using the aging feature fingerprint sequence of a battery cell in the most recent few cycles as the original data, the predicted value of future cycles is estimated through cumulative generation and exponential trend extrapolation model, the predicted value is compared with the retirement health status threshold to obtain the remaining cycle number, the data is updated and the model parameters are re-estimated after each cycle, and the difference between each cell and the median of the remaining cycle number of all cells is calculated based on the median to form a sorting vector.

4. The adaptive equalization control method for the capacity of an energy storage battery pack according to claim 1, characterized in that: The adjustment amount of the output ratio coefficient in step two satisfies: The sum of the output ratio coefficients of all individual units remains 1 before and after the adjustment; When the health status of a single cell is below the first threshold, its adjustment amount is limited to a non-positive value; When the health status of a single entity is above the second threshold, its adjustment amount is allowed to be positive and the upper limit is relaxed.

5. The adaptive equalization control method for the capacity of an energy storage battery pack according to claim 1, characterized in that: In step two, the multi-agent deep deterministic policy gradient framework adopts a centralized training and decentralized execution architecture: during the training phase, each agent acts as an independent agent and shares a centralized commentator network that can access global information; during the execution phase, each agent outputs actions based only on local observation information; and the policy model is trained offline using historical data during the initial deployment phase.

6. The adaptive equalization control method for the capacity of an energy storage battery pack according to claim 1, characterized in that: The implementation of the model prediction control framework in step three is as follows: the prediction time domain width is set to several charge and discharge cycles. At the beginning of each decision cycle, a constrained quadratic programming problem is constructed based on the current state of charge and health. The optimization variable is the output coefficient fine-tuning amount for multiple future cycles. The objective function is to minimize the weighted sum of the macroscopic command tracking error and the output fluctuation penalty term. The constraints include the discharge depth safety threshold and the voltage safety boundary. The fine-tuning amount obtained in the first cycle is superimposed with the macroscopic command as the final output coefficient for this cycle. The solution is rolled over again in the next cycle.

7. The adaptive equalization control method for the capacity of an energy storage battery pack according to claim 1, characterized in that: In step four, the distributed floating energy storage unit satisfies the following: the capacity of each energy anchor point is a percentage of the capacity of a single cell in the battery pack; it uses supercapacitors, lithium-ion capacitors, or small lithium batteries; the operating voltage window matches the nominal voltage of the battery cell; each anchor point is electrically isolated; it is not grounded or directly connected to the main power circuit; and each anchor point is equipped with an independent management system to maintain its state of charge within a preset operating window.

8. The adaptive equalization control method for the capacity of an energy storage battery pack according to claim 1, characterized in that: The cluster relay method in step four includes: converting the final output coefficient into the planned charging and discharging energy of each individual unit before the start of each cycle; each individual unit performs charging and discharging according to the plan during the cycle; allowing temporary deviations without real-time intervention; calculating the total deviation between the actual and planned energy of the entire group during the rest period; and triggering relay equalization if the deviation exceeds the threshold. The relay balancing process includes: identifying sets of surplus and deficit cells, controlling nearby anchor points to draw energy from surplus cells up to the upper limit, disconnecting the anchor points and transmitting the energy to another anchor point via an isolation bus, and then releasing the energy to the deficit cells from that anchor point, repeating this process until the total deviation is less than the threshold or the rest period ends.

9. The adaptive equalization control method for the capacity of an energy storage battery pack according to claim 1, characterized in that: The collaborative operation mode of energy transfer during the rest period in step four is as follows: When it is necessary to process energy transfer between multiple groups of surplus and deficit cells at the same time, as long as the battery local clusters connected to each anchor point are electrically isolated and have no shared path, multiple anchor points are allowed to perform energy absorption or release in parallel. The multiple relay transmissions on the isolation bus adopt time-division multiplexing, allocate non-overlapping time windows for each source anchor point and target anchor point and insert protection intervals. The management system of each anchor point monitors the state of charge and temperature in real time. When the safety threshold is exceeded, participation in parallel exchange is automatically suspended and rejoined after recovery.

10. The adaptive equalization control method for the capacity of an energy storage battery pack according to claim 1, characterized in that: The synchronous retirement determination condition for the battery pack in step five is as follows: when the ratio of the maximum to the minimum number of remaining operational charge-discharge cycles calculated by the system for all battery cells is lower than a set ratio, and the health status assessment value of each cell is lower than the retirement threshold, the synchronous retirement window is determined. When the condition is met for the first time, a retirement takeover report is automatically generated, which includes a tiered list of the remaining available capacity of each cell and a recommendation of tiered utilization scenarios.

Images