An adaptive load distribution method and system for a parallel power supply system

Abstract

The application discloses a kind of adaptive load distribution method and system of parallel power supply system, it is related to power supply system field, including in the detection time window by the positive and negative electrode terminal of each battery unit injection asymmetric wideband alternating current current pulse sequence, voltage fluctuation response and actual injection current are synchronously collected to obtain impedance response sequence;Its input long short-term memory network model outputs electrolyte spatial concentration distribution sequence and extracts extreme value;When receiving discharge instruction, distribute load to target unit with concentration difference less than preset threshold;When receiving idle instruction, control target unit to output electric energy to abnormal unit, drive current flows through the internal bottom area of abnormal unit to excite local joule heat to generate heat convection from bottom to top.The application solves the problem that liquid battery is damaged and scrapped due to blind loading caused by implicit concentration stratification.

Description

Technical Field

[0001] This invention relates to the field of power supply systems, and more specifically, to an adaptive load distribution method and system for parallel power supply systems. Background Technology

[0002] With the development of microgrids, distributed energy storage power stations, and large-scale backup power systems, battery energy storage systems using liquid electrolytes still have significant application value in stationary energy storage scenarios. In energy storage arrays composed of multiple battery cells, battery clusters, or battery cabinets connected in parallel, maintaining the consistency of the operating state of each parallel battery cell and rationally allocating charging and discharging power according to the state of each cell are important factors affecting the safety, available capacity, and service life of the energy storage system.

[0003] For the balancing and control of battery packs, existing technologies have proposed a variety of solutions. For example, Chinese Patent Publication No. CN112054571A, entitled "A Method for SOC Consistency Balancing of Lithium Battery Energy Storage System," addresses the problem of low SOC estimation accuracy during the voltage plateau period of lithium batteries by proposing to construct a "virtual capacity" through an active voltage balancing system and correct the balancing results during non-voltage plateau periods, thereby improving the SOC consistency balancing accuracy of lithium battery energy storage systems.

[0004] Current technologies primarily focus on balancing the terminal voltage, state of charge, capacity differences, or estimated capacity of individual battery cells or battery clusters. The controlled objects are typically measurable or estimable electrical states external to the battery cell. However, in battery energy storage systems using liquid electrolytes, implicit inconsistencies may exist within the battery due to electrolyte fluid distribution. Taking flooded lead-acid batteries, lead-carbon batteries, or other energy storage units with flowing liquid electrolytes as examples, when the battery is in a state of long-term quiescence, float charging, shallow charging and discharging, or irregular low-rate cycling, the electrolyte may stratify due to gravity and insufficient diffusion, resulting in a relatively high electrolyte concentration at the bottom and a relatively low concentration at the top.

[0005] Electrolyte concentration stratification can cause differences in ionic conductivity, electrode reaction conditions, and concentration polarization states at different heights within a battery. These differences are often difficult to reflect directly through external terminal voltage, overall temperature, or conventional SOC estimations. In other words, even if a parallel battery cell exhibits normal terminal voltage and a high SOC externally, it may still have an issue with uneven electrolyte distribution throughout its internal structure.

[0006] When the grid or load side issues a high-power discharge dispatch command, if the parallel energy storage system still distributes the load according to the average allocation, capacity ratio allocation, or simply based on SOC allocation, it may cause battery cells that have already undergone electrolyte concentration stratification to bear a large sudden discharge current. Due to the uneven reaction conditions inside such battery cells, the sudden large current will further amplify the existing concentration stratification. Specifically, when the electrolyte concentration in the top region is low, the number of ions that can participate in charge transfer is insufficient, and the local conductivity decreases. At the moment of high-current discharge, the external load requires the battery to quickly output a large amount of charge, but the top region cannot provide sufficient ion migration and electrochemical reaction support in time, which easily leads to obvious concentration polarization and charge transfer polarization. At this time, in order to maintain the external current output, the local electrode potential will be forced to deviate from the normal reaction range, thereby increasing the probability of side reactions, such as gas evolution reaction or local overheating.

[0007] Meanwhile, the bottom region of the electrolyte, due to its higher concentration and conductivity, bears a greater current flow and reaction load, leading to a further imbalance in the current density distribution across the upper and lower regions of the electrode. The bottom region may experience over-reaction and localized heating, while the top region suffers from insufficient reaction due to inadequate ion supply. After prolonged or repeated exposure to this process, the utilization rate of active materials at different heights of the battery electrode will vary, with some areas experiencing excessive consumption and others chronically underutilized. This results in uneven sulfation distribution, softening or shedding of active materials, increased internal resistance, and decreased usable capacity. Therefore, for parallel battery cells that have already experienced electrolyte concentration stratification, a sudden large current does not act uniformly throughout the entire battery. Instead, it further exacerbates the existing concentration differences, causing uneven current distribution, uneven reaction rates, and uneven heat distribution. This superposition of electrochemical reactions and thermal effects accelerates battery cell capacity decay and increases the risk of premature failure.

[0008] Therefore, there is an urgent need for an adaptive load allocation method suitable for parallel battery energy storage systems using liquid electrolytes, so that the system can obtain state information characterizing the spatial concentration distribution of the electrolyte without damage during the detection phase, and avoid applying sudden high-power loads to battery cells with obvious concentration stratification during the discharge scheduling phase. Summary of the Invention

[0009] The technical problem to be solved by the present invention is to provide an adaptive load distribution method and system for a parallel power supply system, so as to solve the problems mentioned in the background art.

[0010] To achieve the above objectives, the present invention adopts the following technical solution:

[0011] An adaptive load sharing method for a parallel power supply system, applied to a battery energy storage system using liquid electrolyte, includes:

[0012] Within the detection time window, an asymmetric broadband AC current pulse sequence is injected through the positive and negative terminals of each of multiple parallel battery cells. Simultaneously, the corresponding voltage fluctuation response sequence and the actual injected current sequence are acquired at the positive and negative terminals. An impedance response sequence is obtained based on the voltage fluctuation response sequence and the actual injected current sequence. The impedance response sequence is then input into a long short-term memory network model, which outputs the electrolyte spatial concentration distribution sequence within each parallel battery cell. The maximum and minimum concentration values ​​in the electrolyte spatial concentration distribution sequence of each parallel battery cell are extracted.

[0013] Upon receiving a discharge scheduling command, the total load power carried by the discharge scheduling command is allocated to the target parallel battery cell whose difference between the maximum concentration value and the minimum concentration value is less than a preset uniformity threshold. Upon receiving an idle state command, the target parallel battery cell is controlled to output electrical energy to the abnormal parallel battery cell whose difference is greater than or equal to the preset uniformity threshold by adjusting the bidirectional converter. The driving current flows through the bottom region inside the abnormal parallel battery cell to generate bottom-up heat convection by stimulating local Joule heating in the bottom region.

[0014] Furthermore, the asymmetric broadband AC current pulse sequence is a bidirectional current pulse perturbation signal containing multiple components of different frequencies, wherein the amplitude of the forward current pulse is not equal to that of the reverse current pulse, and the net charge of the bidirectional current pulse perturbation signal within one detection period is less than a preset charge threshold; the step of obtaining the impedance response sequence includes:

[0015] The instantaneous current value and voltage fluctuation response sequence during the pulse injection are collected synchronously, and the actual injection current sequence is formed from the instantaneous current value according to the sampling time sequence.

[0016] The voltage fluctuation response sequence and the actual injected current sequence are transformed in the frequency domain to extract the voltage frequency domain complex value and the current frequency domain complex value under multiple frequency components.

[0017] Divide the voltage frequency domain complex value under the same frequency component by the corresponding current frequency domain complex value to calculate the impedance complex value under the corresponding frequency component.

[0018] The impedance response sequence is generated by arranging at least one of the impedance complex values, impedance magnitudes, impedance phase angles, and the real or imaginary parts of the impedance at each frequency component in frequency order.

[0019] Furthermore, the training process of the Long Short-Term Memory network model includes:

[0020] Obtain the sample impedance response sequence from historical operating data and the actual electrolyte spatial concentration distribution sequence corresponding to the sample impedance response sequence;

[0021] The sample impedance response sequence is used as the input feature to input the initial long short-term memory network model, and the output is the predicted distribution sequence;

[0022] The actual electrolyte spatial concentration distribution sequence is used as a supervision label, and the cross-entropy loss function is used to quantify the numerical deviation between the predicted distribution sequence and the supervision label.

[0023] The network parameters of the initial long short-term memory network model are iteratively updated using the gradient descent algorithm until the numerical deviation meets the preset convergence condition, thereby obtaining the trained long short-term memory network model.

[0024] Furthermore, the process of obtaining the actual electrolyte spatial concentration distribution sequence includes:

[0025] Concentration detection probes are deployed at different physical height positions inside a pre-set multi-segment test battery pack. The concentration detection probes include fiber optic refractive index concentration probes or miniature ultrasonic density probes.

[0026] The concentration probe synchronously collects the actual concentration values ​​of the electrolyte at different physical height locations;

[0027] All the actual electrolyte concentration values ​​at the same time point are spliced ​​together in descending order of physical height to generate the actual electrolyte spatial concentration distribution sequence as the supervision label.

[0028] Furthermore, controlling the target parallel battery cell to output electrical energy to the abnormal parallel battery cell by adjusting the bidirectional converter includes:

[0029] Real-time acquisition of the remaining power percentage of the abnormal parallel battery units;

[0030] When the remaining power ratio is less than a preset safety upper limit threshold, the bidirectional converter is controlled to continuously input charging current from the target parallel battery unit to the abnormal parallel battery unit.

[0031] When the remaining power ratio reaches the preset safety upper limit threshold, the pulse width modulation duty cycle of the internal switching device of the bidirectional converter is adjusted to control the output voltage value of the abnormal parallel battery unit to be greater than the receiving voltage value of the target parallel battery unit to reverse the current flow direction. When the remaining power ratio drops to the preset safety lower limit threshold, the current flow direction is reversed again, and the local Joule heating is maintained by alternating charging and discharging.

[0032] Furthermore, during the time period of the alternating charge and discharge, the alternating charge and discharge is paused at a preset time interval, and the asymmetric broadband AC current pulse sequence is re-injected into the abnormal parallel battery cell through the positive and negative terminals to obtain an updated impedance response sequence.

[0033] The updated impedance response sequence is input into the long short-term memory network model, which outputs the current electrolyte spatial concentration distribution sequence and extracts the current maximum concentration value and the current minimum concentration value.

[0034] When the difference between the current maximum concentration value and the current minimum concentration value is less than the preset uniformity threshold, a convection termination command is generated to stop the alternating charging and discharging.

[0035] Furthermore, while performing the alternating charge and discharge, a targeted thermal control mechanism for the abnormal parallel battery cells is activated, including:

[0036] The thermal management system corresponding to the abnormal parallel battery unit is controlled to stop cooling the bottom of the abnormal parallel battery unit, and the thermal management system is controlled to start cooling the top of the abnormal parallel battery unit.

[0037] During the operation of the directional thermal control mechanism, the bottom casing temperature of the abnormal parallel battery unit is collected in real time;

[0038] When the temperature of the bottom outer casing reaches the preset thermal runaway warning threshold, the thermal management system is controlled to resume cooling of the bottom of the abnormal parallel battery cell and a power zeroing command is sent to the bidirectional converter to cut off the charging and discharging current.

[0039] Furthermore, before inputting the impedance response sequence into the long short-term memory network model, a bubble interference screening operation is performed for each parallel battery cell, including:

[0040] A DC bias voltage signal is injected into the corresponding parallel battery cell through the positive and negative terminals and maintained for a preset time period.

[0041] At the termination time node when the DC bias voltage signal is removed, a second asymmetric broadband AC current pulse sequence is injected into the corresponding parallel battery cell through the positive and negative terminals to obtain the second impedance response sequence.

[0042] The first impedance response sequence and the second impedance response sequence are input into a trained convolutional neural network binary classification model to output the discrimination result.

[0043] Furthermore, the initially acquired impedance response sequence and the second impedance response sequence are input into a trained convolutional neural network binary classification model, and the output discrimination result includes:

[0044] The impedance response sequence obtained initially is concatenated with the impedance response sequence obtained second time to form a dual-channel input sequence;

[0045] The dual-channel input sequence is input into the convolutional neural network binary classification model. The convolutional neural network binary classification model extracts the morphological shift features between the first impedance response sequence and the second impedance response sequence, and outputs a classification label characterizing whether there is bubble attachment interference.

[0046] When the classification label indicates interference, the identification result is confirmed to be bubble adhesion interference, and a negative depolarization current pulse is injected into the corresponding parallel battery unit through the positive and negative terminals.

[0047] When the classification label indicates that there is no interference, the identification result is confirmed as the true concentration gradient, and the operation of inputting the initially acquired impedance response sequence into the long short-term memory network model is triggered.

[0048] The present invention also discloses an adaptive load sharing system for implementing the above method in a parallel power supply system, comprising:

[0049] An AC pulse injection and acquisition module is used to inject an asymmetric broadband AC current pulse sequence through the positive and negative terminals of multiple parallel battery cells within a detection time window, and simultaneously acquire the corresponding voltage fluctuation response sequence and actual injected current sequence at the positive and negative terminals, and obtain the impedance response sequence based on the voltage fluctuation response sequence and the actual injected current sequence.

[0050] The state prediction and analysis module is used to input the impedance response sequence into the long short-term memory network model, and the long short-term memory network model outputs the electrolyte spatial concentration distribution sequence inside each parallel battery cell; and extracts the maximum and minimum concentration values ​​in the electrolyte spatial concentration distribution sequence of each parallel battery cell.

[0051] The discharge scheduling and allocation module is used to receive a discharge scheduling instruction and allocate the total load power carried by the discharge scheduling instruction to the target parallel battery cell whose difference between the maximum concentration value and the minimum concentration value is less than a preset uniformity threshold.

[0052] The circulating current charging and discharging control module is used to receive idle state commands and control the target parallel battery cell to output electrical energy to the abnormal parallel battery cell whose difference is greater than or equal to the preset uniformity threshold by adjusting the bidirectional converter. The driving current flows through the bottom region inside the abnormal parallel battery cell to generate local Joule heating in the bottom region to produce heat convection from bottom to top.

[0053] The advantages of this invention over the prior art are:

[0054] Compared to traditional battery management systems, this invention expands the management scope of the energy storage system from external macroscopic parameters such as terminal voltage, remaining capacity ratio, and temperature to the electrolyte concentration distribution within the battery, thereby enabling state identification, load avoidance, and proactive maintenance of parallel battery cells.

[0055] First, this invention enables non-destructive testing of the electrolyte concentration distribution within a battery without disassembling the battery or damaging its sealed structure. Specifically, this invention injects an asymmetric broadband AC current pulse sequence into the positive and negative terminals of the battery cell, and simultaneously acquires the voltage fluctuation response sequence and the actual injected current sequence, obtaining the impedance response sequence based on both.

[0056] This invention can use impedance response sequences to determine whether the electrolyte is in a uniform or stratified state. The principle is that different electrolyte concentration distribution states will cause the battery to generate different impedance feedbacks to AC current disturbances.

[0057] When the electrolyte concentration distribution is relatively uniform, the distribution of conductive ions in the vertical direction inside the battery is relatively consistent, and the charge transfer and ion diffusion conditions in different height regions of the porous electrode are relatively similar. Therefore, after injecting an asymmetric broadband AC current pulse sequence into the battery, the response of each region to AC disturbances is relatively small, and the impedance response sequence measured at the port usually exhibits relatively stable and regular frequency domain characteristics.

[0058] When the electrolyte undergoes gravitational stratification, the ion concentration is higher in the bottom region of the plates and lower in the top region, forming distributed conductive regions with different impedance characteristics inside the battery. At this point, alternating current disturbances of different frequencies will exhibit different response characteristics.

[0059] In the high-frequency band, the AC signal changes rapidly, making it difficult for ions to complete long-distance diffusion or fully participate in electrochemical reactions within a single cycle. Therefore, the high-frequency impedance mainly reflects the ohmic conductivity of the electrolyte. When the electrolyte concentration is high and the ohmic impedance is low in the bottom region, the high-frequency current is more likely to form a current path through the low-impedance region at the bottom, causing the high-frequency equivalent impedance measured at the port to appear lower or significantly deviated.

[0060] In the low-frequency range, the AC signal changes more slowly, allowing ion migration, concentration diffusion, and charge transfer processes to participate more fully in the response. In the top low-concentration region, due to the insufficient number of ions available for reaction and migration, low-frequency perturbations are more prone to ion replenishment hysteresis and enhanced concentration polarization. This phenomenon manifests in port impedance results as low-frequency phase angle shifts, increased diffusion impedance, and a curved or hysteresis-enhanced recovery tail of the impedance curve.

[0061] Therefore, when spatial concentration stratification occurs in the electrolyte, the port impedance response sequence simultaneously includes high-frequency ohmic impedance shifts and low-frequency phase / diffusion characteristic changes. This combined characteristic reflects the difference between the low-concentration, high-polarization region at the top of the battery and the high-concentration, low-impedance region at the bottom. This invention uses a long short-term memory network model to analyze the combination of features in this impedance response sequence, thereby obtaining parameters to characterize the spatial concentration distribution and degree of stratification of the electrolyte.

[0062] The impedance response sequence can be understood as the electrical feedback result of the electrolyte inside the battery to external AC disturbances. Although the battery port can only collect the overall voltage and current, the overall response is actually formed by the superposition of the responses of different regions inside the battery. Therefore, as long as there are differences in conductivity, ion supply capacity, or polarization in different regions inside the battery, these differences will be reflected in the impedance response sequence measured at the port.

[0063] This invention further inputs the impedance response sequence into a long short-term memory network model, which extracts the frequency domain and temporal features from the impedance response and outputs the electrolyte spatial concentration distribution sequence within each parallel battery cell. By extracting the maximum and minimum concentration values ​​from this sequence and calculating the difference between them, a quantitative parameter characterizing the uniformity of electrolyte concentration can be obtained, thereby transforming the internal concentration stratification state of the battery into a control basis that can be used for load distribution.

[0064] Secondly, this invention can allocate the discharge load based on the spatial concentration distribution of the electrolyte. Upon receiving a discharge scheduling command, this invention allocates the total load power to target parallel battery cells where the difference between the maximum and minimum concentration values ​​is less than a preset uniformity threshold. This allows battery cells with more uniform concentration distribution to take on the discharge task first, while abnormal parallel battery cells with obvious concentration stratification are prevented from directly experiencing sudden high-power discharge impacts. Therefore, the risk of localized polarization, gas evolution, temperature rise, uneven aging, or rapid capacity decay caused by insufficient local ion supply in abnormal battery cells can be reduced, improving the safety and reliability of the parallel energy storage system under high-power scheduling conditions.

[0065] Furthermore, this invention enables proactive maintenance of abnormal parallel battery cells exhibiting concentration stratification during grid idle periods. Upon receiving an idle state command, this invention adjusts the bidirectional converter to control the target parallel battery cell to output electrical energy to the abnormal parallel battery cell. Because the abnormal parallel battery cell contains a low-concentration, high-resistance region at the top and a high-concentration, low-resistance region at the bottom, current flows more easily through the low-resistance region at the bottom, generating localized Joule heating. This localized Joule heating raises the temperature and decreases the density of the electrolyte at the bottom, causing it to flow upwards. Simultaneously, the relatively cooler electrolyte at the top replenishes the flow downwards, creating a bottom-up natural thermal convection that promotes mixing of the high-concentration electrolyte at the bottom with the low-concentration electrolyte at the top, thereby alleviating electrolyte concentration stratification.

[0066] Furthermore, this invention monitors the remaining charge percentage of the abnormal parallel battery cells and reverses the current flow direction when the remaining charge percentage reaches a preset safety upper or lower threshold, enabling alternating charging and discharging between the abnormal parallel battery cells and the target parallel battery cells. This method maintains localized Joule heating and thermal convection driving force in the bottom region of the abnormal parallel battery cells while preventing overcharging, over-discharging, or remaining charge percentage deviation caused by prolonged unidirectional charging or discharging of the abnormal parallel battery cells.

[0067] Furthermore, this invention pauses alternating charge and discharge at preset time intervals during the alternating charge and discharge process, and re-detects the impedance of abnormal parallel battery cells. The system inputs the updated impedance response sequence back into the long short-term memory network model, outputs the current electrolyte spatial concentration distribution sequence, and determines the degree of stratification mitigation based on the difference between the current maximum concentration value and the current minimum concentration value. When this difference is less than a preset uniformity threshold, the system generates a convection termination command and stops alternating charge and discharge.

[0068] Furthermore, to improve the efficiency and safety of the thermal convection maintenance process, this invention incorporates a directional thermal control mechanism. While performing alternating charge and discharge, the system controls the thermal management system corresponding to the abnormal parallel battery cell to stop or reduce cooling of the bottom region, and to activate or enhance cooling of the top region. This creates a spatial temperature boundary condition within the abnormal parallel battery cell where the bottom temperature is higher than the top temperature, thereby enhancing the driving force for the upward flow of the heated electrolyte at the bottom and promoting natural thermal convection. Simultaneously, the system collects the bottom casing temperature of the abnormal parallel battery cell in real time. When this temperature reaches a preset thermal runaway warning threshold, the system controls the thermal management system to resume bottom cooling and sends a power zeroing command to the bidirectional converter to cut off the charging and discharging current, reducing the risk of localized overheating.

[0069] Furthermore, this invention performs bubble interference screening before inputting the impedance response sequence into the long short-term memory network model to reduce misjudgments caused by bubble adhesion, local gas evolution, or interface polarization. Specifically, this invention injects a DC bias voltage signal into the corresponding parallel battery cell and maintains it for a preset time period, followed by injecting a second asymmetric broadband AC current pulse sequence to obtain a second impedance response sequence. If the impedance anomaly originates from bubble adhesion, the bubbles may undergo interface state changes under the action of the DC bias electric field, causing a significant morphological shift in the second impedance response sequence compared to the first impedance response sequence. If the impedance anomaly originates from actual concentration stratification, the short-term DC bias voltage is unlikely to significantly change the overall spatial concentration distribution of the electrolyte, and the two impedance response sequences usually maintain a high degree of consistency.

[0070] This invention concatenates the initially acquired impedance response sequence with the second impedance response sequence into a dual-channel input sequence, which is then input into a trained convolutional neural network binary classification model. This model outputs a classification label indicating the presence of bubble adhesion interference. When the classification label indicates the presence of bubble adhesion interference, the system injects a negative depolarization current pulse into the corresponding parallel battery cell to reduce the impact of bubble adhesion or interface polarization on the impedance detection results. When the classification label indicates the absence of bubble adhesion interference, the system confirms that the impedance response corresponds to the true concentration gradient and triggers subsequent recognition by a long short-term memory network model. Therefore, this invention improves the accuracy and reliability of electrolyte concentration distribution identification results.

[0071] In summary, this invention utilizes technologies such as asymmetric broadband AC current pulse detection, impedance response sequence analysis, long short-term memory network model prediction, discharge load allocation based on concentration uniformity, alternating charge and discharge during idle periods to induce thermal convection, directional thermal control safety protection, and bubble interference identification. These technologies enable parallel liquid electrolyte battery energy storage systems to identify internal electrolyte concentration stratification without damaging the battery structure. Based on this identification result, the system can perform load allocation and proactive maintenance, thereby improving the operational safety, load allocation rationality, and long-term service reliability of the energy storage system. Attached Figure Description

[0072] Figure 1 This is a schematic diagram of the overall process of the adaptive load distribution method for the parallel power supply system of the present invention;

[0073] Figure 2 This is a schematic diagram of the system composition of the adaptive load distribution system of the parallel power supply system of the present invention;

[0074] Figure 3 This is a schematic diagram of the impedance detection principle of the present invention, which injects asymmetric broadband AC current pulses into the positive and negative terminals and obtains the impedance response sequence.

[0075] Figure 4This is a schematic diagram of the training process of the long short-term memory network model of the present invention and the construction of labels for the actual electrolyte spatial concentration distribution sequence;

[0076] Figure 5 This is a schematic diagram illustrating the discharge scheduling and allocation based on the uniformity of electrolyte concentration according to the present invention.

[0077] Figure 6 This is a schematic diagram of the present invention for alternating charge and discharge control based on the remaining power ratio of abnormal parallel battery cells;

[0078] Figure 7 This is a schematic diagram of the process of performing periodic retesting and generating convection termination commands during the alternating charge and discharge maintenance of the present invention.

[0079] Figure 8 This is a schematic diagram illustrating the directional thermal control mechanism and bubble interference identification operation performed during the maintenance process of this invention. Detailed Implementation

[0080] The invention will now be described with reference to the accompanying drawings. For ease of understanding, the following embodiments use multiple battery cabinets employing liquid electrolytes connected in parallel to form an energy storage array as an example. In practical applications, the parallel battery units can also be single cells, battery modules, battery clusters, or battery cabinets, as long as they contain a liquid electrolyte that can flow or diffuse slowly, the solution of this invention can be adopted. The concentration values ​​in the electrolyte spatial concentration distribution sequence can be expressed as mass concentration, molar concentration, or concentration equivalent to that obtained by converting electrolyte density. The same dimension is used in the training, inference, and threshold judgment processes to avoid judgment biases between different concentration characterization methods.

[0081] The overall execution process of this invention is as follows: Figure 1 As shown, the system diagram is as follows: Figure 2 As shown, the parallel power supply system includes multiple parallel battery cells, bidirectional converters connected to each parallel battery cell, sampling circuits for acquiring terminal voltage and current, pulse injection circuits for injecting AC pulse disturbance signals, a controller for performing data processing and model inference, and a thermal management system for regulating the local cooling state of the battery cells. To enable controlled energy transfer from the target parallel battery cell to the abnormal parallel battery cell, each parallel battery cell can be connected to a common DC bus via an independent bidirectional DC-DC converter branch, or via a bidirectional converter matrix with gating function. This allows the controller to actively control the current direction and magnitude between different parallel battery cells, preventing uncontrollable circulating currents caused by differences in terminal voltage among multiple battery cells. The controller can be implemented using a battery management system, energy management system, industrial controller, or embedded processor within the energy storage system. Model inference can be performed in the local controller or in an edge computing device connected to the energy storage station.

[0082] In one embodiment, the system selects a detection time window before entering discharge scheduling. The detection time window can be set during the state detection phase after the energy storage system is in a static, lightly loaded, or float-charged state, or before the arrival of the scheduling command, and its length can range from 5 seconds to 300 seconds. This time window should not be too short, as this will lead to insufficient low-frequency impedance characteristics; nor should it be too long, as this will affect the energy storage system's response speed to scheduling commands. For large-scale stationary energy storage stations, the detection time window can be 30 to 180 seconds; for backup power systems requiring rapid response, it can be 5 to 30 seconds.

[0083] Within the detection time window, the controller injects an asymmetric broadband AC current pulse sequence through the positive and negative terminals of multiple parallel battery cells, and simultaneously acquires the voltage fluctuation response sequence and the actual injected current sequence at the same positive and negative terminals. Figure 3 As shown in the diagram, the asymmetric broadband AC current pulse sequence refers to a bidirectional disturbance signal that simultaneously contains positive and negative current pulses and multiple components of different frequencies. The amplitudes of the positive and negative current pulses are not exactly equal. For example, the amplitude of the positive pulse can be 0.02C to 0.20C of the current corresponding to the battery's rated capacity, while the amplitude of the negative pulse can be 0.01C to 0.18C. The ratio of their amplitudes can be controlled between 1.1 and 5.

[0084] It should be noted that this method, which involves injecting current pulses through the positive and negative terminals and acquiring voltage fluctuation responses at these terminals, is essentially a port-type impedance measurement method. In other words, the battery cell is considered an electrochemical impedance spectroscopy (EIS) object with positive and negative terminals. An external detection circuit applies a small current disturbance to these terminals and simultaneously detects the voltage response generated under the disturbance. The impedance information is then obtained based on the relationship between the voltage response and the actual injected current. This method is feasible for both electrochemical impedance spectroscopy and online internal resistance detection, and there is no issue of detection being impossible due to both the injection and sampling positions corresponding to positive and negative terminals. Furthermore, "through the positive and negative terminals" is a general description of the electrical connection location and does not imply that all embodiments can only use two wires for the actual physical wiring. In some embodiments, the current injection circuit and voltage sampling circuit can be arranged separately at the same positive and negative terminal location, for example, using a four-wire Kelvin connection. One set of larger cross-sectional area wires is used to inject an asymmetric broadband AC current pulse sequence into the parallel battery cells, while another set of high input impedance sampling wires is used to simultaneously acquire the voltage fluctuation response between the positive and negative terminals. Because the voltage sampling circuit has a high input impedance, the current flowing through the sampling conductor is very small, and the voltage drop across the sampling conductor can be ignored. Simultaneously, the additional voltage drop caused by the current injected into the conductor and its contact resistance is not directly added to the voltage sampling result. Therefore, using a four-wire Kelvin connection can reduce the influence of conductor resistance, terminal contact resistance, and injected current voltage drop on the impedance measurement results, improving the accuracy of the correspondence between the voltage fluctuation response sequence and the actual injected current sequence. Even in the simplified embodiment using a two-wire connection, the influence of wiring impedance can be reduced through no-load calibration, conductor impedance compensation, and synchronous sampling correction. In the preferred embodiment, a four-wire Kelvin connection is used to complete the port-type impedance measurement.

[0085] The reason for using asymmetric pulses is that perfectly symmetrical AC perturbations primarily reflect the linear impedance characteristics of the battery within a small signal range. However, the stratification of liquid electrolyte concentration causes different ion migration capabilities, local conductivity, and polarization states in different height regions within the battery. These differences are more easily manifested as changes in impedance response morphology under slight asymmetric perturbations. The wideband setting is used to simultaneously observe high-frequency ohmic conductivity characteristics and low-frequency diffusion and polarization characteristics. Frequency components can cover 0.01Hz to 10kHz, and the number of frequency points can range from 5 to 80. For lead-acid or lead-carbon batteries, the high-frequency band can focus on covering 100Hz to 10kHz, and the low-frequency band can focus on covering 0.01Hz to 10Hz.

[0086] To avoid significantly altering the remaining charge percentage of the battery during the detection process, the net charge of the bidirectional current pulse disturbance signal within one detection cycle is controlled within a preset charge threshold. The preset charge threshold can be 0.001% to 0.5% of the rated capacity of the corresponding parallel battery cell, preferably 0.005% to 0.1%. A smaller net charge results in less disturbance to the battery's state of charge; however, if the net charge is too small, the voltage fluctuation response signal may be overwhelmed by sampling noise. Therefore, the actual value can be calibrated based on the battery capacity, sampling accuracy, and the level of electromagnetic interference at the site.

[0087] During pulse injection, the sampling circuit synchronously acquires instantaneous current and voltage fluctuation response values ​​according to a unified sampling clock. The sampling frequency can be 5 to 50 times the highest injection frequency; for example, if the highest frequency is 10 kHz, the sampling frequency can be 50 kHz to 500 kHz. The instantaneous current values ​​are arranged in the order of sampling time to form the actual injected current sequence; the voltage fluctuation response values ​​are arranged in the order of sampling time to form the voltage fluctuation response sequence. Synchronous acquisition reduces phase error and avoids distortion in impedance phase angle calculation due to time offset between voltage and current sampling.

[0088] Subsequently, the controller performs frequency domain transformation on the voltage fluctuation response sequence and the actual injected current sequence. Frequency domain transformation can be achieved using Fast Fourier Transform, Short-Time Fourier Transform, Discrete Fourier Transform, or sweep frequency phase-locked loop detection methods. If the pulse sequence contains multiple discrete frequency components, the voltage and current frequency domain complex values ​​can be extracted at the corresponding frequency points. The controller divides the voltage frequency domain complex value at the same frequency component by the corresponding current frequency domain complex value to obtain the impedance complex value at that frequency component. To improve noise immunity, mean filtering or median filtering can also be applied to the impedance complex values ​​of multiple adjacent detection cycles.

[0089] After obtaining the complex impedance values ​​for each frequency component, the controller generates an impedance response sequence in a fixed order from low to high or high to low frequency. The impedance response sequence can contain only the complex impedance values, or it can contain one or more of the following: impedance magnitude, impedance phase angle, real part of impedance, and imaginary part of impedance. In practical engineering, it is preferable to use the real part of impedance, the imaginary part of impedance, the impedance magnitude, and the impedance phase angle as a common input model for multiple channels. This is because the real part of impedance is more sensitive to changes in ohmic resistance, while the imaginary part and phase angle are more sensitive to diffusion polarization and interface polarization. Multi-channel input helps improve the accuracy of concentration distribution identification.

[0090] After the impedance response sequence is generated, the controller inputs it into the trained Long Short-Term Memory (LSTM) network model. The LTM model is used to infer the electrolyte spatial concentration distribution along the height direction inside the parallel battery cells based on the impedance response features arranged in frequency order. Here, the LTM model is not only used to process time series, but also treats the impedance features at different frequency points as a feature sequence with a fixed arrangement order, extracting the correlation between high-frequency ohmic features, mid-frequency charge transfer features, and low-frequency diffusion polarization features through memory units. The electrolyte spatial concentration distribution sequence can be understood as the concentration estimates corresponding to multiple height positions inside the battery from top to bottom or bottom to top. To keep the model output consistent with the training labels, this embodiment preferably uses a physical height order from high to low, for example, the first position in the sequence corresponds to the top region of the battery, and the last position corresponds to the bottom region of the battery.

[0091] In one embodiment, the Long Short-Term Memory (LSTM) network model includes an input layer, one to three LSM layers, a fully connected layer, and an output layer. The input layer receives a sequence of impedance features arranged in frequency order, with each frequency point corresponding to two to six input features. The number of hidden units in each LSM layer can be 32 to 256, and a random deactivation ratio of 0 to 0.5 can be set between layers to reduce the risk of overfitting. The fully connected layer maps the hidden states output by the LSM layer to concentration prediction results at multiple height locations. The number of height locations can be consistent with the number of concentration probes deployed during the training phase, for example, 4, 6, 8, or 12.

[0092] Because the cross-entropy loss function is used during training, this embodiment first discretizes the actual concentration value into multiple concentration levels, and then the model outputs the probability distribution of the corresponding concentration level for each height position. The number of concentration levels can be from 10 to 200, and the level intervals can be evenly divided according to the actual concentration measurement range of the electrolyte, or denser levels can be set in areas where concentration stratification is likely to occur. During model inference, the controller can choose the concentration level with the highest probability as the output result for that height position, or it can perform a weighted calculation based on the probabilities of each concentration level to obtain a continuous form of concentration estimate, thereby forming a spatial concentration distribution sequence of the electrolyte. This allows the cross-entropy loss function to match the label format, avoiding the ambiguity of the training objective caused by directly using continuous concentration values ​​as cross-entropy labels.

[0093] The training process of a Long Short-Term Memory (LSTM) network model is as follows: Figure 4As shown. During the training phase, the sample impedance response sequences from historical operating data and the corresponding actual electrolyte spatial concentration distribution sequences are first acquired. The sample impedance response sequences can be collected from experimental batteries, test battery packs, or calibrated battery packs in operation. To enable the model to learn different concentration stratification states, samples with different resting times, charge / discharge rates, ambient temperatures, and cycling conditions can be set during the testing process to induce homogeneous, slightly stratified, and heavily stratified states of the electrolyte.

[0094] The actual spatial concentration distribution sequence of the electrolyte was obtained through a pre-designed multi-segment test battery pack. Concentration probes were deployed at different physical heights within this test battery pack. These probes could be either fiber optic refractive index concentration probes or miniature ultrasonic density probes. Fiber optic refractive index concentration probes utilize the characteristic that the electrolyte's refractive index changes with concentration to obtain concentration values, making them suitable for experimental calibration requiring high spatial resolution. Miniature ultrasonic density probes utilize the relationship between ultrasonic propagation speed or echo characteristics and electrolyte density to obtain concentration values, making them suitable for scenarios where changes in electrolyte density are more sensitive.

[0095] Multiple concentration probes simultaneously collect actual electrolyte concentration values ​​at different heights. The controller stitches together all actual concentration values ​​at the same time point in descending order of physical height to generate an actual electrolyte spatial concentration distribution sequence, which is then converted into concentration level labels. Since real-world operating batteries are typically not suitable for long-term probe implantation, this multi-segment test battery pack is primarily used for model training, calibration, and validation. After training, the actual parallel energy storage system only requires port voltage and current responses to infer the internal concentration distribution.

[0096] During training, the controller or training server inputs the sample impedance response sequence as input features into the initial Long Short-Term Memory (LSTM) network model to obtain the predicted distribution sequence. The actual electrolyte spatial concentration distribution sequence, after concentration grading, serves as the supervision label. The cross-entropy loss function is used to quantify the numerical deviation between the predicted distribution sequence and the supervision label. Subsequently, the network parameters of the initial LSM network model are iteratively updated using a gradient descent algorithm. The gradient descent algorithm can be stochastic gradient descent, momentum gradient descent, RMSProp, or Adam optimization algorithm. The learning rate can be 0.0001 to 0.01, the number of samples per batch can be 16 to 256, and the number of iterations can be 50 to 1000. The preset convergence condition can be that the validation set loss decreases by less than 0.1% to 1% within 5 to 30 consecutive iterations, or that the validation set concentration grading prediction accuracy and concentration prediction error meet preset requirements. After the convergence condition is met, the trained LSM network model is obtained.

[0097] After the model inference is completed, the controller extracts the maximum and minimum concentration values ​​from the electrolyte spatial concentration distribution sequence for each parallel battery cell and calculates the difference between them. This difference characterizes the uniformity of electrolyte concentration. The smaller the difference, the closer the concentrations in the upper and lower regions, and the more uniform the ionic conductivity and reaction conditions inside the battery; the larger the difference, the more obvious the electrolyte concentration stratification. The preset uniformity threshold can be determined according to the electrolyte type and battery model, for example, it can be 0.5% to 10% of the rated electrolyte concentration. If the concentration value is expressed in terms of density equivalent, the preset uniformity threshold can be 0.005 g / cm³ to 0.08 g / cm³; if the concentration value is expressed in terms of molar concentration, the preset uniformity threshold can be calculated and determined according to the concentration stratification experimental results of the same battery model. For flooded lead-acid batteries, it is preferable to set this threshold based on the electrolyte density difference; for batteries using other liquid electrolytes, the concentration difference can be calculated based on the ion concentration difference or conductivity.

[0098] When the system receives a discharge scheduling command, the controller reads the total load power carried in the command and allocates power according to the uniformity of the battery concentration in each parallel battery cell, such as... Figure 5 As shown, parallel battery cells whose difference between the maximum and minimum concentration values ​​is less than a preset uniformity threshold are identified as target parallel battery cells; parallel battery cells whose difference is greater than or equal to the preset uniformity threshold are identified as abnormal parallel battery cells. The controller prioritizes allocating total load power to target parallel battery cells, allowing battery cells with more uniform internal concentration distribution to undertake the discharge task.

[0099] In specific allocation, the total load power can be evenly distributed to all target parallel battery cells, or a weighted allocation can be performed based on the remaining charge percentage, rated capacity, current temperature, and allowable discharge rate of the target parallel battery cells. This weighted allocation does not change the core logic of the invention, namely, abnormal parallel battery cells will not directly bear the sudden high-power discharge load. If the number of target parallel battery cells is insufficient to bear the total load power, the controller can return a power limiting suggestion to the energy management system, or reduce the total output power according to the safe discharge rate. If all parallel battery cells are judged to be abnormal, the controller can prohibit high-power discharge and generate maintenance instructions to avoid all battery cells with concentration stratification being subjected to high-rate impacts simultaneously.

[0100] When the system receives an idle state command, it indicates that the energy storage system is in a state where there is no external high-power discharge task or internal maintenance can be performed. At this time, the controller controls the target parallel battery cell to output electrical energy to the abnormal parallel battery cell by adjusting the bidirectional converter. This energy transfer can be completed via a common DC bus, an isolated bidirectional DC-DC converter, or a multi-port bidirectional converter. Since the bottom region inside the abnormal parallel battery cell usually has a higher electrolyte concentration and a lower equivalent impedance, the current entering the abnormal parallel battery cell does not only pass through the bottom region, but the bottom region is more likely to bear a larger proportion of the current flow and reaction load compared to the lower concentration region at the top. When the current passes through the relatively low impedance region at the bottom, it generates local Joule heating, which raises the temperature and decreases the density of the electrolyte at the bottom, causing it to flow upward. The lower temperature and lower concentration electrolyte at the top replenishes downward, forming a bottom-up natural thermal convection, thereby promoting the mixing of the upper and lower electrolytes and alleviating concentration stratification. In some alternative embodiments, controlled energy transfer can also be performed between multiple abnormal parallel battery cells to reduce the cycling losses caused by the target parallel battery cell participating in the maintenance process and to improve the convection efficiency between multiple abnormal parallel battery cells. However, this requires higher requirements for control stability, thermal safety, and protection of abnormal cells. In this embodiment of the invention, the target parallel battery cell outputs electrical energy to the abnormal parallel battery cell. Its advantage is that the electrolyte concentration distribution inside the target parallel battery cell is more uniform, and the terminal voltage recovery process, equivalent internal resistance, and polarization response are relatively stable, which can provide a more stable voltage support and current regulation reference for the bidirectional converter.

[0101] To prevent overcharging or over-discharging of abnormally connected parallel battery cells during maintenance, the controller collects the remaining charge percentage of the abnormally connected parallel battery cells in real time, such as... Figure 6 As shown. The remaining charge percentage can be obtained by the battery management system based on open-circuit voltage, current integral, temperature correction, and capacity calibration results. The preset safety upper limit threshold can be 70% to 95%, preferably 80% to 90%; the preset safety lower limit threshold can be 20% to 60%, preferably 30% to 50%. The safety upper limit threshold needs to be higher than the safety lower limit threshold, and the interval between the two can be 10% to 50% to avoid frequent switching of current direction.

[0102] When the remaining charge percentage of the abnormally connected parallel battery cell is less than a preset safety upper limit threshold, the controller controls the bidirectional converter to continuously input charging current from the target parallel battery cell to the abnormally connected parallel battery cell. This charging current can be set from 0.02C to 0.3C according to the rated capacity of the abnormally connected parallel battery cell. If the current is too small, the bottom Joule heating will be insufficient, and the driving force for thermal convection will be weak; if the current is too large, it may lead to excessively rapid local temperature rise or the risk of side reactions. Therefore, the actual value can be determined based on the bottom temperature, concentration difference, and the allowable rate of the battery manufacturer.

[0103] When the remaining charge percentage of the abnormal parallel battery cell reaches a preset safety upper limit threshold, the controller adjusts the pulse width modulation duty cycle of the switching devices inside the bidirectional converter. This ensures that the output voltage of the abnormal parallel battery cell, calculated through the converter to the circulating current control loop, is greater than the receiving voltage of the target parallel battery cell, thus reversing the current flow direction. The output and receiving voltages here are not simply forced increases in the battery's terminal voltage, but rather controlled energy flow is achieved through the buck-boost regulation of the bidirectional converter, preventing overvoltage at the battery terminals from exceeding permissible limits. At this time, the abnormal parallel battery cell discharges to the target parallel battery cell, while the system maintains a controlled internal circulating current channel, allowing localized Joule heating to continue in the bottom region of the abnormal parallel battery cell. When the remaining charge percentage of the abnormal parallel battery cell drops to a preset safety lower limit threshold, the controller reverses the current flow direction again, causing the target parallel battery cell to recharge the abnormal parallel battery cell. This alternating charge-discharge method maintains localized Joule heating and thermal convection driving force at the bottom, while preventing the abnormal parallel battery cell from prolonged unidirectional charging or discharging. It should be noted that the reversal of current direction during alternating charge and discharge is not used to change the spatial current flow position within the abnormally parallel battery cell, but rather to control the remaining charge ratio of the abnormally parallel battery cell during maintenance, avoiding prolonged unidirectional charging or discharging. After electrolyte concentration stratification occurs in the abnormally parallel battery cell, the bottom region has a higher electrolyte concentration, higher ionic conductivity, and lower equivalent impedance, while the top region has a lower electrolyte concentration and more pronounced polarization. Therefore, regardless of whether the abnormally parallel battery cell is in the charging or discharging direction, the maintenance current entering it tends to cause the relatively low-impedance bottom region to bear a higher proportion of current flow and reaction load. When the current direction changes, the direction of electrochemical reaction and ion migration will change, but the local Joule heating is mainly determined by the current magnitude and local current flow path, and will not disappear due to the reversal of current polarity. Therefore, this invention maintains continuous or intermittent local heating conditions in the bottom region inside the abnormally parallel battery cell through alternating charge and discharge, rather than relying on a fixed current direction to form bottom heating. The fact that the driving current flows through the bottom region should be understood as the bottom region carrying a larger proportion of current than the top region, not that all the current is confined to the bottom region.

[0104] During alternating charge-discharge maintenance, the controller pauses alternating charge-discharge at preset time intervals and re-detects the concentration stratification status of abnormal parallel battery cells, such as... Figure 7As shown. The preset time interval can be from 5 minutes to 120 minutes, preferably from 10 minutes to 60 minutes. The pause time can be from 1 second to 60 seconds, so that the port voltage response tends to stabilize after the large current of charging and discharging ends. After the pause, the controller injects an asymmetric broadband AC current pulse sequence again through the positive and negative terminals of the abnormal parallel battery cell to obtain the updated impedance response sequence. The updated impedance response sequence is input into the trained long short-term memory network model, which outputs the current electrolyte spatial concentration distribution sequence. The controller extracts the current maximum concentration value and the current minimum concentration value and calculates the current difference.

[0105] When the difference between the current maximum concentration value and the current minimum concentration value is less than the preset uniformity threshold, it indicates that the concentration stratification within the abnormal parallel battery cell has been mitigated to an acceptable range. The controller generates a convection termination command to stop the alternating charging and discharging between the target parallel battery cell and the abnormal parallel battery cell, and updates the status of the abnormal parallel battery cell to a candidate cell that can participate in subsequent discharge scheduling. If the current difference is still greater than or equal to the preset uniformity threshold, the controller continues to perform alternating charging and discharging maintenance. To avoid excessively long maintenance times, the controller can also set a maximum maintenance duration, such as 1 hour to 24 hours; if the uniformity state is not restored after the maximum maintenance duration is reached, a maintenance prompt is generated.

[0106] In a further embodiment, while performing alternating charge and discharge, the controller activates a targeted thermal control mechanism for abnormally connected parallel battery cells, such as... Figure 8 As shown. The thermal management system may include a bottom cooling plate, a top cooling duct, a liquid cooling circuit, an air-cooled fan, cooling coils, or temperature-controlled valves. The controller controls the thermal management system corresponding to the abnormal parallel battery cell to stop or reduce bottom cooling and to start or enhance top cooling. This can create a temperature boundary condition inside the abnormal parallel battery cell where the bottom temperature is relatively high and the top temperature is relatively low, enhancing the driving force for the heated electrolyte to flow from the bottom to the top.

[0107] During the operation of the directional thermal control mechanism, the controller collects the bottom casing temperature of the abnormal parallel battery cell in real time. The bottom casing temperature can be collected via thermocouples, platinum resistance thermometers, NTC temperature sensors, or infrared temperature sensors. The preset thermal runaway warning threshold can be determined based on the battery type, for example, it can be between 45°C and 80°C. For lead-acid or lead-carbon energy storage batteries, it is preferably set between 50°C and 65°C. This threshold is not equivalent to the actual thermal runaway temperature, but rather a protection threshold with a safety margin. When the bottom casing temperature reaches the preset thermal runaway warning threshold, the controller immediately controls the thermal management system to resume cooling of the bottom of the abnormal parallel battery cell and sends a power zeroing command to the bidirectional converter, cutting off the charging and discharging current. This process prevents further accumulation of localized Joule heat at the bottom, reducing the risk of localized overheating, gas evolution, or abnormal casing temperature rise.

[0108] In another embodiment, before inputting the impedance response sequence into the long short-term memory network model, the controller performs a bubble interference screening operation for each parallel battery cell. This operation is used to distinguish between the actual electrolyte concentration gradient and impedance anomalies caused by bubble adhesion, local gas evolution, or interface polarization. Because when bubbles adhere to the electrode surface or near the separator, they may also cause phase shifts or increases in port impedance, which may be misjudged by the model as concentration stratification if not screened.

[0109] During bubble interference detection, the controller first injects a DC bias voltage signal through the positive and negative terminals of the corresponding parallel battery cell and maintains it for a preset time period. The DC bias voltage signal can be 0.1% to 5% of the rated voltage of the single cell, and the maintenance time can be 1 second to 120 seconds. During the injection process, a current-limiting circuit limits the bias current to avoid significant charging and discharging or inducing excessive gas evolution. This DC bias voltage is not used to significantly change the battery's state of charge, but rather to change the bubble attachment interface, local double-layer state, or slight polarization state. If the impedance anomaly mainly originates from bubble attachment, the bubble interface state may change after DC bias treatment, resulting in a significant morphological shift between the second impedance response sequence and the first impedance response sequence. If the impedance anomaly originates from the actual concentration gradient, short-term DC bias is usually insufficient to change the overall concentration distribution of the electrolyte inside the battery, and the consistency between the two impedance response sequences will be higher.

[0110] In some embodiments, when performing bubble interference detection, the controller preferably performs cell-level gating on the parallel battery cell to be detected, temporarily isolating the parallel battery cell from other parallel battery cells through an independent detection branch, an isolated detection converter, or an electronic switch, or placing other parallel branches in a high-impedance detection state, thereby preventing the DC bias voltage signal from forming an uncontrollable circulating current among multiple parallel battery cells. Furthermore, the controller calculates the net charge introduced by the DC bias process through current integration and controls the net charge within 0.001% to 0.2% of the rated capacity of the corresponding parallel battery cell, preferably within 0.001% to 0.05%. If the net charge reaches a preset upper limit, or if any parameter such as terminal voltage, current, or temperature exceeds the allowable range, the controller immediately removes the DC bias voltage signal. Thus, the DC bias operation is mainly used to disturb the bubble attachment interface and the local double-layer state without significantly changing the remaining charge ratio of the parallel battery cells.

[0111] At the termination point when the DC bias voltage signal is removed, the controller injects a second asymmetric broadband AC current pulse sequence into the corresponding parallel battery cell through the positive and negative terminals, and acquires a second impedance response sequence. The frequency components, net charge limits, and sampling method of the second pulse sequence can be kept consistent with the initial detection to ensure the comparability of the two impedance response sequences. Subsequently, the controller concatenates the initially acquired impedance response sequence and the second impedance response sequence into a dual-channel input sequence, and inputs it into a trained convolutional neural network binary classification model.

[0112] A convolutional neural network binary classification model can include a dual-channel input layer, 1 to 4 one-dimensional convolutional layers, pooling layers, fully connected layers, and a binary classification output layer. The one-dimensional convolutional layers extract local morphological features of the impedance curve along frequency order, such as abrupt changes in the high-frequency band, tail bending in the low-frequency band, phase angle shifts, and curve shifts between two probes. The kernel length can be 3 to 9, and the number of convolutional channels can be 16 to 128. The binary classification output layer outputs a classification label indicating whether bubble attachment interference exists. The training samples for this model can include impedance dual-channel samples known to have bubble attachment and impedance dual-channel samples known to have a true concentration gradient. The classification loss can be cross-entropy loss, and the model parameters can be iteratively updated using a gradient descent algorithm until the classification accuracy on the validation set reaches a preset requirement.

[0113] When the classification label indicates interference, the controller confirms the identification result as bubble attachment interference and injects a negative depolarization current pulse into the corresponding parallel battery cell through the positive and negative terminals. The negative depolarization current pulse can be set to 0.02C to 0.3C, and its duration can be 0.1 seconds to 30 seconds. This negative pulse is used to reduce local polarization, change the interfacial electric field state near the attached bubble, and promote bubble desorption or disturb the bubble's position. After depolarization is complete, the controller can re-perform impedance probing to avoid inputting transient impedance anomalies caused by bubbles into the long short-term memory network model.

[0114] When the classification label indicates the absence of interference, the controller confirms the identification result as the true concentration gradient and triggers the input of the initially acquired impedance response sequence into the long short-term memory network model. This improves the reliability of the concentration distribution identification results and reduces misjudgments caused by local bubbles or interface anomalies.

[0115] This invention also provides an adaptive load distribution system for parallel power supply systems, such as... Figure 2As shown, the system includes an AC pulse injection and acquisition module, a state prediction and analysis module, a discharge scheduling and allocation module, and a circulating current charge and discharge control module. The AC pulse injection and acquisition module can be composed of a pulse current source, a current sampling resistor, a Hall current sensor, a voltage sampling circuit, an analog-to-digital converter, and a synchronous clock circuit. It is used to inject an asymmetric broadband AC current pulse sequence into multiple parallel battery cells within the detection time window, and simultaneously acquire the voltage fluctuation response sequence and the actual injected current sequence, thereby generating an impedance response sequence.

[0116] The state prediction and analysis module can be implemented using an embedded processor, digital signal processor, edge computing unit, or industrial control computer. This module stores the trained Long Short-Term Memory (LSTM) network model and inputs the impedance response sequence into it to obtain the electrolyte spatial concentration distribution sequence within each parallel battery cell. The state prediction and analysis module also extracts the maximum and minimum concentration values ​​from the concentration distribution sequence and determines whether the parallel battery cell belongs to the target parallel battery cell or an abnormal parallel battery cell based on the difference between the two values.

[0117] The discharge scheduling and allocation module receives discharge scheduling commands from external energy management systems or grid dispatching systems and reads the total load power. Based on the concentration uniformity provided by the state prediction and analysis module, this module allocates the total load power to target parallel battery cells, ensuring that cells with more uniform concentration distribution handle the discharge output. This module can also output power limiting suggestions or maintenance prompts when there are insufficient target parallel battery cells.

[0118] The circulating current charge / discharge control module, upon receiving an idle state command, controls the energy flow between the target parallel battery cell and the abnormal parallel battery cell by adjusting the bidirectional converter. Based on the remaining charge ratio of the abnormal parallel battery cell, this module controls the pulse width modulation duty cycle of the bidirectional converter, causing the current to flow in reverse when it reaches a preset upper and lower safety threshold, thus forming an alternating charge / discharge process. This module can also be integrated with the thermal management system to perform bottom cooling shutdown, top cooling, bottom temperature monitoring, and over-temperature protection during maintenance.

[0119] In a more complete system embodiment, the system may further include a bubble interference detection module and a periodic retest termination module. The bubble interference detection module is used to perform DC bias voltage injection, second impedance detection, dual-channel input sequence splicing, convolutional neural network binary classification judgment, and negative direction depolarization current pulse injection. The periodic retest termination module is used to pause the maintenance process at preset time intervals during alternating charge and discharge, reacquire the impedance response sequence of abnormal parallel battery cells, and determine whether to generate a convection termination command based on the current concentration distribution output by the long short-term memory network model.

[0120] Through the above embodiments, the present invention can infer the spatial concentration distribution inside a liquid electrolyte battery by utilizing port impedance response without disassembling the battery or damaging the sealed structure, and use this internal state for load distribution in a parallel energy storage system. During the discharge phase, the system avoids abnormal parallel battery cells with obvious concentration stratification. During the idle phase, it utilizes the alternating charge and discharge between the target parallel battery cell and the abnormal parallel battery cell to induce localized Joule heating at the bottom and bottom-up thermal convection. Simultaneously, the effectiveness and safety of the maintenance process are ensured through periodic retesting, directional thermal control, and bubble interference identification.

[0121] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An adaptive load allocation method for a parallel power supply system, applied to a battery energy storage system using liquid electrolyte, characterized in that, include: Within the detection time window, an asymmetric broadband AC current pulse sequence is injected through the positive and negative terminals of each of multiple parallel battery cells. Simultaneously, the corresponding voltage fluctuation response sequence and the actual injected current sequence are acquired at the positive and negative terminals. An impedance response sequence is obtained based on the voltage fluctuation response sequence and the actual injected current sequence. The impedance response sequence is then input into a long short-term memory network model, which outputs the electrolyte spatial concentration distribution sequence within each parallel battery cell. The maximum and minimum concentration values ​​in the electrolyte spatial concentration distribution sequence of each parallel battery cell are extracted. Upon receiving a discharge scheduling command, the total load power carried by the discharge scheduling command is allocated to the target parallel battery cell whose difference between the maximum concentration value and the minimum concentration value is less than a preset uniformity threshold. Upon receiving an idle state command, the target parallel battery cell is controlled to output electrical energy to the abnormal parallel battery cell whose difference is greater than or equal to the preset uniformity threshold by adjusting the bidirectional converter. The driving current flows through the bottom region inside the abnormal parallel battery cell to generate bottom-up heat convection by stimulating local Joule heating in the bottom region.

2. The method according to claim 1, characterized in that, The asymmetric broadband AC current pulse sequence is a bidirectional current pulse perturbation signal containing multiple components of different frequencies, wherein the amplitude of the forward current pulse is not equal to that of the reverse current pulse, and the net charge of the bidirectional current pulse perturbation signal within one detection period is less than a preset charge threshold; the steps for obtaining the impedance response sequence include: The instantaneous current value and voltage fluctuation response sequence during the pulse injection are collected synchronously, and the actual injection current sequence is formed from the instantaneous current value according to the sampling time sequence. The voltage fluctuation response sequence and the actual injected current sequence are transformed in the frequency domain to extract the voltage frequency domain complex value and the current frequency domain complex value under multiple frequency components. Divide the voltage frequency domain complex value under the same frequency component by the corresponding current frequency domain complex value to calculate the impedance complex value under the corresponding frequency component. The impedance response sequence is generated by arranging at least one of the impedance complex values, impedance magnitudes, impedance phase angles, and the real or imaginary parts of the impedance at each frequency component in frequency order.

3. The method according to claim 1, characterized in that, The training process of the Long Short-Term Memory network model includes: Obtain the sample impedance response sequence from historical operating data and the actual electrolyte spatial concentration distribution sequence corresponding to the sample impedance response sequence; The sample impedance response sequence is used as the input feature to input the initial long short-term memory network model, and the output is the predicted distribution sequence; The actual electrolyte spatial concentration distribution sequence is used as a supervision label, and the cross-entropy loss function is used to quantify the numerical deviation between the predicted distribution sequence and the supervision label. The network parameters of the initial long short-term memory network model are iteratively updated using the gradient descent algorithm until the numerical deviation meets the preset convergence condition, thereby obtaining the trained long short-term memory network model.

4. The method according to claim 3, characterized in that, The process of obtaining the actual electrolyte spatial concentration distribution sequence includes: Concentration detection probes are deployed at different physical height positions inside a pre-set multi-segment test battery pack. The concentration detection probes include fiber optic refractive index concentration probes or miniature ultrasonic density probes. The concentration probe synchronously collects the actual concentration values ​​of the electrolyte at different physical height locations; All the actual electrolyte concentration values ​​at the same time point are spliced ​​together in descending order of physical height to generate the actual electrolyte spatial concentration distribution sequence as the supervision label.

5. The method according to claim 1, characterized in that, Controlling the output of electrical energy from the target parallel battery cell to the abnormal parallel battery cell by adjusting the bidirectional converter includes: Real-time acquisition of the remaining power percentage of the abnormal parallel battery units; When the remaining power ratio is less than a preset safety upper limit threshold, the bidirectional converter is controlled to continuously input charging current from the target parallel battery unit to the abnormal parallel battery unit. When the remaining power ratio reaches the preset safety upper limit threshold, the pulse width modulation duty cycle of the internal switching device of the bidirectional converter is adjusted to control the output voltage value of the abnormal parallel battery unit to be greater than the receiving voltage value of the target parallel battery unit to reverse the current flow direction. When the remaining power ratio drops to the preset safety lower limit threshold, the current flow direction is reversed again, and the local Joule heating is maintained by alternating charging and discharging.

6. The method according to claim 5, characterized in that, During the time period of the alternating charge and discharge, the alternating charge and discharge is paused at a preset time interval, and the asymmetric broadband AC current pulse sequence is re-injected into the abnormal parallel battery cell through the positive and negative terminals to obtain the updated impedance response sequence. The updated impedance response sequence is input into the long short-term memory network model, which outputs the current electrolyte spatial concentration distribution sequence and extracts the current maximum concentration value and the current minimum concentration value. When the difference between the current maximum concentration value and the current minimum concentration value is less than the preset uniformity threshold, a convection termination command is generated to stop the alternating charging and discharging.

7. The method according to claim 5, characterized in that, While performing the alternating charge and discharge, a targeted thermal control mechanism for the abnormal parallel battery cells is activated, including: The thermal management system corresponding to the abnormal parallel battery unit is controlled to stop cooling the bottom of the abnormal parallel battery unit, and the thermal management system is controlled to start cooling the top of the abnormal parallel battery unit. During the operation of the directional thermal control mechanism, the bottom casing temperature of the abnormal parallel battery unit is collected in real time; When the temperature of the bottom outer casing reaches the preset thermal runaway warning threshold, the thermal management system is controlled to resume cooling of the bottom of the abnormal parallel battery cell and a power zeroing command is sent to the bidirectional converter to cut off the charging and discharging current.

8. The method according to claim 1, characterized in that, Before inputting the impedance response sequence into the long short-term memory network model, a bubble interference screening operation is performed for each parallel battery cell, including: A DC bias voltage signal is injected into the corresponding parallel battery cell through the positive and negative terminals and maintained for a preset time period. At the termination time node when the DC bias voltage signal is removed, a second asymmetric broadband AC current pulse sequence is injected into the corresponding parallel battery cell through the positive and negative terminals to obtain the second impedance response sequence. The first impedance response sequence and the second impedance response sequence are input into a trained convolutional neural network binary classification model to output the discrimination result.

9. The method according to claim 8, characterized in that, The initially acquired impedance response sequence and the second impedance response sequence are input into a trained convolutional neural network binary classification model, and the output discrimination results include: The impedance response sequence obtained initially is concatenated with the impedance response sequence obtained second time to form a dual-channel input sequence; The dual-channel input sequence is input into the convolutional neural network binary classification model. The convolutional neural network binary classification model extracts the morphological shift features between the first impedance response sequence and the second impedance response sequence, and outputs a classification label characterizing whether there is bubble attachment interference. When the classification label indicates interference, the identification result is confirmed to be bubble adhesion interference, and a negative depolarization current pulse is injected into the corresponding parallel battery unit through the positive and negative terminals. When the classification label indicates that there is no interference, the identification result is confirmed as the true concentration gradient, and the operation of inputting the initially acquired impedance response sequence into the long short-term memory network model is triggered.

10. An adaptive load sharing system for implementing the method of claim 1 in a parallel power supply system, characterized in that, include: An AC pulse injection and acquisition module is used to inject an asymmetric broadband AC current pulse sequence through the positive and negative terminals of multiple parallel battery cells within a detection time window, and simultaneously acquire the corresponding voltage fluctuation response sequence and actual injected current sequence at the positive and negative terminals, and obtain the impedance response sequence based on the voltage fluctuation response sequence and the actual injected current sequence. The state prediction and analysis module is used to input the impedance response sequence into the long short-term memory network model, and the long short-term memory network model outputs the electrolyte spatial concentration distribution sequence inside each parallel battery cell; and extracts the maximum and minimum concentration values ​​in the electrolyte spatial concentration distribution sequence of each parallel battery cell. The discharge scheduling and allocation module is used to receive a discharge scheduling instruction and allocate the total load power carried by the discharge scheduling instruction to the target parallel battery cell whose difference between the maximum concentration value and the minimum concentration value is less than a preset uniformity threshold. The circulating current charging and discharging control module is used to receive idle state commands and control the target parallel battery cell to output electrical energy to the abnormal parallel battery cell whose difference is greater than or equal to the preset uniformity threshold by adjusting the bidirectional converter. The driving current flows through the bottom region inside the abnormal parallel battery cell to generate local Joule heating in the bottom region to produce heat convection from bottom to top.

Images