A method for regulating the parallel charging current of multiple chargers based on a battery parallel system
By using CAN bus communication between the BMS and the charger and deep learning control, the problems of high hardware cost, poor scalability and current surge oscillation in lithium battery parallel systems are solved, realizing flexible, safe and efficient parallel charging of multiple chargers.
Patent Information
- Application Number
- CN202611610765.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-19
- Publication Date
- 2026-11-17
AI Technical Summary
In existing lithium battery parallel charging systems, multiple chargers connected in parallel suffer from high hardware costs, poor scalability, insufficient safety, and current surges and oscillations, making it impossible to achieve flexible, safe, and efficient charger expansion and smooth current transition.
By communicating with multiple chargers through the Battery Management System (BMS), and using a CAN bus current limiting percentage adjustment method, combined with deep learning and intelligent diagnostic technology, dynamic control and failure detection of the chargers can be achieved, avoiding hardware current detection circuits and enabling plug-and-play and flexible expansion.
It reduces hardware costs, improves system stability and equipment lifespan, enables flexible expansion of multi-charger systems and smooth current transition, and solves the problems of current surge and oscillation.
Smart Images

Figure CN123121906A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of lithium battery technology, and in particular to a method for regulating the parallel charging current of multiple chargers based on a battery parallel system. Background Technology
[0002] In lithium battery parallel charging applications, current technologies generally employ a single high-power charger to charge multiple batteries in parallel systems. To meet fast charging demands, this approach necessitates continuously increasing the charger's rated power, leading to a sharp rise in charger hardware costs and posing a potential threat to battery safety and cycle life. Regarding system scalability, each additional battery module often requires re-evaluation and replacement of charger power devices (such as charging MOSFETs), resulting in high R&D costs and hindering plug-and-play and flexible expansion capabilities. Furthermore, existing solutions lack smooth transition control of the total charging current during the startup phase of multi-charger parallel charging or the shutdown phase of a single charger, easily causing current surges or oscillations that affect system stability and equipment lifespan. Therefore, achieving flexible, safe, and efficient multi-charger parallel charging control at a lower hardware cost has become a pressing technical challenge in this field. Summary of the Invention
[0003] This invention overcomes the shortcomings of the prior art and provides a method for regulating the parallel charging current of multiple chargers based on a battery parallel system.
[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows: This invention discloses a method for regulating the charging current of multiple chargers in a parallel battery system, applied to a battery management system (BMS). The BMS communicates with multiple chargers via a CAN bus, and includes the following steps: Step S1: When a charger is detected to be connected, the current limiting percentage is set to the initial current limiting percentage, and a current limiting message containing the initial current limiting percentage is generated and sent to the charger. Step S2: Determine whether the battery communication in the parallel system is disconnected or whether the battery is in a protection or recovery state; if yes, reset the current limiting percentage to the initial current limiting percentage, send the corresponding current limiting message, and then repeat this step; if no, proceed to step S3. Step S3: Obtain the maximum battery current of the parallel system and calculate the absolute value of the difference between the maximum battery current and the preset target current; if the absolute value of the difference is not greater than the preset deviation threshold, maintain the current current limiting percentage unchanged and jump to step S5; if it is greater, execute step S4. Step S4: Compare the maximum battery current with the preset target current. If the maximum current of the battery is greater than the preset target current, then when the current current limiting percentage is greater than the preset lower limit percentage, the current current limiting percentage will be reduced by a preset adjustment step size; otherwise, the current current limiting percentage will remain unchanged. If the maximum current of the battery is less than the preset target current, then when the current current limiting percentage is less than 100%, the current current limiting percentage will be increased by the preset adjustment step size; otherwise, the current current limiting percentage will remain unchanged. A current-limiting message is generated based on the adjusted current-limiting percentage and sent to the charger; Step S5: Determine if the charger is faulty. If so, end the adjustment; otherwise, return to step S2.
[0005] Preferably, the preset target current is 11A and the preset deviation threshold is 3A.
[0006] Preferably, the initial current limiting percentage is 25%, the preset lower limit percentage is 25%, and the preset adjustment step size is 25%.
[0007] Preferably, the maximum battery current is the maximum value among the charging currents of all batteries in the parallel system.
[0008] Preferably, the current limiting message is a CAN communication message used to instruct the charger to limit its output current to the value corresponding to the current limiting percentage of the charger's rated current.
[0009] Preferably, the BMS obtains the output current of each charger through CAN communication to determine the maximum current of the battery, and the charger does not need to be equipped with a hardware current detection circuit for charging current measurement.
[0010] Preferably, determining whether the charger is faulty specifically includes: During the initial power-on self-test of the parallel system, a set of synthetic samples with clear failure labels is generated based on the characteristic parameter boundaries of various typical charger failure modes pre-stored in the BMS; the synthetic samples are then mixed with the historical feature sample set collected under historical normal charging conditions to form a hybrid training set. The isolated forest is trained using the hybrid training set, thereby constructing a decision tree structure that can distinguish between normal sample clusters and various known failed sample clusters in the initial stage of the model through random cutting. From the continuous CAN messages uploaded by the target charger cycle, extract multiple frames of data within a preset time window, analyze and calculate the cumulative degradation characteristics of electrical stress, including at least: the impulse density of pulse current spikes per unit time, the thermal stress cycle factor based on the joint weighting of shell temperature and output current, and the eye diagram opening attenuation coefficient reflecting the signal quality of the communication physical layer; to form the degradation feature vector to be detected at the current moment. The degraded feature vector to be detected is input into the isolated forest model, which traverses each decision tree and records the number of path edges it traverses in each decision tree to be completely isolated as the isolation depth. At the same time, during the path search process, the proportion of synthetic samples included in the branches that the corresponding sample falls into at the split node is counted, thereby assigning a failure tendency index to the corresponding path. After traversal, the arithmetic mean of the isolation depths of the degenerate feature vector to be detected under all decision trees is calculated to obtain the average isolation depth, and the arithmetic mean of the failure tendency index on all paths is calculated to obtain the average failure tendency index. The average isolation depth is mapped to a basic anomaly score through a normalization factor, and then the average failure tendency index is used as a weight to correct the anomaly score to obtain the failure risk score. The failure risk score is compared with a dynamic failure decision threshold, which is the upper quantile calculated in real time based on the failure risk score distribution of all chargers in the current parallel system within the same time window. At the same time, it is determined whether each feature in the electrical stress cumulative degradation feature vector exceeds its respective preset absolute safety physical boundary. If the failure risk score exceeds the dynamic failure decision threshold, or any one of the electrical stress cumulative degradation features exceeds its absolute safety physical boundary, then the target charger is finally determined to have failed.
[0011] Preferably, after the rate limiting percentage remains unchanged or is adjusted, but before sending the message, the following steps are also included: The system collects time-series data from the historical charging process of the parallel charging system, forming three sets of parallel data streams: the first data stream is a sequence of current limiting percentage instructions with control cycle granularity; the second data stream is an instantaneous frequency sequence extracted from the output current values of each charger at the corresponding moment through Hilbert-Huang transform; and the third data stream is a sequence of load mutation event labels recorded by the battery management system at the corresponding moment. The three sets of parallel data streams are aligned by timestamps to construct a multi-stream input sample set, which is used to train a convolutional long short-term memory network model that incorporates attention gating. Obtain the current limiting percentage instructions of the current control cycle and several previous control cycles, the instantaneous frequency of the output current value of each charger extracted by Hilbert-Huang transform, and the load change event label, and generate the current multi-stream input feature vector according to the structure of the multi-stream input sample set; The current multi-stream input feature vector is input into the trained convolutional long short-term memory network model with fusion attention gating. After model inference, the predicted value of the current spike in the next control cycle is output. The predicted value of the current spike is the overshoot of the instantaneous current peak that may occur in the next control cycle relative to the current steady-state current value. Obtain a preset current spike tolerance limit, calculate the difference between the current spike prediction value and the current spike tolerance limit, and use the difference as the feedforward compensation amount; while maintaining the adjusted current limiting percentage obtained in step S4 or the current limiting percentage that remains unchanged, subtract the current limiting percentage correction value corresponding to the feedforward compensation amount to obtain the current limiting percentage after feedforward correction. The current limiting percentage after feedforward correction is subjected to boundary limiting processing with the preset lower limit percentage and 100%. If the current limiting percentage after feedforward correction is lower than the preset lower limit percentage, the preset lower limit percentage is used; if it is higher than 100%, 100% is used. A current limiting message is generated based on the current limiting percentage after limiting processing and sent to the charger.
[0012] This invention addresses the technical deficiencies in the prior art and has the following beneficial effects: This invention achieves the technical effect of replacing high-cost hardware stacking with pure software algorithm control. Using a BMS as the host, it implements unified current-limiting percentage control over multiple identical chargers via CAN communication. By completely eliminating the need for hardware current detection circuits at the charger end, it reduces the hardware manufacturing cost and R&D adaptation threshold of multi-charger systems. Simultaneously, this closed-loop control mechanism can precisely and smoothly transition the total charging current at each stage, including charging initiation, dynamic adjustment, and charger exit, fundamentally suppressing current surges and oscillations, significantly improving system stability and equipment lifespan. Furthermore, based on simple percentage commands, it enables plug-and-play and flexible expansion of any number of chargers without adjusting any hardware parameters due to changes in the number of batteries connected in parallel, fundamentally solving the problems of high cost, poor scalability, and insufficient safety in multi-charger systems. Attached Figure Description
[0013] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other embodiments can be obtained from these drawings without creative effort.
[0014] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation
[0015] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.
[0016] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.
[0017] This invention provides a method for regulating the parallel charging current of multiple chargers based on a battery parallel system, which is applied to a battery management system (BMS). The BMS acts as the control master, communicating bidirectionally with multiple parallel chargers of the same model and specifications (as slaves) via a CAN bus. Each charger is uniformly controlled by the BMS, and each charger itself does not require additional high-precision hardware current detection circuitry. Its output current is autonomously adjusted by the charger based on the received current-limiting message, and the specific current value is reported to the BMS in real time via CAN messages through its built-in conventional detection function.
[0018] The overall control flow of the method of the present invention is as follows: Figure 1 As shown, the specific steps include: Step S1 (Initialization Phase): When the BMS detects that one or more chargers are connected to the parallel system via the CAN bus, it triggers the entire charging control process. To ensure the safety of the charging process startup and avoid current surges at power-on, the BMS first sets a low initial current limit percentage, preferably 25% of the charger's rated current. Subsequently, the BMS encapsulates this initial current limit percentage into a standard CAN communication message, i.e., a current-limiting message, and broadcasts it to the CAN bus. Upon receiving this message, all connected chargers will limit their respective output current to 25% of their rated current.
[0019] Step S2 (Status Check and Loop Judgment): After sending the initial current limiting message, the BMS enters a continuous monitoring loop. First, the BMS determines whether the overall state of the parallel system is suitable for continuing charging. Specific judgment conditions include: whether battery communication is disconnected, or whether the battery is in a protection state (such as over-temperature protection, overcharge protection, short-circuit protection, etc.) or in the transition process of protection recovery. If any condition is met (Y), it indicates that the current system state is unstable or unsuitable for charging. At this time, the BMS will forcibly reset the current limiting percentage to the initial current limiting percentage (25%) and resend the corresponding current limiting message to bring all chargers back to a safe low-power output state. Then, the BMS will repeat this step until the system state returns to normal. If none of the above conditions are met (N), it indicates that the system communication is stable and the battery is in a state that can be charged normally, and the process proceeds to the next step, S3.
[0020] Step S3 (Current Deviation Judgment): The BMS acquires the charging current of all batteries in the parallel system in real time via the CAN bus and selects the one with the largest value, defining it as the "maximum battery current". The BMS has preset target current and deviation threshold; preferably, the target current is 11A and the deviation threshold is 3A. Next, the BMS calculates the absolute value of the difference between the "maximum battery current" and the preset target current (11A). If the absolute value of the difference is not greater than the preset deviation threshold (3A), it means that the current total charging current is close to the target value, the system has entered a steady state, and no adjustment is needed. At this time, the BMS maintains the current current limiting percentage unchanged and directly jumps to step S5. If the absolute value of the difference is greater than the preset deviation threshold (3A), it means that the actual current deviates significantly from the target and dynamic intervention is required. The process then proceeds to the next step, S4.
[0021] Step S4 (Dynamic Adjustment Logic): In this step, the BMS performs current boosting or current reduction operations based on the direction of the current deviation.
[0022] The BMS first compares the "maximum battery current" with the "preset target current": Scenario A (Current Limit Required): If the "Maximum Battery Current" is greater than the "Preset Target Current," it indicates that the total charging current is too high. To protect the battery, the current limit percentage needs to be reduced. However, before performing the reduction operation, the BMS performs a boundary check: determining whether the current current limit percentage is greater than a preset lower limit percentage. In this embodiment, the preset lower limit percentage is preferably 25%. If the current current limit percentage is greater than 25%, the BMS reduces it by a preset adjustment step, preferably 25%; if the current current limit percentage is equal to or lower than 25%, to avoid the charger being completely shut off or reverse-charged, the BMS will maintain the current current limit percentage unchanged and will not reduce it further.
[0023] Scenario B (Current Boost Required): If the "Maximum Battery Current" is less than the "Preset Target Current," it indicates that the total charging current has room for improvement, which can speed up the charging process. The BMS will determine whether the current current limit percentage is less than 100%. If it is less than 100%, the BMS will increase it by the preset adjustment step size of 25%; if the current current limit percentage has already reached 100%, it means that the charger is already at full power output, and the current current limit percentage will remain unchanged.
[0024] After making the decision on whether to adjust or not, the BMS will generate a new current limiting message based on the final determined current limiting percentage and prepare to send it to each charger.
[0025] In addition, before the BMS generates a current-limiting message based on the final determined current-limiting percentage and sends it to the charger, in order to cope with current spikes in the next control cycle that may be caused by sudden load changes or dynamic system adjustments, this invention introduces a prediction feedforward correction mechanism based on deep learning. The specific implementation details are as follows: First, an attention-gated convolutional long short-term memory network model needs to be trained offline. The data acquisition and construction process is as follows: During the historical charging process of the parallel system, three sets of parallel time-series data streams are simultaneously collected using a data logger. The first data stream is a sequence of current-limiting percentage commands issued by the BMS at the control cycle granularity, representing the system's input control intent. The second data stream is the instantaneous frequency sequence extracted after performing a Hilbert-Huang transform on the output current value reported by each charger at the corresponding time, which best characterizes the non-stationary changes in current. This instantaneous frequency is extremely sensitive to sudden current fluctuations and spike precursors, containing richer dynamic information than the original current amplitude signal. The third data stream is a sequence of load mutation event labels recorded by the BMS at the corresponding time, representing abrupt changes in load or system state, such as "a new charger suddenly connects" or "a high-power load is momentarily disconnected." These three sets of data streams are strictly aligned according to their timestamps to form a multi-stream input sample set for training the model.
[0026] During actual online prediction, the BMS acquires the current limiting percentage commands for the current control cycle and several preceding consecutive control cycles, the instantaneous frequency extracted from the current signal via Hilbert-Huang transform at the corresponding moment, and the recorded load mutation event labels. Following the same structure as during training, it generates a "current multi-stream input feature vector" in real time. This vector is then input into the trained model for inference. The model automatically extracts local features through its internal convolutional layers and captures long-term dependencies through long short-term memory (LSTM) layers. Finally, with the assistance of attention gating, it accurately outputs the "predicted current spike value for the next control cycle." This predicted value specifically refers to the overshoot of the instantaneous current peak value relative to the current steady-state current value that may occur in the next control cycle.
[0027] Next, the BMS obtains a preset upper limit for current spike tolerance that the system can tolerate. This upper limit is determined based on battery specifications and system design safety margins. The BMS calculates the difference between the predicted overshoot and the upper limit. If the predicted overshoot exceeds the upper limit, this difference is used as a feedforward compensation; otherwise, the compensation is zero. Finally, based on the current limiting percentage calculated in step S4 or kept constant, a current limiting percentage correction value corresponding to the feedforward compensation is subtracted in advance to obtain the "current limiting percentage after feedforward correction". To ensure safety, the corrected current limiting percentage also needs to undergo boundary limiting processing with a preset lower limit percentage (25%) and upper limit percentage (100%). Finally, the BMS generates and sends a current limiting message based on the final current limiting percentage after the limiting processing.
[0028] In this embodiment, the operation of "determining whether the charger is faulty" adopts an intelligent diagnostic method, which includes two stages: offline model construction and online real-time diagnosis.
[0029] Offline Model Building Phase: The model building process is executed during the initial power-on self-test of the parallel system or during firmware upgrades via the BMS maintenance interface. First, the feature parameter boundaries of various typical charger failure modes pre-stored in the BMS non-volatile memory are retrieved. These failure modes include, for example, increased output ripple due to aging of the output electrolytic capacitor, a sharp drop in efficiency and increased heat generation due to soft breakdown of the power MOSFET, and deterioration in signal quality due to CAN transceiver performance degradation. Based on these boundaries, a batch of synthetic samples with clear failure labels is generated through data synthesis. These synthetic samples are then mixed with a large set of historical feature samples collected during normal charging conditions to form an information-rich hybrid training set. This hybrid training set is used to train the isolation forest. The core of the training is to randomly select a feature dimension and randomly choose a cutoff value between the maximum and minimum values of that dimension, recursively bisecting the sample space. Due to the introduction of synthetic samples with clear failure characteristics, the trained multiple isolation trees can not only quickly isolate abnormal unknown samples but also form a unique decision tree branch structure that can distinguish normal sample clusters from various known failure sample clusters.
[0030] Online real-time diagnostic phase: During charging, the BMS extracts multiple frames of data within a sliding time window from the periodically uploaded continuous CAN messages of each connected target charger. This data is analyzed, and a set of degradation characteristics reflecting the early cumulative effects of electrical stress is calculated. These characteristics are more sensitive to early failures than conventional voltage, current, and temperature metrics. Specifically, these include: Impulse density of pulse current spikes per unit time: This is calculated by counting the number of pulse spikes within a preset time window where the charger's output current exceeds 110% of its rated current. The duration and amplitude of each spike are integrated, summed, and then divided by the time window length. A sudden increase in this value reflects degradation of the charger's internal feedback control loop.
[0031] The thermal stress cycle factor, weighted by a combination of case temperature and output current, calculates the cumulative absolute value of the product of the "case temperature change rate" and the "output current change rate" over a time window, rather than solely considering temperature values. This factor accurately quantifies the impact of thermomechanical stress caused by load fluctuations on the device and is a typical precursor to solder joint fatigue in power devices.
[0032] The eye diagram opening attenuation coefficient, reflecting the signal quality of the physical layer of communication, is used by the BMS's CAN controller to perform signal integrity analysis on the target charger's messages and extract the eye diagram of its differential signal. The attenuation ratio of the vertical opening of the eye diagram relative to the standard template is calculated at the optimal sampling point. A continuous decrease in this coefficient indicates that the performance of the CAN transceiver or bus connector is degrading, potentially leading to intermittent communication failures. These features constitute the "degradation feature vector to be detected" at the current moment. This degradation feature vector is then input into the constructed isolated forest model, which traverses each decision tree in the model. During the traversal, not only is the path length (isolation depth) from the root node to the isolated leaf node recorded, but an intermediate quantity is also introduced: the failure propensity index. When each internal node of each tree splits, the BMS queries the proportion of the aforementioned synthetic samples contained in the branch that the node belongs to. The higher the proportion of synthetic samples in the branch the vector falls into, the more the system accumulates it along the path and ultimately assigns a higher "failure propensity index." This index correlates the model's structural information with prior knowledge of failure in real time.
[0033] After traversal, the average isolation depth and average failure propensity index under all decision trees are obtained. First, using the average isolation depth and a normalization factor related to the number of samples in the model, the basic anomaly score is calculated. A score close to 1 indicates that the sample is very easily isolated, i.e., highly anomaly. Next, the average failure propensity index is used as a correction weight to strengthen the basic anomaly score. For example, if a sample is isolated quickly (high basic anomaly score) and its path is full of failed samples (high average failure propensity index), its final "failure risk score" will be pushed to an extremely high level.
[0034] Finally, a failure determination is made. The failure determination method of this invention employs both dynamic thresholds and physical hard boundaries to ensure extremely high reliability. Regarding the dynamic threshold, the BMS collects the failure risk scores of all online chargers in the current parallel system, forming a score distribution, and calculates the upper quantile of this distribution in real time as the "dynamic failure determination threshold." Simultaneously, regarding the physical hard boundaries, the system checks whether the aforementioned calculated degradation characteristics have exceeded their independently preset "absolute safety physical boundaries," such as an eye diagram attenuation coefficient falling below a certain absolute lower limit that would lead to complete communication failure. Ultimately, if the failure risk score of the target charger exceeds the dynamic threshold, or any degradation characteristic exceeds its absolute safety physical boundary, the system ultimately determines that the target charger has failed.
[0035] If step S5 determines that the charger has failed, the BMS will immediately stop sending current-increase commands to the failed charger and may issue an alarm or attempt to safely disconnect it, thus ending the entire adjustment process. If the charger has not failed, the process unconditionally returns to step S2 to continue monitoring and adjustment for the next cycle, forming a complete, safe, and intelligent closed-loop negative feedback control system.
[0036] In another preferred embodiment of the invention, it further includes: The output current time series of each charger within a preset sliding time window is collected, and the scalar current sequence of each charger is reconstructed into a current state point cloud in a high-dimensional phase space using the time delay coordinate embedding method. The embedding dimension and time delay are determined according to the rated switching frequency of the charger. For each charger's current state point cloud, its continuous homology is calculated. Specifically, the Vietoris-Rips simplicial complex manifold corresponding to the point cloud is constructed, and the one-dimensional homology group of the Vietoris-Rips simplicial complex manifold is calculated from its appearance to its extinction to generate the corresponding continuous barcode image. Each horizontal line segment in the image represents a one-dimensional homology class, with its starting scale being the birth scale and its ending scale being the death scale. In the continuous barcode image, the horizontal line segment corresponding to the H1-dimensional homology class where the difference between the birth scale and the death scale is greater than the preset divergence persistence threshold is determined as the H1 bar representing the current divergence, and the line segment length of the H1 bar is extracted as the divergence persistence length; if multiple H1 bars appear in the continuous barcode image of a certain charger, the maximum line segment length among them is taken as the divergence persistence length of the charger. The branching duration of each charger is input into a preset nonlinear attenuation mapper. This mapper outputs a current-limiting percentage attenuation factor with the branching duration as the independent variable. The nonlinear attenuation mapper has the following characteristics: when the branching duration is within the range of zero to the preset transition length, the attenuation factor remains zero; when the branching duration exceeds the preset transition length, the attenuation factor increases with the increase of the branching duration in a superlinear trend and gradually approaches the upper limit of saturation attenuation. The attenuation factor is multiplied by the current current-limiting percentage after adjustment in step S4 or kept unchanged to obtain the current-limiting percentage after attenuation. Using the attenuated current limiting percentage of each charger as the final current limiting command, an independent current limiting message is generated and sent to the corresponding charger via the CAN bus for execution. This actively breaks the topology structure of the H1 strip in the current distribution of the parallel system, reducing the H1 strip of the current state point cloud of all chargers to within the preset divergence persistence threshold, thus completing the system topology reshaping.
[0037] Specifically, after the BMS performs step S4 to dynamically adjust the current limiting percentage and before generating and sending the current limiting message, for each online charger, the BMS collects a continuous output current time series within a preset sliding time window. The length of the sliding time window can be set to cover several control cycles to ensure sufficient information about the system's dynamic behavior is captured. To reveal the deep geometric structures in the current signal that cannot be intuitively identified on the time-domain waveform, this embodiment uses a time-delay coordinate embedding method to reconstruct the scalar current sequence of a single charger from a one-dimensional time axis to a high-dimensional phase space, forming a discrete point set called a "current state point cloud".
[0038] Continuous homology is a technique used in topological data analysis to detect topological invariants in point cloud data at different spatial scales. One-dimensional homology groups directly correspond to "ring-like" or "void-like" structures within the point cloud. The calculation process is as follows: First, based on the current state point cloud of a charger, its Vietoris-Rips simplex complex manifold is constructed step by step. This construction process involves setting a progressively increasing proximity scale parameter. At this scale, if the distance between any two points is less than the scale, an edge is connected between them; if any three points are connected pairwise to form a triangle, a triangular facet is filled; if any four points are connected pairwise to form a tetrahedron, a tetrahedron is filled… and so on, forming a complex structure that continuously "grows" with the scale, composed of simplexes such as points, edges, triangles, and tetrahedrons—that is, the Vietoris-Rips simplex complex manifold. As the scale increases from zero, the complex manifold undergoes a complete evolutionary process from a completely discrete set of points, gradually developing connected components and loop structures, until it finally merges into a completely connected cluster—a process of "from emergence to demise." Throughout this process, the persistent homology algorithm tracks and records the birth and annihilation of each one-dimensional homology class (which can be considered a topological feature of a "hole" or "loop"): when a "loop" first forms in the complex, its corresponding scale is recorded as the "birth scale"; when the "loop" disappears after being completely filled by a subsequent higher-dimensional simplex, its corresponding scale is recorded as the "death scale." All detected one-dimensional homology classes are plotted on a two-dimensional graph according to their birth and death scales, where each horizontal line segment represents a one-dimensional homology class, with the left endpoint's x-coordinate being the birth scale and the right endpoint's x-coordinate being the death scale. This is the so-called "persistent barcode map."
[0039] After generating the persistent barcode image, topological structures that represent stable and persistent divergences within the system are selectively identified. The specific selection criterion is to compare the length of each horizontal line segment in the image (the difference between the death and birth scales) with a preset divergence persistence threshold. Only H1-dimensional homology classes with segment lengths greater than this threshold are considered to represent persistent current divergence patterns indicating true physical meaning, defined as "H1 bars," and their segment lengths are extracted as the "divergence persistence length." If multiple H1 bars meeting this condition appear in the persistent barcode image of a charger, it means that multiple intertwined circulating current patterns exist simultaneously in the phase space trajectory of the charger, indicating a more complex degree of system divergence. In this case, this scheme takes the maximum segment length as the "divergence persistence length" of the charger to represent its most severe degree of divergence.
[0040] To transform this abstract topology metric into executable control commands, this embodiment designs a preset nonlinear attenuation mapper. This mapper maps the divergence duration to a current-limiting percentage attenuation factor, with the core design principle being a "sluggish activation, strong suppression, and safe saturation" control strategy. Specifically, the mapper's characteristics are as follows: when the divergence duration is small, falling within the range of zero to a preset transition length, the system considers this normal point cloud morphology fluctuation, not constituting a topology threat. Therefore, its output attenuation factor is always zero, without triggering any additional current-limiting intervention, thus avoiding misadjustment during normal charging. However, once the divergence duration exceeds this preset transition length, it indicates that a topology divergence has been truly established and begins to threaten system stability. At this point, the attenuation factor increases rapidly with a superlinear trend as the divergence duration further increases, thereby applying strong current suppression to the charger that has generated severe divergence, aiming to quickly disintegrate its internal topology loop structure. Meanwhile, to prevent the charger's output current from dropping to zero and losing its current sharing function due to over-adjustment, the growth of this attenuation factor is not unlimited. It will eventually gradually approach a preset saturation attenuation limit, thereby ensuring that the charger still bears part of the load in the system.
[0041] Finally, the calculated attenuation factor of each charger is multiplied by the adjusted current limiting percentage or the unchanged current limiting percentage obtained after step S4 to obtain the "attenuated current limiting percentage". The BMS uses this as an independent final current limiting command for each charger, generates a unicast current limiting message, and sends it to the corresponding charger for execution via the CAN bus. From a topology perspective, the direct physical effect of this operation is to actively "cut off" the closed-loop trajectory that constitutes the H1 strip in the current state point cloud of the charger that is the source of instability by selectively and nonlinearly reducing the current, thus breaking the topology that maintains the current divergence and causing the point cloud shape of the charger to be topologically reshaped. The ideal convergence result of this process is that the H1 strip of the current state point cloud of all chargers is reduced to within the preset divergence persistence threshold, that is, there is no longer any persistent and stable topological circulating current in the system. Thus, with minimal current regulation cost, a qualitative transformation is achieved in the parallel system from "superficial balance but internal topological divergence" to "true current sharing at the topology level", improving the long-term operational reliability of the system.
[0042] In another preferred embodiment of the invention, it further includes: The total parallel current signal, formed by the superposition of the output currents of each charger in the parallel system, is collected. Symmetric geometric mode decomposition is performed on the total parallel current signal to obtain multiple symmetric geometric mode components arranged from high to low symmetric geometric energy. The symmetric geometric entropy of each symmetric geometric mode component is calculated. The symmetric geometric entropy is a measure of the concentration of the symmetric spectrum of the symmetric geometric mode component at each symmetric frequency point. Symmetric geometric mode components whose symmetric geometric entropy exceeds a preset nonlinear modal entropy threshold are marked as nonlinear modal components. For each labeled nonlinear mode component, its symplectic structure features are extracted; the symplectic structure features include the dominant symplectic frequency of the nonlinear mode component, the symplectic geometric phase corresponding to the dominant symplectic frequency, and the symplectic geometric spectral energy corresponding to the nonlinear mode component in the symplectic geometric mode decomposition. The extracted symplectic structural features are associated with the charger identifier that generates the nonlinear modal component to form a nonlinear modal feature tuple; Monitor the symplectic geometric entropy in each nonlinear modal feature tuple. When the symplectic geometric entropy of any nonlinear modal feature tuple exceeds the preset modal instability boundary, determine that the charger corresponding to the nonlinear modal feature tuple is a modally unstable charger. Immediately freeze the current current limiting percentage of the modally unstable charger and extract the symplectic structural features in the nonlinear modal feature tuple associated with the modally unstable charger to form the unstable mode symplectic structural feature package to be canceled. The unstable mode symmetric structure feature packet is broadcast to all other normal chargers in the parallel system via the CAN bus; the unstable mode symmetric structure feature packet contains the dominant symmetric frequency, symmetric geometric phase, and symmetric geometric spectrum energy; the normal chargers that receive the unstable mode symmetric structure feature packet calculate the anti-phase symmetric geometric phase based on the difference between the symmetric geometric phase of their current output current and the symmetric geometric phase in the unstable mode symmetric structure feature packet, so that the anti-phase symmetric geometric phase differs from the symmetric geometric phase in the unstable mode symmetric structure feature packet by half a symmetric period; Each normal charger uses the anti-phase symplectic geometric phase as the target phase, the dominant symplectic frequency in the unstable mode symplectic structural feature packet as the modulation frequency, and the symplectic geometric spectrum energy in the unstable mode symplectic structural feature packet as the amplitude reference to generate anti-phase symplectic cancellation mode components. The anti-phase symplectic cancellation mode components are then superimposed on their own output current command, thereby actively cancelling the unstable nonlinear modes generated by the mode unstable charger at the level of the total parallel current.
[0043] It should be noted that in the complex electromagnetic and thermal coupling environment of multiple chargers operating in parallel, even if the main regulating circuit has stabilized the average total current near the target, individual chargers may still generate nonlinear current oscillation modes with a defined symplectic geometric structure due to internal power device parameter drift, insufficient feedback loop phase margin, or external disturbances. These modes are not random noise, but rather structured disturbances carrying specific frequencies, phases, and energy. Once formed, they are difficult to eliminate by conventional proportional or step current limiting regulation. Instead, they create nonlinear coupling between chargers through a shared DC bus, resulting in a continuous, intermittent oscillating envelope in the total parallel current, exacerbating device stress and inducing protection malfunctions. Therefore, in this embodiment, the total parallel current signal formed by the natural superposition of the output currents of each charger on the DC bus is first acquired. The total current signal is subjected to symplectic geometric mode decomposition (SMD), an adaptive signal processing method based on the symplectic geometric structure of Hamiltonian systems. SMD decomposes complex nonlinear signals into multiple symplectic geometric mode components arranged from high to low energy according to their inherent symplectic geometric energy. Each component is essentially an eigenmode function with a definite symplectic frequency and symplectic geometric phase. After obtaining each component, its symplectic geometric entropy is calculated. The symplectic geometric entropy measures the degree of concentration of the component's distribution in the symplectic frequency domain: if the symplectic spectral energy of a component is highly concentrated at a few symplectic frequency points, its symplectic geometric entropy is low, representing a relatively regular deterministic mode; conversely, if the symplectic spectral energy is dispersed, the symplectic geometric entropy is high. When the symplectic geometric entropy of a component exceeds a preset nonlinear mode entropy threshold, it indicates that the component has deviated from the regular mode category and exhibits chaotic or complex nonlinear behavior characteristics, thus being labeled as a "nonlinear mode component."
[0044] For each labeled nonlinear modal component, the BMS extracts three elements to form its symplectic structural features: the dominant symplectic frequency of the component, the symplectic geometric phase corresponding to the dominant symplectic frequency, and the symplectic geometric spectral energy of the component in the symplectic geometric mode decomposition result. The extracted symplectic structural features are bound to the charger identifier that generated the component, forming a traceable nonlinear modal feature tuple. Subsequently, the BMS continuously monitors the symplectic geometric entropy in each nonlinear modal feature tuple. When the symplectic geometric entropy of any feature tuple exceeds the preset modal instability boundary, it means that the nonlinear oscillation of the charger has developed from a weakly chaotic state to a severe unstable state, posing a direct threat to the system. At this time, the BMS immediately determines that the charger is a modally unstable charger and freezes its current limiting percentage to prevent further deterioration. Simultaneously, symplectic structural features are extracted from the nonlinear modal feature tuples associated with the modally unstable charger, packaged into an unstable mode symplectic structural feature package to be canceled, which contains a clearly defined dominant symplectic frequency, symplectic geometric phase, and symplectic geometric spectral energy.
[0045] The unstable mode symmetric structure feature packet is then broadcast via the CAN bus to all other chargers in the parallel system that are determined to be normal. Each normal charger receiving this feature packet first reads the symmetric geometric phase of its current output current and calculates the difference between it and the unstable mode symmetric geometric phase in the feature packet. Based on this, an anti-phase symmetric geometric phase is calculated, ensuring that this anti-phase phase is precisely half a symmetric period different from the unstable mode's symmetric geometric phase, achieving complete phase alignment. Each normal charger then uses this anti-phase symmetric geometric phase as the target phase, the dominant symmetric frequency in the feature packet as the modulation frequency, and the symmetric geometric spectrum energy as the amplitude reference to generate an anti-phase symmetric cancelling mode component internally, and directly superimposes this component into its own output current command. Since the anti-phase cancelling components injected collaboratively by all normal chargers form a vector sum at the level of the total parallel current, its effect is equivalent to introducing a compensation wave with the same amplitude and frequency as the unstable nonlinear mode but opposite phase onto the DC bus, thereby physically and actively canceling the harmful nonlinear oscillations excited by the mode-unstable charger. This process fundamentally dismantles the energy loop that sustains the nonlinear modes, allowing the total parallel current to stabilize and the system to return to a safe operating range.
[0046] In another preferred embodiment of the invention, it further includes: S101. Initialize the quantum particle swarm: Map the current limiting percentage of each charger in the parallel system to a quantum individual. Each quantum individual is described by its quantum pose probability amplitude and quantum pose phase angle. The square of the modulus of the quantum pose probability amplitude represents the probability that the charger takes a certain current limiting percentage value in the feasible solution space. The quantum pose phase angle represents the search bias of the current limiting percentage of the charger in the solution space. Taking the current limiting percentage adjusted in step S4 as the center, Gaussian sampling initializes the quantum pose probability amplitude of each quantum individual. At the same time, the quantum pose phase angle is uniformly and randomly initialized in the range of zero to two π. S102. For each quantum individual, its quantum pose probability amplitude is mapped to a real-value current limiting percentage through a probability amplitude decoder. The probability amplitude decoder linearly maps the quantum pose probability amplitude within a preset lower limit percentage to 100%. Based on the real-value current limiting percentages of all chargers, the two objective function values of parallel circulating current loss and bus voltage fluctuation are calculated. Parallel circulating current loss is characterized by the sum of squares of the deviations between the output currents of each charger, and bus voltage fluctuation is characterized by the ratio of the peak-to-peak value of the bus voltage to the average value of the bus voltage within the current control cycle. S103. Based on the two objective function values of all quantum individuals in the current iteration, construct the Pareto non-dominated solution set of the joint objective; combine the real-valued current limiting percentage, parallel circulating current loss value and bus voltage fluctuation value corresponding to the same quantum individual in the current iteration into a solution vector. If the solution vector is not dominated by any existing solution vector in the Pareto non-dominated solution set, then include it in the Pareto non-dominated solution set, and remove the existing solution vectors in the Pareto non-dominated solution set that are dominated by the solution vector. S104. Select the solution vector with the smallest crowding distance from the Pareto non-dominated solution set as the global attractor. Convert the real-valued current limiting percentage corresponding to the global attractor into the global attractor probability amplitude through the inverse probability amplitude encoder. Update the quantum pose probability amplitude of each quantum individual using the quantum gate rotation operator. The rotation angle is determined by the difference between the current quantum pose phase angle of the quantum individual and the target phase angle corresponding to the global attractor probability amplitude. After the update, apply a random perturbation to the quantum pose phase angle of each quantum individual to avoid the search getting trapped in a local minimum. S105. Repeat steps S102 to S104 until the iteration converges. Select the solution vector that best balances the parallel circulating current loss and bus voltage fluctuation from the final Pareto non-dominated solution set. Extract the real-value current limiting percentage of each charger and use it as the independent current limiting percentage of each charger to generate an independent current limiting message. Send the message to the corresponding charger for execution via the CAN bus.
[0047] It should be noted that once the main control loop stabilizes the total charging current near the target value, how to rationally distribute this total amount to each charger to minimize circulating current at the system level and achieve the most stable bus voltage is a deeper challenge. In practice, due to the unavoidable dispersion in the output characteristics, cable voltage drop, and connector contact resistance of each charger, simply issuing the same current limiting percentage command to all chargers often triggers parallel circulating current. This not only causes ineffective power flow between chargers and reduces system efficiency, but also exacerbates voltage fluctuations on the DC bus, affecting the power supply quality of other sensitive loads. This is a typical multi-objective optimization problem, and its solution space grows exponentially with the number of chargers. Traditional rule-based or linear programming methods are difficult to find satisfactory compromise solutions within the real-time control cycle. Therefore, this scheme first abstracts the current limiting percentage of each charger to be allocated in the parallel system as a quantum individual. In the quantum particle swarm optimization algorithm, a quantum individual is a quantized representation of a candidate solution. It is no longer a fixed numerical value but is described by two core parameters: the quantum pose probability amplitude, whose squared magnitude represents the probability that the charger will tend to adopt a specific current limiting percentage at the current search moment; and the quantum pose phase angle, which determines the charger's search direction and oscillation mode. During initialization, the quantum pose probability amplitude of each quantum individual is Gaussian sampled around the adjusted current limiting percentage determined in step S4, ensuring that the initial search range is concentrated near the output of the main control loop, guaranteeing the physical rationality of the search and its convergence speed. Simultaneously, the quantum pose phase angle is uniformly and randomly initialized within the range of zero to two π to provide sufficient diversity to the search process.
[0048] In each round of the iterative search, probability amplitude decoding is performed on each quantum individual. The abstract quantum pose probability amplitude is converted into a specific real-valued current limiting percentage within a preset lower limit percentage to 100% range through linear mapping, thus restoring it to an executable physical instruction. Based on the current real-valued current limiting percentage allocation scheme of all chargers, two key objective function values are calculated: parallel circulating current loss, quantified by the sum of squared deviations between the output currents of each charger; and bus voltage fluctuation, measured by the ratio of the peak-to-peak value of the DC bus voltage to the average voltage within the current control cycle. These two objective functions evaluate the merits of the current allocation scheme from two dimensions: energy loss and power quality.
[0049] Next, BMS constructs a Pareto-nondominated solution set for the joint objective. Pareto nondomination refers to the condition in multi-objective optimization where solution vector A is not inferior to solution vector B on all objectives, and is strictly superior to B on at least one objective; in this case, A is said to dominate B. The solution vectors corresponding to all quantum individuals in the current iteration (including real-valued current limiting percentage, circulating current loss, and voltage fluctuation) are traversed. Solutions not dominated by any solution in the existing nondominated solution set are included, while old solutions dominated by new solutions are removed. This maintains a frontier where each solution represents a compromise that cannot improve one objective without worsening another.
[0050] To guide the entire quantum particle swarm towards the Pareto front, the BMS selects the solution with the smallest crowding distance from the current non-dominated solution set as the global attractor. Crowding distance is an indicator of the sparseness of solution distribution in the target space; selecting the smallest helps explore regions on the front that have not yet been fully explored, preventing premature aggregation. The real-valued current-limiting percentage corresponding to this global attractor is mapped back to the quantum pose probability amplitude space using an inverse probability amplitude encoder. Subsequently, the quantum pose probability amplitude of each quantum individual is updated using a quantum gate rotation operator. The rotation angle is determined by the difference between the individual's current phase angle and the target phase angle corresponding to the global attractor, causing the individual to rotate and move closer to the global attractor in quantum space. After the update, a small random perturbation is applied to the phase angle of each quantum individual to maintain the ability to escape local minima.
[0051] The above steps are iterated repeatedly until the convergence condition is met. Finally, from the stable Pareto non-dominated solution set, the solution vector that achieves the best trade-off between parallel circulating current loss and bus voltage fluctuation is selected, and the real-valued current limiting percentage corresponding to each charger is extracted. The BMS uses this as the independent current limiting command for each charger, generates unicast current limiting messages, and sends them to be executed via the CAN bus, thereby realizing a closed-loop upgrade from unified allocation to independent optimization, reducing system circulating current loss and smoothing bus voltage fluctuation while maintaining the total current target.
[0052] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for regulating the parallel charging current of multiple chargers based on a battery parallel system, applied to a battery management system (BMS), wherein the BMS communicates with multiple chargers via a CAN bus, characterized in that... Includes the following steps: Step S1: When a charger is detected to be connected, the current limiting percentage is set to the initial current limiting percentage, and a current limiting message containing the initial current limiting percentage is generated and sent to the charger. Step S2: Determine whether the battery communication in the parallel system is disconnected or whether the battery is in a protection or recovery state; if yes, reset the current limiting percentage to the initial current limiting percentage, send the corresponding current limiting message, and then repeat this step; if no, proceed to step S3. Step S3: Obtain the maximum battery current of the parallel system and calculate the absolute value of the difference between the maximum battery current and the preset target current; if the absolute value of the difference is not greater than the preset deviation threshold, maintain the current current limiting percentage unchanged and jump to step S5; if it is greater, execute step S4. Step S4: Compare the maximum battery current with the preset target current. If the maximum current of the battery is greater than the preset target current, then when the current current limiting percentage is greater than the preset lower limit percentage, the current current limiting percentage will be reduced by a preset adjustment step size; otherwise, the current current limiting percentage will remain unchanged. If the maximum current of the battery is less than the preset target current, then when the current current limiting percentage is less than 100%, the current current limiting percentage will be increased by the preset adjustment step size; otherwise, the current current limiting percentage will remain unchanged. A current-limiting message is generated based on the adjusted current-limiting percentage and sent to the charger; Step S5: Determine if the charger is faulty. If so, end the adjustment; otherwise, return to step S2.
2. The method according to claim 1, characterized in that, The preset target current is 11A, and the preset deviation threshold is 3A.
3. The method according to claim 1, characterized in that, The initial current limiting percentage is 25%, the preset lower limit percentage is 25%, and the preset adjustment step size is 25%.
4. The method according to claim 1, characterized in that, The maximum current of the battery is the maximum value among the charging currents of all batteries in the parallel system.
5. The method according to claim 1, characterized in that, The current limiting message is a CAN communication message used to instruct the charger to limit its output current to the value corresponding to the current limiting percentage of the charger's rated current.
6. The method according to claim 1, characterized in that, The BMS obtains the output current of each charger through CAN communication to determine the maximum current of the battery, and the charger does not need to be equipped with a hardware current detection circuit for charging current measurement.
7. The method according to claim 1, characterized in that, Determining whether a charger is malfunctioning includes: During the initial power-on self-test of the parallel system, a set of synthetic samples with clear failure labels is generated based on the characteristic parameter boundaries of various typical charger failure modes pre-stored in the BMS; the synthetic samples are then mixed with the historical feature sample set collected under historical normal charging conditions to form a hybrid training set. The isolated forest is trained using the hybrid training set, thereby constructing a decision tree structure that can distinguish between normal sample clusters and various known failed sample clusters in the initial stage of the model through random cutting. From the continuous CAN messages uploaded by the target charger cycle, extract multiple frames of data within a preset time window, analyze and calculate the cumulative degradation characteristics of electrical stress, including at least: the impulse density of pulse current spikes per unit time, the thermal stress cycle factor based on the joint weighting of shell temperature and output current, and the eye diagram opening attenuation coefficient reflecting the signal quality of the communication physical layer; to form the degradation feature vector to be detected at the current moment. The degraded feature vector to be detected is input into the isolated forest model, which traverses each decision tree and records the number of path edges it traverses in each decision tree to be completely isolated as the isolation depth. At the same time, during the path search process, the proportion of synthetic samples included in the branches that the corresponding sample falls into at the split node is counted, thereby assigning a failure tendency index to the corresponding path. After traversal, the arithmetic mean of the isolation depths of the degenerate feature vector to be detected under all decision trees is calculated to obtain the average isolation depth, and the arithmetic mean of the failure tendency index on all paths is calculated to obtain the average failure tendency index. The average isolation depth is mapped to a basic anomaly score through a normalization factor, and then the average failure tendency index is used as a weight to correct the anomaly score to obtain the failure risk score. The failure risk score is compared with a dynamic failure decision threshold, which is the upper quantile calculated in real time based on the failure risk score distribution of all chargers in the current parallel system within the same time window. At the same time, it is determined whether each feature in the electrical stress cumulative degradation feature vector exceeds its respective preset absolute safety physical boundary. If the failure risk score exceeds the dynamic failure decision threshold, or any one of the electrical stress cumulative degradation features exceeds its absolute safety physical boundary, then the target charger is finally determined to have failed.
8. The method according to claim 1, characterized in that, After the rate limiting percentage remains unchanged or is adjusted, and before sending the message, the following steps are also included: The system collects time-series data from the historical charging process of the parallel charging system, forming three sets of parallel data streams: the first data stream is a sequence of current limiting percentage instructions with control cycle granularity; the second data stream is an instantaneous frequency sequence extracted from the output current values of each charger at the corresponding moment through Hilbert-Huang transform; and the third data stream is a sequence of load mutation event labels recorded by the battery management system at the corresponding moment. The three sets of parallel data streams are aligned by timestamps to construct a multi-stream input sample set, which is used to train a convolutional long short-term memory network model that incorporates attention gating. Obtain the current limiting percentage instructions of the current control cycle and several previous control cycles, the instantaneous frequency of the output current value of each charger extracted by Hilbert-Huang transform, and the load change event label, and generate the current multi-stream input feature vector according to the structure of the multi-stream input sample set; The current multi-stream input feature vector is input into the trained convolutional long short-term memory network model with fusion attention gating. After model inference, the predicted value of the current spike in the next control cycle is output. The predicted value of the current spike is the overshoot of the instantaneous current peak that may occur in the next control cycle relative to the current steady-state current value. Obtain a preset current spike tolerance limit, calculate the difference between the current spike prediction value and the current spike tolerance limit, and use the difference as the feedforward compensation amount; while maintaining the adjusted current limiting percentage obtained in step S4 or the current limiting percentage that remains unchanged, subtract the current limiting percentage correction value corresponding to the feedforward compensation amount to obtain the current limiting percentage after feedforward correction. The current limiting percentage after feedforward correction is subjected to boundary limiting processing with the preset lower limit percentage and 100%. If the current limiting percentage after feedforward correction is lower than the preset lower limit percentage, the preset lower limit percentage is used; if it is higher than 100%, 100% is used. A current limiting message is generated based on the current limiting percentage after limiting processing and sent to the charger.
