Battery dynamic balancing method and system based on operating state

Abstract

This invention discloses a battery dynamic balancing method and system based on operating status. The battery dynamic balancing method includes: determining the cell digital fingerprint of each cell based on the collected operating condition data of each cell and the electrochemical parameters of the cells identified online; predicting the balancing current sequence and balancing switch state matrix sequence of the balancing branch within a preset period based on the cell digital fingerprint and a multi-objective dynamic cost function; and dynamically balancing the battery based on the balancing current sequence and balancing switch state matrix sequence; the cost function is determined by weighted summation based on the difference between the current cell SOC and the average SOC, the balancing energy loss, the predicted temperature rise during balancing, the aging factor penalty term, and the balancing time required, with weight coefficients dynamically determined according to different operating conditions. This invention achieves dynamic battery balancing.

Description

Technical Field

[0001] This invention relates to the field of battery equalization technology, and in particular to a battery dynamic equalization method and system based on operating status. Background Technology

[0002] Traditional equalization technologies face significant challenges in electric vehicle high-voltage battery packs, primarily due to issues such as trigger lag, energy waste, and lack of health management, leading to ineffective energy transfer and even accelerated aging. Current technologies, typically driven by fixed rules under complex real-world vehicle operating environments, struggle to achieve a globally optimal dynamic balance between energy transfer efficiency, equalization speed, battery life, and system safety throughout the battery's entire lifecycle and operating conditions. Furthermore, existing technologies have failed to address the issue of when equalization should be initiated.

[0003] Publication No. CN121036261A proposes a battery equalization control system that constructs a comprehensive consistency index to dynamically trigger equalization demand. This solves the problems of slow equalization speed, low efficiency, and strong dependence on large data samples in existing technologies. However, it has shortcomings: 1) The comprehensive consistency index lacks mechanistic support: the index mainly relies on conventional state variables such as voltage and SOC, without explicitly modeling the internal aging mechanisms of the cells (such as lithium plating and SEI growth), making it difficult to accurately reflect equalization priority during the accelerated aging stage of the battery. 2) Dimensionality reduction processing loses key physical information: the dimensionality reduction method used to reduce the complexity of high-dimensional states may filter out details such as transient polarization differences and local thermal coupling between cells, failing to accurately capture the dynamic evolution of inconsistencies under fast charging or pulse conditions. 3) No closed-loop learning and strategy iteration are formed: the system only performs open-loop or semi-open-loop rolling optimization, lacking the ability to continuously correct and learn model parameters and optimization weights online based on the actual equalization execution effect (dispersion changes, temperature rise, energy consumption).

[0004] Publication No. CN114865747A proposes an online balancing method for the state of charge (SOC) power scheduling of energy storage battery clusters. This invention determines the charge / discharge state of the energy storage battery clusters based on power scheduling commands and estimates the online balancing time in real time using the SOC of each cluster, its charge / discharge limit power, and the power scheduling command value. This invention considers the charge / discharge limit power of the battery clusters, enabling online balancing of the SOC during power scheduling. However, it has shortcomings: 1) It relies on a single SOC index, ignoring multi-dimensional states: Power is allocated only based on the SOC of each battery cluster, without considering key factors such as internal resistance, capacity decay rate, and temperature. When SOCs are similar but internal resistance increases significantly, overheating or accelerated aging is likely. 2) The balancing granularity is coarse, failing to address inconsistencies within the cluster: The method focuses on power scheduling at the battery cluster level and cannot reach the cell or module level. When there is severe dispersion between cells within a cluster, this method is ineffective. 3) Lack of optimization objectives for balanced energy consumption and system efficiency: Only pursuing rapid SOC convergence without establishing a cost function for balanced energy consumption or overall system efficiency may result in a large amount of energy flowing between clusters and being wasted by power converter losses.

[0005] Publication No. CN114547379A proposes a dynamic flow graph online equilibrium partitioning method and storage medium based on game strategy. The method sets the target partition block obtained in the previous time step as a game participant; for all game participants, a strategy set is set for each participant based on all possible actions of the newly added vertex; a dynamic flow graph online equilibrium partitioning game model is constructed based on equilibrium degree (ED) and modularity (MD); the utility value of each participant under all strategies is calculated using the dynamic flow graph online equilibrium partitioning game model; the online equilibrium partitioning model is solved to obtain the Nash equilibrium solution; the target block is updated according to the strategy of each participant corresponding to the Nash equilibrium solution; and the online equilibrium partitioning of the dynamic flow graph is completed. However, it has shortcomings: 1) It cannot output the physical instructions required for battery balancing: the Nash equilibrium solution provides a static or semi-static partitioning scheme and cannot generate actual execution instructions of the battery system such as balancing current, duty cycle, or switching sequence. 2) It completely ignores battery physical constraints and safety boundaries: it does not consider any battery-specific safety operating constraints such as voltage, current, temperature, and topology, and direct application will cause serious safety risks.

[0006] As can be seen from the above, under current technology, the industry typically determines the start and stop of battery equalization by judging whether the voltage difference exceeds a threshold. However, the strategy of starting equalization based on the threshold is rather vague, resulting in equalization starting either too early or too late. Starting equalization too early will lead to energy waste; starting equalization too late will result in an excessively large battery voltage difference, thus accelerating aging. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings and defects of existing technologies and provide a battery dynamic balancing method and system based on operating status that can comprehensively consider multiple factors to determine when balancing should be initiated. This invention is based on an online balancing method for high-voltage battery packs using battery pack operating status perception and adaptive strategy updates. It upgrades battery balancing from traditional passive voltage / capacity matching to an intelligent energy scheduling problem that integrates online estimation of electrochemical state and dynamic optimization decision-making. Furthermore, it comprehensively considers the beneficial and potential impacts of balancing throughout the battery's entire life cycle and across all operating conditions to provide the optimal balancing scheme.

[0008] According to one aspect of this application, a battery dynamic balancing method based on operating state is provided, comprising:

[0009] Based on the collected operating condition data of each battery cell and the electrochemical parameters of the battery cells identified online, the digital fingerprint of each battery cell is determined.

[0010] Within a preset period, based on the cell digital fingerprint and multi-objective dynamic cost function, the equalization current and equalization switch state of the equalization branch in the future time window are predicted, and the battery is dynamically equalized according to the equalization current sequence and the equalization switch state matrix sequence; the cost function is determined based on the difference between the current cell SOC and the average SOC, the equalization energy loss, the temperature rise predicted by equalization, the aging factor penalty term, and the time required for equalization.

[0011] Preferably, the expression for the cost function is as follows:

[0012] ;

[0013] In the formula, Represents the cost function, Indicates the current cell SOC. Indicates the average SOC of the battery cell. Indicates the equalization current. Indicates the equalization resistance. This represents the temperature rise predicted in equilibrium. This indicates a penalty for aging factors. This indicates the time required for equilibrium to be reached. , , , , Indicates the weighting coefficient. Indicates the number of battery cells. This indicates the cell number.

[0014] Preferably, the cell digital fingerprint includes at least the cell's physical serial number, real-time terminal voltage, real-time temperature, open-circuit voltage, estimated SOC value, ohmic internal resistance, polarization internal resistance, polarization time constant, current capacity health (SOH), capacity decay rate, and aging stress index; preferably, the current capacity health is determined based on the capacity decay, which is determined by weighted summation of the cell's cumulative throughput and usage time.

[0015] Preferably, the prediction of the equalization current sequence and equalization switch state matrix sequence of the equalization branch in the future time window is performed under preset constraints. The preset constraints include current constraints, temperature constraints, voltage constraints, and topological constraints. The current constraints include that the equalization current is not greater than the upper limit of the hardware equalization current. The temperature constraints include that the sum of the current temperature of the cell and the corresponding temperature rise is not greater than the safe temperature of the cell. The voltage constraints include that the cell voltage is not greater than the maximum charging voltage during charging and not less than the minimum discharging voltage during discharging. The topological constraints include the combination of equalization switch states determined by the equalization hardware, including constraints on energy transfer between adjacent cells or constraints on preventing the formation of short-circuit loops after the equalization switches are turned on.

[0016] Preferably, the prediction of the balanced current sequence and balanced switch state matrix sequence of the balanced branch in the future time window is processed using an interior-point solver or a pre-trained deep neural network as an approximate optimizer, including:

[0017] Construct the Lagrange function based on the cost function:

[0018] ;

[0019] N is the number of cells to be balanced. Given an N-dimensional Boolean matrix, stored using adjacency list compression encoding. The number of discrete points is determined by the prediction time domain and the discrete time step. Representing variables The cost function, These are optimization variables, including the equalization current sequence and the equalization switch state; These are slack variables, ensuring that the iteration point is always within the feasible region; It is a positive scalar, controlling the barrier function term. The strength; It is a Lagrange multiplier. This means relaxing the equality constraints into the cost function. Represents inequality constraints, limiting The feasible domain includes current, temperature, and voltage constraints;

[0020] Solve the problem using Newton's method. The corrected equation is used to calculate the search direction;

[0021] Iterative updates are performed based on the search direction to reduce variables. Once the accuracy requirements are met, the optimal equalization current sequence and equalization switch state within the future time window are output through a rolling solution process.

[0022] Preferably, the aging factor penalty term is calculated online using the electrochemical model SPMe. The SPMe model takes the current cell digital fingerprint and predicted equilibrium current and temperature as inputs, outputs preset indicators for the cell within a future time window, and integrates these as a lifetime penalty term. The preset indicators include at least a lithium plating risk index and an increase in SEI film thickness. Preferably, the calculation model for the aging factor penalty term includes:

[0023] ;

[0024] In the formula, Indicates battery cell At discrete time The lithium plating risk index, Indicates battery cell No. The thickness increment of the SEI film at discrete time points; , Indicates non-negative weight coefficients. This represents the discrete time step.

[0025] Preferably, the dynamic balancing of the battery based on the balancing current sequence and the balancing switch state matrix sequence includes:

[0026] The initial duty cycle is determined based on the predicted equalization current of the equalization branch in the future time window, the upper limit of the hardware equalization current, and the steady-state gain of the equalization DC / DC converter. A PID controller is used to adjust the initial duty cycle to obtain the target duty cycle by using the actual sampled equalization current as feedback. The PWM signal is determined based on the target duty cycle, and the equalization current of the equalization branch is controlled by the action of the equalization switch on the equalization branch driven by the PWM signal.

[0027] Preferably, the dynamic balancing of the battery based on the balancing current sequence and the balancing switch state matrix sequence includes:

[0028] During the balancing operation, the target parameters are monitored in real time to determine whether they exceed the preset boundary. If they do, the balancing operation is terminated. During the balancing process, faults are identified based on balancing anomalies, the fault tree experience database is queried to determine the fault type, and the upper limit of the balancing current is determined based on the fault type. The target parameters include at least temperature rise data, real-time current, voltage, and temperature.

[0029] Preferably, after each balancing cycle, the target data is stored locally and / or in the cloud for offline learning and updating of the model weight index. The target data includes the cell's digital fingerprint, operating condition indicator, SOC distribution histogram, and balancing current sequence for each balancing cycle, change in SOC standard deviation after balancing, balancing energy loss and maximum temperature rise, and cell capacity decay dispersion.

[0030] According to another aspect of this application, a battery dynamic balancing system is provided, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the battery dynamic balancing method based on the operating state as described.

[0031] This invention, based on the digital fingerprint of the battery cell obtained through online identification and minimizing the dynamic multi-objective cost function, determines the equalization current sequence and equalization switching state matrix sequence. It performs preventative equalization from a mechanistic perspective, effectively intervening in aging at its root, and achieving a shift from equalizing surface states (SOC, voltage difference) to intervening in the root causes of aging. This realizes mechanistic-level preventative equalization. Furthermore, in the cost function-based equalization decision, the goal is to maximize the overall system lifespan, comprehensively considering the long-term aging consequences of battery use from multiple factors, achieving adaptive equalization decisions that take into account long-term aging consequences, thus achieving long-term optimization. In addition, a self-learning library is established based on the data from each equalization operation, enabling feedback optimization based on actual results. This transforms the fixed equalization system into a self-learning and optimizing intelligent battery health management agent, realizing a smart battery system with self-learning capabilities. Attached Figure Description

[0032] Figure 1 This is a flowchart of the battery dynamic balancing method based on operating state according to the present invention. Detailed Implementation

[0033] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0034] This invention constructs a closed-loop system of "perception-decision-execution-learning". In the perception layer, the battery management system (BMS) identifies the multi-dimensional digital fingerprint of each battery cell online through high-frequency sensors and embedded algorithms. In the decision layer, the BMS integrates the global state of the battery pack and the operating conditions of the vehicle to solve a dynamic multi-objective optimization problem and generate the optimal balancing strategy. In the execution and learning layers, the balancing command is precisely executed through a closed-loop control circuit with safety monitoring, and the execution effect data is fed back to the strategy knowledge base to realize the online learning and continuous evolution of the system strategy.

[0035] See Figure 1 As shown in the embodiment of this application, the battery dynamic balancing method based on operating state includes:

[0036] Based on the collected operating condition data of each battery cell and the electrochemical parameters of the battery cells identified online, the digital fingerprint of each battery cell is determined.

[0037] Within a preset period, based on the cell digital fingerprint and minimizing the multi-objective dynamic cost function, the equalization current sequence and equalization switch state matrix sequence of the equalization branch in the future time window are predicted, and the battery is dynamically equalized according to the equalization current sequence and equalization switch state matrix sequence; the cost function is determined based on the difference between the current cell SOC and the average SOC, the equalization energy loss, the temperature rise predicted by equalization, the aging factor penalty term, and the time required for equalization.

[0038] According to embodiments of this application, in order to monitor the electrochemical parameters of each cell online, a sensing layer of this application is pre-constructed to monitor and sense the electrochemical parameters of each cell, so as to monitor and sense the multi-dimensional battery operating status, aiming to achieve in-depth quantitative diagnosis of the root causes of cell inconsistencies. For example, a high-precision synchronous sampling circuit is set up for temperature sampling. For instance, each battery module in the battery pack is configured with a slave controller whose ADC (analog-to-digital converter) supports synchronous sampling of the voltage of all series-connected cells (±1mV accuracy) and at least two temperature points, with a synchronization timing error of less than 1 microsecond, to capture the true differences in the voltage response of each cell during current surges. Alternatively, a high-frequency current sensing device is set up, with Hall current sensors with a bandwidth of not less than 10kHz configured at the total positive / total negative terminals of the battery pack to capture charging and discharging pulse current signals. The sampled values ​​are aligned with the voltage sampled values ​​of each slave controller via high-speed CAN FD or daisy-chain communication at the millisecond level.

[0039] After obtaining the operating status data of each cell from the sensing layer, online identification of electrochemical parameters can be performed based on a pre-established embedded online electrochemical parameter identification algorithm. This includes identifying parameters such as ohmic polarization resistance, polarization internal resistance, and time constant in the equivalent circuit model, as well as predicting the capacity decay trend of the cell to determine its capacity health status, capacity decay rate, and aging stress index. The capacity health status of the cell is obtained from its capacity decay, which can be determined by comparing the current capacity with the factory nominal capacity and combining it with the cumulative throughput. A simplified empirical model is applied to determine and use this model to calculate and update the capacity decay slope online.

[0040] ;

[0041] in, This is the capacity decay amount. For cumulative throughput, For usage time, It is a coefficient for cumulative throughput. It is a coefficient for usage time.

[0042] According to embodiments of this application, the cell digital fingerprint includes the cell physical number, real-time terminal voltage, real-time temperature, estimated open-circuit voltage, SOC estimate based on Kalman filtering, online identified ohmic internal resistance, polarization internal resistance, polarization time constant, current capacity health, capacity decay rate, and comprehensive aging stress index.

[0043] According to embodiments of this application, the equilibrium current is determined at the decision-making level through a constructed cost function. This is a multi-objective dynamic optimization decision function based on operating conditions and overall status. It transforms the equilibrium problem into a real-time optimization problem, with a cost function. It is recalculated and minimized in each cycle (e.g., 1 second), including:

[0044] ;

[0045] The cost function described in this application takes into account the average Factors such as the difference in energy levels, energy loss during equilibrium, temperature rise, aging, and equilibrium time. Represents the average The gap. Among them, It is the current battery cell , It is the average of all battery cells . The energy representing the equilibrium loss, of which, It is a current balancing mechanism. It is a balanced resistance. This is the temperature rise predicted by equilibrium. It is a penalty item for aging factors. It is the time required for equilibrium to be reached.

[0046] Among them, in order to be applicable to all working conditions, the coefficients corresponding to each factor are... , , , , It can be dynamically adjusted under different working conditions, and can be optimized based on feedback from the experience base during use.

[0047] For example, consider the driving conditions: Set the consistency weight to the highest level (e.g., 1.0). Set the (time weight) to a low level (e.g., 0.1). (Efficiency weight) should be moderate (e.g., 0.3) to ensure priority is given to power output. Fast charging condition: (Efficiency weight) and The aging weight is significantly increased (e.g., 0.8, 0.7) to optimize energy efficiency and control additional lifespan loss caused by balanced heat production. The weighting of temperature rise is also increased accordingly.

[0048] Stationary or slow charging conditions: The time weight is reduced to a minimum (e.g., 0.05) to allow for long-term fine-grained balancing. (Efficiency weight) remains at a high level (e.g., 0.6).

[0049] According to an embodiment of this application, when the decision layer determines the combination of equalization current and equalization switch state based on the cost function in online optimization, the objective variable of its online optimization is the combination of equalization current and equalization switch state within a future time window. The solution is performed based on preset current, temperature, voltage, and topology constraints. The solution algorithm can use an interior-point solver for fast rolling optimization, including:

[0050] Set optimization variables :

[0051] ;

[0052] By optimizing variables This allows for the optimization of the current sequence and switching matrix state of each equalization branch within a future time window (e.g., the next 60 seconds). , These are the battery systems The equalization current in each cell and The equalization switching state of each cell, i.e., the equalization switching matrix configuration sequence.

[0053] Set the constraints as follows:

[0054] Current constraint: ≤ , To balance the upper limit of current and temperature constraints in the hardware: , No. The temperature of each battery cell It is the first Temperature rise corresponding to each battery cell This refers to the safe temperature of the battery cell. Voltage constraint: During charging, the cell voltage... During discharge, . It is the first The voltage of each battery cell This is the maximum charging voltage limit. This is the minimum discharge voltage limit. Topology constraint: The combination of switching states determined by specific balancing hardware (such as an adjustable topology for adjacent / arbitrary cells).

[0055] Solve using a solution algorithm:

[0056] First, construct the Lagrange function. ;in, It is the objective function. These are the optimization variables (equalization current sequence and equalization switch state). These are slack variables, and their dimension equals the number of inequality constraints. This ensures that the iteration point is always within the feasible region; It is a positive scalar, controlling the barrier function term. The intensity. As iterations proceed... Gradually decreasing it to 0, the influence of the obstacle term disappears, and the solution eventually converges to the true optimal solution of the original problem. When the th... Slack variables of individual cells By moving away from 0, we can ensure that the iteration point is strictly within the feasible region; It is a Lagrange multiplier. The term "relaxes" the equality constraints into the objective function. It is a primitive inequality constraint, restricting... The feasible domain includes constraints such as current, temperature, and voltage;

[0057] Secondly, to solve the KKT system, Newton's method is used in each iteration to solve for the relationship between the KKT system and the KKT system. The corrected equation is used to calculate the search direction; iterative updates are performed, updating variables along the search direction to reduce... This process is repeated until the solution meets the accuracy requirements. Finally, the result is output and the solution is iterated continuously. The solution yields the optimal equalization current sequence and the column order of the equalization switch state matrix for a future time window (e.g., 60 seconds). The current command corresponding to the current moment is sent to the execution layer. After entering the next cycle, the problem is reinitialized with the latest state, and the optimization continues in a rolling manner.

[0058] When embedded computing power is limited, a pre-trained deep neural network can be used as an approximate optimizer, directly mapping the optimization strategy from the input of the perception layer, significantly reducing the online computational load. For example, when embedded computing power is limited, such as when the main controller's clock speed is below 200MHz, a pre-trained deep neural network can be used instead of the interior-point method. The neural network can be implemented using the following structure:

[0059] The network structure has an input layer dimension of 9+3 (9 dimensions of digital fingerprints for all battery cells, plus 3 dimensions of operating condition encoding), a hidden layer of 3 fully connected layers with 256 neurons per layer, using ReLU activation function, and an output layer with linear activation and an output dimension of N×1 (i.e., the recommended equalization current at the current moment). During training, an offline interior-point solver is used to generate a large number of samples covering various operating conditions, aging states, and SOC distributions. The training objective is to minimize the mean squared error and constraint violation penalty. During inference, a single forward computation takes less than 5 milliseconds.

[0060] According to embodiments of this application, when the decision layer performs equilibrium optimization, it simultaneously performs rolling optimization of lifetime prediction, such as calculating an aging acceleration penalty term. This is used for equalization current optimization calculations. The optimization of the aging acceleration penalty term can be calculated online using a simplified, reduced-order electrochemical model (SPMe). This SPMe model takes the current digital fingerprint and predicted equalization current and temperature as inputs, and directly outputs mechanistic indicators such as the lithium plating risk index and SEI film thickness increment for each cell within the future time window, integrating these as the lifetime penalty term. For example, the following formula can be applied:

[0061] ;

[0062] Indicates battery cell At discrete time The lithium plating risk index (reflected by the amount of lithium ion loss output by the SPMe model; the higher the value, the higher the risk of lithium plating). Indicates the first The thickness increment of the SEI film at discrete time points. , This represents a non-negative weighting coefficient used to balance the risk of lithium plating with the contribution of SEI film growth to aging (it needs to be calibrated according to battery characteristics). This represents the discrete time step.

[0063] ;

[0064] Wherein, the molar mass of the SEI film is =0.1 kg / mol, density of SEI film =1.7g / cm 3 , It is Faraday's constant; , These represent the current time and the previous time, respectively. Indicates an infinitesimally small time interval Inside, the amount of charge flowing through a unit area of ​​the electrode due to the SEI film formation reaction.

[0065] As an example, the decision layer of this application may include a main controller (MC) employing a multi-core embedded processor, such as an ARM Cortex-M7 or RISC-V architecture, running a real-time operating system (RTOS). The main controller integrates an interior-point solver or a pre-trained deep neural network (DNN) for inference.

[0066] According to embodiments of this application, by outputting a balanced current sequence and a combination of balanced switch states through the decision layer, the execution layer can perform closed-loop execution and safety monitoring of the balanced operation, ensuring the safe and accurate implementation of the intelligent strategy. Specifically, when performing the balanced operation, precise control of the balanced current can be achieved by predicting and feeding back the balanced current, using PID control to provide feedback on the current, and establishing a closed-loop balanced current control with prediction and feedback.

[0067] In one embodiment, the feedforward control is based on the reference current of the optimized output. Based on the known transfer function model of the balanced DC / DC converter, the initial duty cycle command is calculated. During feedback control, a PID controller is used, based on the actual sampled balanced current. As feedback, the duty cycle is fine-tuned to achieve precise current tracking, and its control cycle matches the switching frequency of the equalization hardware. For example:

[0068] According to the reference current issued by the decision-making level (Current time value) and the transfer function of the DC / DC converter (Identified as a first-order hysteresis model at the factory), calculate the initial duty cycle. :

[0069] ;

[0070] in, For DC / DC steady-state gain, This indicates the upper limit of the device's maximum current.

[0071] Calculate duty cycle increment :

[0072] ;

[0073] in, , , , For parameters,

[0074] The final duty cycle is It is limited to the range [0,1].

[0075] According to the embodiments of this application, when performing equalization operation, multi-level safety and anomaly diagnosis detection is further performed based on a multi-level safety and anomaly diagnosis mechanism. For example, a safety check is performed before executing any equalization command, and predictive safety protection is performed. If the predicted temperature rise exceeds the limit, the original equalization strategy is not executed / stopped. A fault tree experience base is established for abnormal responses, which is used to learn online using the strategy knowledge base, judge the fault type to avoid safety risks, and output the equalization operation strategy determined in this case.

[0076] According to embodiments of this application, at the execution layer, the equalization execution circuit can employ a bidirectional DC / DC converter and an equalization switch matrix topology to support energy transfer between any battery cells. Preferably, each equalization branch is connected in series with a precision sampling resistor (accuracy ±1%) for current feedback.

[0077] According to embodiments of this application, in the learning layer, a fault tree experience base is established, and online learning is conducted based on the strategy knowledge base. This includes recording inputs, outputs, short-term results, and long-term results, including continuously recording inputs such as the cell digital fingerprint and operating conditions corresponding to each decision, outputs of the balancing strategy, changes in standard deviation, energy consumption, and other short-term results, as well as the long-term results of the dispersion of the capacity decay of the group of cells after a period of time. Data is periodically uploaded to the cloud or a big data platform, and historical data is used to retrain and update the aging model parameters in the decision module, optimize the target weight mapping relationship, and even approximate optimize the neural network, achieving offline learning and model updates. Later, the updated model parameters can be incrementally deployed to the vehicle battery management system (BMS) via OTA wireless upgrades, achieving continuous collective evolution of the vehicle's balancing strategy.

[0078] This can be achieved by connecting to the vehicle management system (BMS) via a cloud-based big data platform through a 4G or 5G communication network, storing historical data and performing offline training, and then deploying the updated model parameters to the vehicle controller via OTA wireless upgrade.

[0079] When implementing the equalization method of this application, in one embodiment, under typical driving conditions, the vehicle travels at a constant speed of 60 km / h, and the battery pack SOC is distributed between 42% and 58%, with a standard deviation of 12%. The sensing layer identifies that the internal resistance of cell #3 is relatively high (15% higher than the average), and the SOH is 91%. The decision layer uses the driving condition coefficient. =1.0, =0.3, =0.2, =0.1, =0.1, prediction time domain 60 seconds, convergence within 15 milliseconds using interior point method, output equalization current: energy is transferred from high SOC cells (#1, #5) to low SOC cells (#3, #8), maximum current 1.2A. The execution layer achieves current tracking through PID control, with an actual root mean square error less than 0.05A. At the end of driving, the SOC standard deviation drops to 5%, the equalization energy loss is 23J, and there is no temperature rise alarm. In fast charging mode, when the vehicle is connected to a 400V / 250A fast charging pile, charging starts from 20% SOC. The decision layer identifies the fast charging mode and switches the weighting coefficients accordingly. =0.6, =0.8, =0.7, =0.7, =0.05. The decision-making team predicted that performing high-current balancing would cause the temperature of cell #2 to exceed 55°C, so it automatically reduced its balancing current to 0.3A and increased the duty cycle of the cooling fan. After charging was completed, the aging penalty statistics showed that the SEI increase during this charging process was 18% lower than the default strategy.

[0080] The following detailed description of the present application's solution, in conjunction with specific embodiments, will provide a complete picture.

[0081] I. Synchronous Collection of Multi-Dimensional Operational Status Data

[0082] At the start of each equalization control cycle (e.g., 1 second), the master controller broadcasts a synchronization pulse, triggering all slave controllers to synchronize sampling. Within 50 microseconds of receiving the synchronization pulse, each slave controller completes voltage sampling of all series-connected cells within its module and temperature sampling of at least two temperature points (e.g., positive and negative tabs, and the middle of the cell), and timestamps the sampled data locally. The Hall current sensor continuously acquires the bus current at a sampling rate of 10kHz. The master controller reads the average current value within the most recent 1 millisecond via direct memory access (DMA) and aligns it with the voltage sampling time using a timestamp. Each slave controller uploads voltage and temperature data to the master controller via a daisy chain, with a total transmission delay of less than 2 milliseconds.

[0083] II. Embedded online identification of electrochemical parameters

[0084] The main controller independently performs preset electrochemical parameter identification for each cell, including:

[0085] Equivalent circuit model parameter identification: Based on the first-order RC equivalent circuit model, the ohmic internal resistance is identified online using the recursive least squares (RLS) method. Polarization resistance and time constant A sliding window was used for identification, with a window length of 10 seconds and a forgetting factor of 0.99.

[0086] Capacity degradation trend prediction: Real-time recording of the cumulative throughput of each cell (Unit: Ah) and usage time (Unit: Day), calculate ;coefficient and The initial values ​​were 0.02 (% / Ah) and 0.001 (% / day), respectively, with coefficients... and The current capacity health is determined every 30 days by calibrating using cloud data. .

[0087] III. Constructing a digital fingerprint for battery cells and optimizing the current balancing strategy

[0088] For each cell ( =1,…, ), construct its digital fingerprint For multidimensional vectors, including:

[0089] Physical number Real-time terminal voltage Unit is mV; real-time temperature Unit: °C; Open circuit voltage The SOC estimate is obtained by looking up the current SOC and the OCV-SOC table. The internal resistance was obtained by fusing voltage, current, and temperature data using an extended Kalman filter (EKF); The unit is mΩ; polarization resistance The unit is mΩ, and the polarization time constant is... Unit: s; Current capacity health Capacity decay rate The unit is % / day.

[0090] The digital fingerprint is updated every second and stored in the main controller's circular buffer for use by the decision-making layer. Each control cycle... (e.g., 1 second) The main controller reads the digital fingerprints of all current battery cells and calculates the cost function. Cost function The weighting coefficients are dynamically adjusted according to different working conditions.

[0091] Table 1 is an example coefficient table.

[0092] Driving conditions 1.0 0.3 0.2 0.1 0.1 Fast charging conditions 0.6 0.8 0.7 0.7 0.05 Let it sit / slow charge 0.8 0.6 0.3 0.2 0.05

[0093] The operating condition identification is obtained by the vehicle control unit (VCU) sending a flag bit via the CAN bus. The above coefficient values ​​can be optimized quarterly based on the cloud-based experience library.

[0094] Optimize variables Prediction time domain =60 seconds, discretization step size =1 second, then the number of discrete time points K is: / =60. Current constraint ≤ , It is the upper limit of hardware equalization current; temperature constraint: ≤55℃; Voltage constraint: during charging ≤4.2V, during discharge ≥2.8V; Topology constraints: For adjacent cell balancing topologies, only ≥2.8V is allowed. =1 Energy transfer between cell pairs; For any cell-adjustable topology, the combination constraint that the switching matrix does not generate short-circuit loops is satisfied (pre-stored legal state table).

[0095] Fast rolling optimization is performed using an interior-point method solver based on optimization variables and line conditions. Since the problem is nonlinear and constrained, a primal-dual interior-point method is used for solution: as mentioned earlier, the obstacle Lagrangian function is constructed, and the obstacle parameters are initially set... Set the value to 10, and multiply by 0.1 in each iteration; solve the KKT system. In each Newton iteration, calculate the Grange function. about The gradient and Hessian matrix are used to solve the linear equations to obtain the search direction. Iterative updates, using backtracking search to ensure... >0, decreases after updating the variable. ;when < And the degree of constraint violation is less than Convergence is determined at the time of the test, and the target cell, balancing path, and dynamic balancing current curve are output. The optimal balancing current sequence and balancing switch state matrix sequence at the current time k=1 are sent to the execution layer. The optimization problem is reinitialized with the latest state in the next control cycle.

[0096] Using the SPMe model with the current cell digital fingerprint and predicted equilibrium current and temperature sequences as input, and based on the aging penalty term calculation model, the lithium plating risk index and SEI film thickness increment for each cell within the future time window are directly output for lifetime prediction rolling optimization. Lithium plating risk index This value characterizes the amount of lithium-ion loss; a higher value indicates a higher risk of lithium plating. The formula for calculation is as follows: . The lithium-ion concentration on the surface of the negative electrode. The ratio of the equilibrium concentration of lithium ions on the surface of the negative electrode to that of the negative electrode is defined as the lithium plating risk index.

[0097] IV. Closed-loop equalization current control

[0098] For each equalization branch, the control cycle is 100 microseconds (matched to the DC / DC switching frequency of 10kHz):

[0099] The initial duty cycle is calculated based on the aforementioned formula for initial duty cycle, and feedforward control is performed thereon. Based on the initial duty cycle feedback control calculation, the actual equalization current is sampled at a 100-microsecond cycle. Calculate the final duty cycle (parameter set to...). =0.5, =0.05, =0.01, integral limit is ±0.1), using a PID controller, the duty cycle is converted into a high-resolution PWM signal (such as 16-bit resolution) to drive the equalization switch action of the equalization branch.

[0100] Before performing the balancing operation, a multi-level safety and anomaly diagnosis mechanism is executed through the main controller:

[0101] Level 1: Pre-execution check. Detects the temperature rise predicted by the equilibrium adjustment. Does it exceed the threshold? If the limit is exceeded, balancing of that branch is disabled, and the event is recorded.

[0102] Level 2: Real-time monitoring. Within each PWM cycle, monitor whether the actual current, voltage, and temperature exceed safety limits: current > 2.1A, voltage > 4.25V or < 2.75V, temperature > 60℃. If any limit is exceeded, immediately shut down the corresponding switch and report the fault.

[0103] Level 3: Fault Tree Diagnosis. When an abnormal shutdown occurs, the fault tree experience base (stored in EEPROM) is queried. For example, if a battery cell frequently triggers excessively high voltage, the diagnosis is "abnormally increased internal resistance," and it is recommended to replace the battery cell and reduce the upper limit of the balancing current.

[0104] V. Strategy Knowledge Base and Online Learning Process

[0105] After each equalization cycle, the main controller packages and stores the target data in a local circular queue (e.g., with a capacity of 1000 entries), such as the digital fingerprints of all current battery cells, operating condition indicators, and SOC distribution histogram; the equalization current sequence output by the decision layer (only the current value); the change in SOC standard deviation within 1 second after execution, the energy loss during this equalization, and the maximum temperature rise; and records the dispersion (standard deviation) of capacity decay of each battery cell after every 1000 kilometers of driving or every 100 fast charging cycles. Once the local storage is full, the data is uploaded to the cloud via the 4G module.

[0106] The cloud server performs offline training once a week, including: using data from the last 30 days to... Medium coefficient , Nonlinear least squares fitting is performed to update global parameters in the cloud; using "improved cell consistency after balancing" and "cumulative aging penalty" as reward functions, Bayesian optimization is employed to adjust coefficients under different operating conditions. , , , , The system generates a new mapping table; or collects the best solution data generated by the interior-point solver over the past 7 days (covering edge conditions), incrementally trains a deep neural network (DNN approximation optimizer), and retains 20% of the old model weights to prevent catastrophic forgetting. The updated model parameters are packaged into a differential upgrade package (less than 50KB) in the cloud and sent to the fleet vehicles via the vehicle-to-everything (V2X) network. Vehicles will automatically and silently upgrade the system the next time they stop and their State of Charge (SOC) is greater than 20%; if the upgrade fails, it will roll back to the previous version. The entire upgrade process does not affect the basic safety functions of the equalization system.

[0107] According to an embodiment of this application, a battery dynamic balancing system is further provided, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the battery dynamic balancing method based on the operating state as described.

[0108] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. It will be apparent to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the spirit or basic characteristics of the present invention. Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects. The scope of the present invention is defined by the appended claims rather than the foregoing description, and therefore all changes falling within the meaning and scope of the equivalents of the claims are intended to be included within the present invention.

[0109] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for dynamic balancing of batteries based on operating conditions, characterized in that, include: Based on the collected operating condition data of each battery cell and the electrochemical parameters of the battery cells identified online, the digital fingerprint of each battery cell is determined. Within a preset period, based on the cell digital fingerprint and multi-objective dynamic cost function, the equalization current sequence and equalization switch state matrix sequence of the equalization branch within the future time window are predicted, and the battery is dynamically equalized according to the equalization current sequence and equalization switch state matrix sequence. The cost function is determined by weighted summation based on the difference between the current cell SOC and the average SOC, the equalization energy loss, the temperature rise predicted for equalization, the aging factor penalty term, and the equalization time, which are dynamically determined according to different operating conditions.

2. The battery dynamic balancing method based on operating state according to claim 1, characterized in that, The expression for the cost function is as follows: ； In the formula, Represents the cost function, Indicates the current cell SOC. Indicates the average SOC of the battery cell. Indicates the equalization current. Indicates the equalization resistance. This represents the temperature rise predicted in equilibrium. This indicates a penalty for aging factors. This indicates the time required for equilibrium to be reached. , , , , Indicates the weighting coefficient. Indicates the number of battery cells. This indicates the cell number.

3. The battery dynamic balancing method based on operating state according to claim 1, characterized in that, The cell digital fingerprint includes at least the cell's physical serial number, real-time terminal voltage, real-time temperature, open-circuit voltage, estimated state of charge (SOC), ohmic internal resistance, polarization internal resistance, polarization time constant, current capacity health (SOH), capacity decay rate, and aging stress index. Preferably, the current capacity health is determined based on the capacity decay, which is determined by weighted summation of the cell's cumulative throughput and usage time.

4. The battery dynamic balancing method based on operating state according to claim 1, characterized in that, The prediction of the equalization current sequence and equalization switch state matrix sequence of the equalization branch in the future time window is performed under preset constraints, including current constraints, temperature constraints, voltage constraints, and topological constraints. The current constraints include that the equalization current is not greater than the upper limit of the hardware equalization current; the temperature constraints include that the sum of the current temperature of the cell and the corresponding temperature rise is not greater than the safe temperature of the cell; the voltage constraints include that the cell voltage is not greater than the maximum charging voltage during charging and not less than the minimum discharging voltage during discharging; the topological constraints include the combination of equalization switch states determined by the equalization hardware, including constraints on energy transfer between adjacent cells or constraints on preventing the formation of short-circuit loops after the equalization switch is turned on.

5. The battery dynamic balancing method based on operating state according to claim 1, characterized in that, The prediction of the balanced current sequence and balanced switch state matrix sequence of the balanced branch in the future time window is processed using an interior-point solver or a pre-trained deep neural network as an approximate optimizer, including: Construct the Lagrange function based on the cost function: ； N is the number of cells to be balanced. Let be an N-dimensional Boolean matrix. The number of discrete points is determined by the prediction time domain and the discrete time step. Represents optimization variables The cost function, These are optimization variables, including the equalization current sequence and the equalization switch state; These are slack variables, ensuring that the iteration point is always within the feasible region; It is a positive scalar, controlling the barrier function term. The strength; It is a Lagrange multiplier. This means relaxing the equality constraints into the cost function. Represents inequality constraints, limiting The feasible domain includes current, temperature, and voltage constraints; Solve the problem using Newton's method. The corrected equation is used to calculate the search direction; Iterative updates are performed based on the search direction to reduce variables. Once the accuracy requirements are met, the optimal equalization current sequence and equalization switch state within the future time window are output through a rolling solution process.

6. The battery dynamic balancing method based on operating state according to claim 1, characterized in that, The aging factor penalty term is calculated online using the electrochemical model SPMe. The SPMe model takes the current cell digital fingerprint and predicted equilibrium current and temperature as input, outputs preset indicators for the cell within a future time window, and integrates these as a lifetime penalty term. The preset indicators include at least a lithium plating risk index and an increase in SEI film thickness. Preferably, the calculation model for the aging factor penalty term includes: ； In the formula, Indicates battery cell At discrete time The lithium plating risk index, Indicates battery cell In the The thickness increment of the SEI film at discrete time points; , Indicates non-negative weight coefficients. This represents the discrete time step.

7. The battery dynamic balancing method based on operating state according to claim 1, characterized in that, The dynamic balancing of the battery based on the balancing current sequence and the balancing switch state matrix sequence includes: The initial duty cycle is determined based on the predicted equalization current of the equalization branch in the future time window, the upper limit of the hardware equalization current, and the steady-state gain of the equalization DC / DC converter. A PID controller is used to adjust the initial duty cycle to obtain the target duty cycle by using the actual sampled equalization current as feedback. The PWM signal is determined based on the target duty cycle, and the equalization current of the equalization branch is controlled by the action of the equalization switch on the equalization branch driven by the PWM signal.

8. The battery dynamic balancing method based on operating state according to claim 1, characterized in that, The dynamic balancing of the battery based on the balancing current sequence and the balancing switch state matrix sequence includes: During the balancing operation, the target parameters are monitored in real time to determine whether they exceed the preset boundary. If they do, the balancing operation is terminated. During the balancing process, faults are identified based on balancing anomalies, the fault tree experience database is queried to determine the fault type, and the upper limit of the balancing current is determined based on the fault type. The target parameters include at least temperature rise data, real-time current, voltage, and temperature.

9. The battery dynamic balancing method based on operating state according to claim 1, characterized in that, After each balancing cycle, the target data is stored locally and / or in the cloud for offline learning and updating of the model weight index. The target data includes the cell's digital fingerprint, operating condition indicator, SOC distribution histogram, and the balancing current sequence for each balancing cycle, the change in SOC standard deviation after balancing, balancing energy loss and maximum temperature rise, and the dispersion of cell capacity decay.

10. A battery dynamic balancing system, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the battery dynamic balancing method based on the operating state as described in any one of claims 1 to 9.

Images