Intelligent current sharing control method of energy storage system direct current convergence cabinet

Abstract

The invention provides an intelligent current-sharing control method for a direct current convergence cabinet of an energy storage system, and the method comprises the steps: constructing a fast and slow channel state prediction structure through high-precision multi-source data collection and verification, and carrying out the optimization of a short-term dynamic current instruction and a long-term current-sharing weight; timely observing and compensating systematic deviation by using Kalman filtering and hysteresis logic, and dynamically adjusting the priority of each battery cluster; fast and slow channel weighted fusion is realized in combination with the real-time load rate, and finally a high-adaptability current distribution instruction is generated; and a parameter self-updating mechanism is introduced to cope with the structural mismatch risk. According to the invention, the current sharing precision, the dynamic response speed and the robustness of long-term operation of the energy storage system are effectively improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent current sharing control and energy management of energy storage systems, and particularly relates to an intelligent current sharing control method for a direct current busbar cabinet of an energy storage system. BACKGROUND

[0002] As an important infrastructure for supporting new energy consumption and power peak shaving, the safe and efficient operation of an energy storage system puts higher requirements on the current sharing performance and health management of multiple battery packs. At present, the current distribution scheme of the mainstream direct current busbar cabinet of the energy storage system mainly relies on a single time scale current sharing control method, which statically or quasi-dynamically adjusts the current of each branch through centralized or distributed current sharing devices. This kind of method usually takes the battery cluster voltage, current or SOC estimated by BMS as the distribution basis, and cooperates with a simple weighted or proportional distribution rule, aiming to realize instantaneous current balance and prevent individual battery overload; In recent years, with the complication of battery technology and energy storage scenarios, some researches introduce model predictive control (MPC), sliding window regression, health state adaptive weighting and other multi-time scale control theories, realizing multi-level current distribution from second-level fast response to hour-level health management. Among them, the industry has applied extended Kalman filter, fuzzy self-tuning parameter, SOC balancing strategy and other methods to realize the correction of the state error of a single level. Some high-end systems also integrate self-learning algorithms to realize parallel scheduling of short-term dynamics and long-term capacity evaluation; The above disclosed technologies are suitable for large energy storage clusters and multi-cluster battery hybrid scenarios, and can maintain a certain degree of current sharing performance and system life under system startup, load change and temperature rise conditions. However, the current multi-time scale current sharing control is mainly based on open-loop prediction strategy, and lacks dynamic closed-loop correction of prediction errors in the multi-level fusion process. In actual operation, core state parameters such as SOC and SOH are affected by complex factors such as sensor drift, model mismatch and thermal runaway, and their long-term trend prediction at the minute and hour level has significant accumulated errors. These errors gradually penetrate the whole process of current distribution with the strategy fusion, eventually leading to increased circulating current between branches, local cluster capacity decay and system imbalance, which cannot meet the current sharing accuracy and control robustness under the dynamic changes of complex loads and health trends. SUMMARY

[0003] The present application provides an intelligent current sharing control method for a direct current busbar cabinet of an energy storage system to solve the above technical problems.

[0004] The technical solution of the present application is as follows: an intelligent current sharing control method for a direct current busbar cabinet of an energy storage system, comprising: S1: Real-time acquisition of output current, voltage and circuit breaker status parameters of each battery cluster branch in the DC combiner cabinet of the energy storage system, and synchronous acquisition of SOC and temperature data provided by the battery management system to form a multi-dimensional operating status dataset. S2: Construct a dual-channel state prediction structure based on the collected multi-dimensional data: The fast channel uses a model prediction control framework with a sampling frequency of 1 second to generate short-term current commands, and the slow channel outputs the benchmark current sharing weight through a sliding window regression model updated every minute. The fast channel input includes real-time current and voltage data, and the slow channel input includes SOC decay trend and temperature accumulation effect parameters. S3: Calculate the error residual sequence between the actual discharge depth and the predicted discharge depth of each battery cluster, use extended Kalman filtering to recursively estimate the residual dynamics, extract the systematic deviation components and generate the error observation matrix, wherein the dimension of the error observation matrix is ​​consistent with the number of battery clusters; S4: When the cumulative error of a battery cluster in the error observation matrix exceeds the preset threshold, the hysteresis logic threshold mechanism is triggered: Only when the consistency of the error direction reaches 90% within three consecutive sampling periods, a correction signal opposite to the error direction is generated. The amplitude of the correction signal is linearly negatively correlated with the cumulative error. S5: The correction signal is injected into the weight optimization process of the slow channel in reverse to dynamically adjust the schedulable priority coefficient of the corresponding battery cluster in the next sliding window period. The adjustment range of the priority coefficient is exponentially positively correlated with the amplitude of the correction signal and the adjustment range is limited to the interval [0.8, 1.2]. S6: Perform weighted fusion of fast and slow channel outputs based on the current system load rate: when the load rate is higher than 80%, a linear combination of 0.7 fast channel weight + 0.3 slow channel weight is used; when the load rate is lower than 30%, a combination of 0.3 fast channel weight + 0.7 slow channel weight is used; for intermediate load rates, the weight allocation ratio is determined by linear interpolation. S7: Send the fused current distribution command to the power regulation unit and monitor the deviation between the actual output of each branch and the command value in real time. When the deviation exceeds 5% for 10 seconds, start the feedback optimization mechanism: tune the model predictive control parameters of the fast channel online. The tuning step size is positively correlated with the integral value of the deviation. S8: Periodically detect the distribution of singular values ​​in the error observation matrix. When the ratio of the largest singular value to the second largest singular value exceeds 10:1, it is determined that there is a structural mismatch in the system. At this time, the parameter self-update process of the multi-timescale coupled controller is started. The update content includes the sliding window length, Kalman filter covariance matrix and hysteresis threshold.

[0005] The present invention provides an intelligent current sharing control method for a DC combiner cabinet in an energy storage system, which has the following beneficial effects: (1) This invention achieves high-precision dynamic current sharing under fast and slow dual-channel coordinated regulation by constructing a multi-time-scale coupled control architecture. The fast channel adopts a high-bandwidth model predictive control framework to quickly respond to transient changes in current, voltage and temperature of each branch at a second-level sampling frequency, ensuring the real-time power matching capability of the system under sudden load or start-up and shutdown conditions. The slow channel is based on a sliding window regression model to perform minute-level modeling of the health trend of the battery cluster (such as SOC / SOH evolution), generating a benchmark current sharing weight for long-term capacity consistency, effectively guiding the system to converge to an equilibrium state. The two channels focus on short-term dynamic response and long-term lifetime optimization objectives, respectively, forming a complementary mechanism in the time dimension, which significantly improves the stability and consistency of multi-battery cluster collaborative operation under complex operating conditions. (2) This invention innovatively introduces an online error compensation unit, which uses extended Kalman filtering to recursively estimate the residual sequence between the actual discharge depth and the predicted value, extracts the deviation components with systematic characteristics, and injects them as correction signals into the weight optimization process of the slow channel to dynamically adjust the schedulable priority of the corresponding battery clusters. This closed-loop feedback mechanism breaks through the limitations of the traditional one-way prediction-execution process, and can actively identify and correct trend deviations caused by BMS estimation errors or environmental disturbances, preventing individual branches from being continuously overloaded or underutilized. At the same time, combined with the hysteresis logic threshold design, the compensation action is triggered only when the residual direction is consistent within several consecutive cycles, avoiding misadjustment caused by instantaneous noise or short-term fluctuations, and significantly enhancing the robustness and decision reliability of the control system. In addition, the strategy fusion layer adaptively adjusts the proportional coefficient of the fast and slow channel outputs according to the current load rate of the system, prioritizing response speed in high power demand scenarios and strengthening the balancing guidance role in low load periods, so that the overall control strategy has the ability to adapt to operating conditions, taking into account both efficiency and durability. Attached Figure Description

[0006] Figure 1 A flowchart of an intelligent current sharing control method for a DC combiner cabinet in an energy storage system according to the present invention; Figure 2 This is a sub-flowchart of an intelligent current sharing control method for a DC combiner cabinet in an energy storage system according to the present invention. Figure 3 This is another sub-flowchart of the intelligent current sharing control method for a DC combiner cabinet of an energy storage system according to the present invention. Detailed Implementation

[0007] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0008] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0009] like Figure 1 As shown, this invention provides an intelligent current sharing control method for a DC combiner cabinet in an energy storage system, specifically including: like Figure 1 As shown, this application provides an intelligent current sharing control method for a DC combiner cabinet in an energy storage system, specifically including: S1: Real-time acquisition of output current, voltage and circuit breaker status parameters of each battery cluster branch in the DC combiner cabinet of the energy storage system, and synchronous acquisition of SOC and temperature data provided by the battery management system to form a multi-dimensional operating status dataset. S2: Construct a dual-channel state prediction structure based on the collected multi-dimensional data: The fast channel uses a model prediction control framework with a sampling frequency of 1 second to generate short-term current commands, and the slow channel outputs the benchmark current sharing weight through a sliding window regression model updated every minute. The fast channel input includes real-time current and voltage data, and the slow channel input includes SOC decay trend and temperature accumulation effect parameters. S3: Calculate the error residual sequence between the actual discharge depth and the predicted discharge depth of each battery cluster, use extended Kalman filtering to recursively estimate the residual dynamics, extract the systematic deviation components and generate the error observation matrix, wherein the dimension of the error observation matrix is ​​consistent with the number of battery clusters; S4: When the cumulative error of a battery cluster in the error observation matrix exceeds the preset threshold, the hysteresis logic threshold mechanism is triggered: Only when the consistency of the error direction reaches 90% within three consecutive sampling periods, a correction signal opposite to the error direction is generated. The amplitude of the correction signal is linearly negatively correlated with the cumulative error. S5: The correction signal is injected into the weight optimization process of the slow channel in reverse to dynamically adjust the schedulable priority coefficient of the corresponding battery cluster in the next sliding window period. The adjustment range of the priority coefficient is exponentially positively correlated with the amplitude of the correction signal and the adjustment range is limited to the interval [0.8, 1.2]. S6: Perform weighted fusion of fast and slow channel outputs based on the current system load rate: when the load rate is higher than 80%, a linear combination of 0.7 fast channel weight + 0.3 slow channel weight is used; when the load rate is lower than 30%, a combination of 0.3 fast channel weight + 0.7 slow channel weight is used; for intermediate load rates, the weight allocation ratio is determined by linear interpolation. S7: Send the fused current distribution command to the power regulation unit and monitor the deviation between the actual output of each branch and the command value in real time. When the deviation exceeds 5% for 10 seconds, start the feedback optimization mechanism: tune the model predictive control parameters of the fast channel online. The tuning step size is positively correlated with the integral value of the deviation. S8: Periodically detect the distribution of singular values ​​in the error observation matrix. When the ratio of the largest singular value to the second largest singular value exceeds 10:1, it is determined that there is a structural mismatch in the system. At this time, the parameter self-update process of the multi-timescale coupled controller is started. The update content includes the sliding window length, Kalman filter covariance matrix and hysteresis threshold.

[0010] Step S1: Real-time acquisition of output current, voltage, and circuit breaker status parameters of each battery cluster branch in the DC combiner cabinet of the energy storage system, and simultaneous acquisition of SOC and temperature data provided by the battery management system to form a multi-dimensional operating status dataset. Specifically, this includes: S1.1: Based on a high-precision current sensor and voltage sampling module, the output current and voltage of each battery cluster branch are synchronously collected to obtain the original current and voltage data sequence. The sampling frequency is not less than 10Hz and the sampling accuracy is not less than 0.5%FS to ensure the real-time performance and accuracy of data acquisition. The input conditions are the analog signals output by the current sensors and voltage sampling modules of each battery cluster branch in the DC combiner cabinet of the energy storage system. Synchronous acquisition is required to ensure the comparability of each physical quantity under the same time reference. A high-precision Hall effect current sensor (parameters: range ±500A, linearity better than 0.5%FS) is used to collect the instantaneous output current of each battery cluster branch, realizing non-contact measurement of large DC current, and converting the analog signal into a digital quantity for subsequent processing; Furthermore, the voltage at each branch terminal is collected by a precision voltage divider sampling module (parameters: insulation withstand voltage ≥1500V, sampling error ≤0.5%FS) to realize real-time monitoring of the DC bus voltage by branch, and the sampled signal is quantized into a voltage value sequence by an analog-to-digital conversion unit; Furthermore, by using synchronous acquisition control logic (parameters: acquisition frequency above 10Hz, time synchronization accuracy ±1ms), the current and voltage sampling process is triggered and aligned, so that the two sensor outputs have a unified timestamp, forming a pair of original current and voltage data with consistent timing, avoiding phase mismatch problems caused by acquisition delay. Furthermore, a digital filtering algorithm (such as finite impulse response (FIR) filter, with parameters of filter order 16 and passband cutoff frequency of 2Hz) is used to denoise the original current and voltage data sequence, suppress high-frequency interference components, and maintain the low-frequency trend signal without distortion, thus ensuring measurement accuracy. By using high-resolution data packaging and indexing, the filtered current and voltage data are stored as a multi-dimensional array structure according to branch number and timestamp, thereby realizing the encapsulation and efficient retrieval of the original operating parameters; For example, in a DC combiner cabinet of an energy storage station, each battery cluster branch is equipped with a LEM brand Hall current sensor (rated range ±300A, output sensitivity 2.5mV / A) and a voltage sampling module based on 0.1% high-precision resistor voltage division. The acquisition controller is set to a sampling period of 0.05s, a sampling accuracy of 0.3%FS, and adopts a dual-channel synchronous trigger mode. The current signal is processed by a 16th-order FIR low-pass filter, with the coefficients selected using a Hanning window based on the window function design method, and a cutoff frequency of 2Hz; the voltage signal is synchronously processed with the same filtering parameters to ensure amplitude-frequency consistency. Taking a certain period of sampling value as an example, the branch current corresponding to a current sensor output of 2.750V is... = A, The voltage sampling module outputs 1.234V, which is then divided by a voltage divider. Converted = V. The collected data, after being encapsulated with a unified timestamp, is stored in the controller buffer. It can be merged with the circuit breaker status dataset in S1.2 to form a complete multi-dimensional operating status dataset. The measured branch operating parameters show a significant improvement in input accuracy in subsequent state prediction and current sharing control. S1.2: The on / off status of each branch circuit breaker in the DC combiner cabinet is monitored in real time using a digital input module to obtain the circuit breaker status signal. The status signal is processed by de-jitter filtering to form a Boolean on / off indicator, which is used to determine whether the branch is in an effective operating state. For the on / off status of each branch circuit breaker in the DC combiner cabinet, a digital input module with high-speed sampling capability (sampling frequency ≥ 1kHz, input channel isolation voltage ≥ 2.5kV) is used to realize the real-time acquisition of status signals and generate the original on / off status data vector indexed by the channel number. Furthermore, by using a jitter-removing filtering algorithm (parameters: time window length = 20ms, threshold determination rule = fixed logic threshold method), the instantaneous jitter of the original on / off state signal is eliminated, and a stable determination signal vector is obtained; Furthermore, a threshold logic mapping processing algorithm (parameters: high-level judgment value = 24V, low-level judgment value = 0V) is employed to convert the stability judgment signal vector into a Boolean state flag matrix, where the matrix elements take values ​​of... Indicates the path status, with values ​​ranging from 0 to 1. Indicates an open circuit state; Furthermore, branch validity verification is performed by comparing the Boolean status identifier matrix with the current sampling results item by item. If there is a logical conflict between the status identifier and the current value of a branch (e.g., the branch is identified as a closed circuit but the current is zero), the branch is marked as an abnormal operating state and an abnormal label vector is generated. Furthermore, by performing a logical AND operation between the anomaly marker vector and the Boolean state identifier matrix, the set of valid operating branches is output as the circuit breaker state component of the multi-dimensional operating state dataset. Through a step-by-step processing method of digital input acquisition, de-jitter filtering, logic mapping, and validity verification, the original status signal of the circuit breaker is converted into a Boolean on / off indicator that can be directly used for system status identification, thereby enabling accurate judgment and real-time updating of the effective operating status of the branch. For example, an 8-channel digital input module is configured in the DC combiner cabinet of the energy storage system. The sampling frequency is set to 1kHz, the isolation voltage is set to 3kV, the time window length is set to 20ms, and the logic judgment thresholds are 24V for high level and 0V for low level. During the acquisition phase, the original sampling vector of the 8 circuit breaker status signals is [24.5, 0.1, 24.3, 24.4, 0.0, 24.5, 0.2, 24.6] (unit: V). After de-jitter filtering, the stable signal vector is [24.5, 0.0, 24.3, 24.4, 0.0, 24.5, 0.0, 24.6]. A Boolean matrix [1,0,1,1,0,1,0,1] is generated through logical mapping. The validity verification process compares the current sampling vector (unit: A) [100,0,98,101,0,99,0,102] of the same period and finds no logical conflicts. The anomaly marker vector is [0,0,0,0,0,0,0,0]. The result of the logical AND operation yields the set of valid operating branches [1,0,1,1,0,1,0,1]. This set is directly incorporated into the multi-dimensional operating status dataset to ensure that the circuit breaker state components are consistent with the actual physical operating state, significantly improving the reliability of subsequent state prediction and current sharing control. S1.3: Establish a communication connection with the battery management system (BMS) through the CAN bus and Modbus protocol interface, and periodically acquire the SOC (State of Charge) and temperature data of each battery cluster based on a predefined data frame format. The SOC data update cycle is 1 second and the temperature data update cycle is 5 seconds to ensure the continuity and synchronization of state perception. In the main step of data acquisition for DC combiner cabinet, in order to meet the requirements for obtaining SOC and temperature of battery clusters, the communication target is set to the battery management system (BMS) of each battery cluster, and the input signals include standard SOC estimates and temperature measurements. A dual-protocol mutual backup communication mechanism with high stability and high anti-interference capability is achieved by adopting a physical layer connection based on the CAN bus protocol and an application layer communication scheme based on the Modbus protocol (parameters: baud rate 500kbps, data bits 8 bits, parity bit 1 bit, stop bit 1 bit). Furthermore, by loading a predefined data frame format (parameters: start identifier 0xAA, data length fixed at 8 bytes, CRC check polynomial 0xA001), the structured parsing of SOC and temperature data is achieved, and the original state information after separating the frame header, data field and frame tail is obtained; Furthermore, through a periodic query mechanism (parameters: SOC query period 1s, temperature query period 5s), a data acquisition task is triggered and numerical fields are extracted from the BMS response frame to form a time-series SOC vector and temperature vector. Furthermore, for the collected time series data, an outlier detection algorithm (parameter: 3σ statistical criterion) is used to filter out abrupt changes in data to ensure the continuity and validity of the input data; Furthermore, by using a timestamp marking algorithm (parameter: accuracy ±1ms), the SOC and temperature data are associated with the current, voltage and circuit breaker status data already collected in the DC combiner cabinet, so as to realize the synchronous perception of multi-source operating status. Through the above-mentioned communication and data processing chain derivation, the sensor acquisition results of the previous step are transformed into continuous, synchronous and anomaly-filtered SOC and temperature operating status indicators, so as to realize multi-dimensional and accurate perception of the working status of the battery cluster. For example, in an energy storage station operation scenario, each battery cluster in the combiner cabinet is equipped with an independent BMS unit, the CAN bus baud rate is set to 500kbps, the physical layer uses twisted-pair shielded cable, and the maximum number of nodes is set to 32. The application layer uses Modbus RTU mode, with an address range of 1~32. Function code 0x03 reads the holding register, where the SOC register address is set to 0x0100, the temperature register address is set to 0x0200, and the register values ​​are 16-bit unsigned integers. The SOC value unit is 0.1%, and the temperature unit is 0.1℃. Periodic data acquisition tasks are triggered by timer interrupts. SOC data is updated once per second, and temperature data is updated every 5 seconds. After acquisition, an outlier removal algorithm is used. According to the 3σ principle, when the increase or decrease of SOC in a certain period exceeds three times the standard deviation of the average of the previous 10 periods, it is judged as an anomaly and removed, and the anomaly event is recorded. The timestamp association employs a global clock synchronization mechanism, stamping the SOC and temperature values ​​with millisecond-level timestamps and indexing them against the current, voltage, and circuit breaker status data of the current cycle. During the verification phase, the associated multi-dimensional status data is input into the slow channel of the dual-channel status prediction structure as the basis for generating the health status trend feature vector. The system demonstrates significantly improved current sharing consistency and enhanced circulating current suppression during long-term operation. S1.4: The collected raw current, voltage, circuit breaker status, SOC and temperature data are timestamped to construct a multi-source heterogeneous data set under a unified time base. The timestamp accuracy is controlled within ±1ms to support the time-series correlation analysis of subsequent modules. S1.5: The validity of the multi-source data set is verified based on the data integrity verification mechanism. If the data of a certain battery cluster branch is found to be missing or abnormal, the data repair process is triggered: the missing value is compensated by interpolation of adjacent branch data, and the abnormal data is marked for subsequent diagnostic module processing, thereby forming a multi-dimensional operating status dataset with complete structure and controllable quality.

[0011] Step S2: Constructing a dual-channel state prediction structure based on the collected multi-dimensional data: The fast channel uses a model prediction control framework with a 1-second sampling frequency to generate short-term current commands, while the slow channel outputs benchmark current sharing weights through a sliding window regression model updated every minute. The fast channel input includes real-time current and voltage data, while the slow channel input includes parameters related to SOC decay trend and temperature accumulation effects. Specifically, this includes: S2.1: Normalize the real-time collected output current and voltage data of each battery cluster to eliminate the differences in physical dimensions between different branches and obtain a standardized current and voltage vector with unified dimensions, so as to provide input features for fast-channel model predictive control. S2.2: Based on the standardized current and voltage vectors, a model predictive control framework with a sampling frequency of 1 second is constructed. A rolling optimization strategy is used to generate short-term current commands. The optimization objective function includes minimizing the rate of change of current commands and power tracking error, and outputs a fast-channel current prediction sequence. Based on the normalized current-voltage vector, a model predictive control (MPC) method is adopted (parameters: sampling period 1 second, prediction time domain). =10. Control Time Domain =5. Constraints include upper and lower limits of current and rate of change limits, to achieve rolling optimization calculation of fast-channel short-term current commands; Furthermore, the objective function is constructed using a method (parameter: weight coefficients). Used for minimizing the rate of change of current command, weighting coefficient (Used for minimizing power tracking error), achieving dual-index coupled optimization, and obtaining the cost function matrix required for optimization solution; Furthermore, numerical optimization is performed using the following objective function expression:

[0012] in, This represents the change in current command between adjacent sampling periods. For the predicted branch output power, The target reference power; Furthermore, a quadratic programming (QP) optimization algorithm is used (parameter: iteration tolerance set to 1×10^). 6. With a maximum number of iterations of 100, the objective function is solved under constraints, and a fast-channel current prediction sequence that satisfies the constraints is generated. Furthermore, a state-space discretization method (parameter: sampling period of 1 second) is adopted to transform the continuous operation characteristics of the prediction model into a discrete mathematical model, so as to support the dynamic update of MPC in the rolling optimization process. By using a rolling optimization strategy, the first control quantity in each optimal current command sequence obtained from the prediction time domain calculation is output to the fast-channel execution module to form a fast-channel current prediction sequence, thereby achieving short-term dynamic response optimization. For example, in a certain energy storage system, the normalized current-voltage vector is input into the MPC module, and the prediction time domain is set. =12 and control time domain =4, current upper and lower limits are constrained to ±200A, current change rate is limited to ±20A / s, and power target is set at 500kW. The objective function weighting coefficients are used. =0.4、 =0.6 is used for dual-index optimization, and the optimization formula is the MathML expression mentioned above. The tolerance for quadratic programming is set to... The maximum number of iterations is 120. During operation, the fast-channel model completes prediction, solution, and output within a sampling period per second. The resulting current prediction sequence remains smooth under dynamic load fluctuations, the instantaneous power tracking error is significantly reduced, and the transient response performance and circulating current suppression effect of the system are significantly improved. S2.3: Extract features from the SOC decay trend and temperature accumulation effect parameters provided by the battery management system, model the long-term capacity decay behavior using a sliding window regression model, and extract the battery cluster health status trend feature vector as the input variable for slow channel current sharing weight calculation. S2.4: Based on the trend feature vector of battery cluster health status, perform sliding window regression model calculation, use the least squares method to optimize weight parameters, and output a benchmark current sharing weight matrix updated every minute. This matrix represents the capacity scheduling priority of each battery cluster under long-term operation. S2.5: The fast-channel current prediction sequence and the slow-channel reference current sharing weight matrix are time-aligned to construct a fusion interface for the dual-channel state prediction structure, generating a joint state prediction vector containing short-term dynamics and long-term trends, which serves as the input basis for the error observation and compensation mechanism.

[0013] like Figure 2 As shown, step S3 involves calculating the error residual sequence between the actual discharge depth and the predicted discharge depth of each battery cluster, recursively estimating the residual dynamics using an extended Kalman filter, extracting the systematic deviation component, and generating an error observation matrix. The dimension of the error observation matrix is ​​consistent with the number of battery clusters. Specifically, this includes: S3.1: Based on the real-time discharge current and voltage data of each battery cluster, combined with the SOC status information, an ampere-hour integration and open-circuit voltage correction fusion algorithm is executed to obtain a high-precision actual depth of discharge value; Based on the high-precision current sensor output data of each battery cluster branch and the data collected by the voltage sampling module, a time-series real-time discharge current sequence and a real-time terminal voltage sequence are constructed. Combined with the SOC state value periodically output by the battery management system, a three-dimensional time series input including current, voltage, and state of charge is formed as the raw data basis for calculating the depth of discharge. An ampere-hour integration method (parameters: sampling frequency ≥ 10Hz, integration time window synchronized with fast channel sampling period) is used to calculate the cumulative discharge of each battery cluster in the current time period. The ampere-hour integration calculation formula is as follows:

[0014] in, This is the real-time discharge current (in A). The negative cube of is the conversion factor from seconds to hours. The current integration window length (in seconds), and the integration result. Cumulative discharge amount (unit: Ah); Furthermore, the accuracy of the ampere-hour integration results is corrected using an open-circuit voltage correction method (parameters: open-circuit voltage-capacity curve provided by the battery manufacturer). This method first eliminates transient voltage offsets under loaded conditions using a terminal voltage and internal resistance estimation model. Then, the corrected terminal voltage value is mapped to the corresponding open-circuit voltage value, and the corresponding theoretical depth of discharge is derived by combining it with the SOC reference curve. ; A weighted fusion algorithm (parameter: the weighting coefficient α depends on the sensor accuracy and the goodness of fit of the calibration curve) is used to achieve a comprehensive estimate of the ampere-hour integrated discharge quantity and the discharge quantity derived from the open-circuit voltage. The fusion formula is as follows:

[0015] in This is the actual depth of discharge with high precision (in Ah or corresponding percentage), and this value serves as the benchmark actual value for subsequent error residual calculation; The output of the above fusion algorithm The time series transforms the multi-source raw data from the previous step into a corrected high-precision discharge depth characterization, ensuring that the basic data in the error observation stage are accurate and reliable. For example, during the operating cycle of a certain energy storage system, the real-time sampling frequency of the battery cluster branch is configured to 20Hz, the sampling accuracy to 0.3%FS, and the integration period to 1 second. The cumulative discharge during this cycle is calculated using ampere-hour integration. Ah. The open-circuit voltage curve calibration value shows that the discharge depth corresponding to this terminal voltage is... Ah, the sensor accuracy evaluation weighting coefficient α is set to 0.6, and the high-precision discharge depth is calculated according to the fusion formula. = Ah. This value, after being normalized and converted to a percentage, is used as the actual discharge depth input of the system in that cycle. Verification results show that this method can significantly improve the accuracy of discharge depth calculation under multiple operating conditions, ensuring the stability of subsequent error calculation and Kalman filter estimation; S3.2: Calculate the theoretical depth of discharge for each battery cluster at the current time scale based on the short-term current command output by the fast-channel model predictive control framework and the benchmark current sharing weight output by the slow-channel sliding window regression model. S3.3: Calculate the point-by-point difference between the actual discharge depth value and the theoretical discharge depth value to generate an error residual sequence, so as to quantify the deviation distribution between the model prediction and the actual operation. The input data includes the high-precision actual depth of discharge value obtained by S3.1 and the theoretical depth of discharge value calculated by S3.2. The two are corresponding according to the battery cluster number under the same time reference and sampling frequency. A point-by-point difference calculation method (parameters: sampling period of 1 second, timestamp accuracy of ±1ms) is used to extract the difference between the actual discharge depth and the theoretical discharge depth of the same battery cluster at the same sampling time, and obtain a preliminary deviation dataset. Furthermore, a residual vector is generated by using a residual construction algorithm (parameters: difference sequence length N, sequence window length M), and the difference data is arranged in time series to form the error time series for each battery cluster. Furthermore, by using a denoising filtering method (parameters: three-point moving average, boundary conditions using mirror expansion), high-frequency noise suppression of the residual sequence is achieved, and a smoothed residual curve is generated to improve the accuracy of error quantization. Furthermore, by using the distribution feature extraction method, the mean and variance indices of each residual curve are calculated, and a residual statistical matrix containing bias and fluctuation characteristics is output, providing prior distribution information for subsequent state space modeling; Through the above chain calculation, the difference between the actual and theoretical discharge depth values ​​is transformed into an error residual sequence that reflects the deviation distribution between the prediction model and actual operation, thereby realizing a quantitative evaluation of the performance of the dual-channel state prediction structure. For example, in an energy storage system containing 8 battery clusters, assuming a sampling period of 1 second, the actual depth of discharge (DHD) is measured in Ah, and the theoretical DHD is predicted and output jointly by the fast and slow channels. At sampling time t=100 seconds, the actual DHD of a certain battery cluster is 45.382 Ah, and the theoretical DHD is 45.215 Ah. The point-by-point difference calculation formula is used:

[0016] in, This is the actual depth of discharge value. This represents the theoretical depth of discharge. Substituting the values ​​yields the error. for Ah. A matrix is ​​constructed from the difference sequences of 600 consecutive seconds. A three-point moving average filter is used to process the boundaries, and the matrix is ​​expanded using a mirror strategy to obtain a smooth residual curve. The mean of the curve is... Ah, the variance is (Ah²). This residual sequence, used as the observation input in subsequent extended Kalman filter state modeling, can significantly improve the accuracy of systematic deviation component extraction and ensure the robustness of the multi-timescale coupled controller in long-term equilibrium strategies; S3.4: Based on the error residual sequence, construct the state space model of the extended Kalman filter, where the system state vector includes the discharge depth deviation, SOC estimation error and temperature drift term, so as to recursively estimate the dynamic characteristics of the residual; S3.5: The error residual sequence is recursively updated and covariance corrected using the state-space model of the extended Kalman filter, and the systematic deviation components of each battery cluster are extracted to form a deviation estimation vector; S3.6: Arrange the deviation estimation vectors of each battery cluster according to the branch number to construct the error observation matrix. The dimension of the error observation matrix is ​​consistent with the number of battery clusters, which provides the input basis for the subsequent hysteresis logic threshold mechanism. After obtaining the deviation estimation vector of each battery cluster, the branch number index mapping rule (parameter: number interval [1, N], where N is the number of battery clusters) is used to realize the sequential arrangement of the deviation estimation vector in physical location; Furthermore, a matrix construction method is used (parameters: matrix dimension N×1, element type floating-point, precision not less than...). This allows the sorted bias estimation vectors to be written row by row into the corresponding row positions of the matrix structure, forming a basic observation array; Furthermore, a matrix dimension matching algorithm (parameter: target dimension N×M, M is the number of bias features per cluster) is used to expand or truncate the basic observation array to ensure that the column dimension of the error observation matrix is ​​consistent with the number of feature parameters of each battery cluster, and zero-value placeholders are filled in the missing column positions to maintain the stability of numerical calculation. Furthermore, a matrix normalization processing algorithm (parameters: normalization interval [-1,1], standardization method is z-score) is adopted to perform column-independent standardization operation on all elements of the observation matrix to eliminate the influence of different feature dimensions on subsequent logical threshold judgment and obtain the normalized error observation matrix. Furthermore, by using the matrix timestamp binding method (parameter: clock synchronization accuracy ±1ms), a unified timestamp of the current sampling period is appended to the metadata tag area of ​​the error observation matrix to ensure the correlation and traceability of the observation results in the time domain. Through the matrix construction and normalization process described above, the deviation estimation vector from the previous step is transformed into a structured error observation matrix, thereby providing the technical effect of input for the subsequent hysteresis logic threshold mechanism. For example, in an energy storage system containing 8 battery clusters, the deviation estimation vector output by the extended Kalman filter is [-0.005, 0.012, -0.009, 0.004, -0.003, 0.015, -0.007, 0.002], and the branch number index range is [1, 8]. This vector is arranged in numerical order in the corresponding row positions of the 8×1 base array using a mapping rule. The target error observation matrix dimension is set to 8×3, where the deviation feature parameters for each cluster are, in order, depth of discharge deviation, SOC estimation error, and temperature drift. When performing matrix dimension matching, the base array is extended by two columns, and missing elements are filled with 0, forming:

[0017] Subsequently, column-wise z-score normalization is performed on the matrix, where the mean and standard deviation are calculated column-wise and the standardization formula is applied:

[0018] in, For matrix elements, This is the mean of the elements in this column. The standard deviation of the elements in this column is used. After normalization, an error observation matrix with all column values ​​ranging from [-1, 1] is obtained, and the timestamp of the current sampling period, 2024-07-30 18:25:15.123, is bound to its metadata tag area to prepare the input for subsequent hysteresis logic threshold judgment. In this embodiment, the normalization process ensures that the parameters of different columns are consistent on the numerical scale, and the direction can be directly compared with the threshold during hysteresis logic judgment, which significantly improves the stability of the judgment.

[0019] like Figure 3 As shown, step S4: When the cumulative error of a battery cluster in the error observation matrix exceeds a preset threshold, a hysteresis logic threshold mechanism is triggered: Only when the consistency of the error direction reaches 90% within three consecutive sampling periods, a correction signal opposite to the error direction is generated. The amplitude of this correction signal is linearly negatively correlated with the cumulative error. Specifically, it includes: S4.1: Perform cluster-by-cluster error accumulation calculation on each row of data in the error observation matrix, extract the cumulative error of each battery cluster in three consecutive sampling periods, and use the sliding window integration method to perform weighted summation on the error sequence to obtain the error accumulation vector, so as to characterize the degree of systematic deviation of each battery cluster in the slow channel prediction model. The input conditions include an error observation matrix calculated by an extended Kalman filter. The number of rows in this matrix corresponds to the number of battery clusters in the DC combiner cabinet of the energy storage system, and the number of columns corresponds to the number of consecutive sampling periods. The algorithm processes the error residual data sequence of each row in the matrix, all of which have been appended with precise timestamps. A sliding window integration method (window length: 3 sampling periods, weight decay factor: 0.8) is used to realize the local cumulative calculation function of the clustered error sequence; Furthermore, through weighted summation (the weights are set based on the latest cycle having the largest weight and the historical cycle weights decreasing successively), the error value is accumulated and decayed over time, resulting in a three-level cycle error accumulation result with time smoothing characteristics. Furthermore, matrix row vector reduction operation (target dimension: 1×N, where N is the number of battery clusters) is used to compress the cumulative error of each cluster after weighted integration into a single value to form an error accumulation vector; Furthermore, a noise suppression filtering algorithm (type: second-order low-pass filter, cutoff frequency: 0.25 times the sampling frequency) is used to smooth the error accumulation vector and generate low-noise systematic deviation amplitude estimation results. Furthermore, using the normalization transformation algorithm (normalization interval: [ [1,1], reference value: maximum absolute error accumulation), realizes the dimensionless processing of the error accumulation amplitude, and obtains a standardized error accumulation vector that can be directly used for hysteresis threshold comparison; Through the above chain-like processing of integration and weighted summation, the original multi-cycle error data is transformed into a quantitatively clear and noise-suppressed error accumulation index, thereby realizing a quantifiable expression of the degree of systematic deviation in the slow channel prediction model. For example, in an energy storage system containing 8 battery clusters, the sampling period is set to 1 second, the sliding window length is fixed at 3 periods, and the weight decay factor is set to 0.8. Taking the 4th battery cluster as an example, its error observations in three consecutive periods are 0.05, 0.08, and 0.12 (unit: SOC percentage), respectively. Applying the weighted integral formula in sequence:

[0020] The numerator is the sum of weighted error values, and the denominator is the sum of weighting coefficients, resulting in a weighted average cumulative error of 0.087. After processing with a second-order low-pass filter, the output smooth value is 0.084. This is then normalized (based on the maximum smooth value of 0.15 for the entire cluster) to obtain a standardized cumulative error value of 0.56. This value is used in subsequent hysteresis threshold comparisons to determine whether there are any systematic deviations in the fourth cluster that require compensation. Tests show that this method significantly improves the stability of error assessment and reduces the false trigger rate under various operating conditions. S4.2: Perform error direction consistency analysis based on the error accumulation vector. Extract the error direction in each sampling period through the sign function and calculate the percentage of error direction consistency in three consecutive periods. The consistency of direction is judged by whether the output value of the sign function is consistent. If the error direction consistency ratio reaches 90% or more in three consecutive periods, it is determined that the battery cluster has a persistent prediction deviation. The sign function extraction method (parameter: error accumulation vector as input) is adopted to identify the error sign in each sampling period and generate direction sequence data to characterize the error direction value of each battery cluster in each period. Furthermore, by using a continuous period consistency calculation method (parameters: direction sequence, analysis window length = 3), the percentage of error direction consistency within three sampling periods is calculated, and a direction consistency vector is obtained, where each element of the vector corresponds to the consistency percentage of a single battery cluster. Furthermore, by using a consistency threshold comparison method (parameter: threshold set to 90%), the percentage of each battery cluster in the directional consistency vector is determined, and a persistent prediction deviation flag matrix is ​​generated. This matrix is ​​arranged by branch number and identifies battery clusters that meet the threshold condition. Furthermore, through a logical decision method (parameters: direction consistency flag matrix, error accumulation vector), the persistent deviation of battery clusters that meet the direction consistency condition is confirmed, and a persistent prediction deviation list is generated as the input signal for the hysteresis logic threshold mechanism. By using symbol function extraction and consistency analysis, the error accumulation vector from the previous step is transformed into a persistent prediction deviation judgment result, thereby achieving a technical evaluation of the long-term prediction error direction stability of the battery cluster. For example, during the operation cycle of a large energy storage system containing 12 battery clusters, the sign function extraction method is used to determine the sign of the error accumulation vector [m1, m2, …, m12]. Defined as:

[0021] in, This represents the cumulative error of the battery cluster within the corresponding sampling period. A consistency percentage calculation is performed on the direction sequence with a window length of 3. The calculation shows that when the direction consistency percentage of a battery cluster is 100% and the cumulative error is 5Ah, it is marked as a persistent prediction deviation. This persistent prediction deviation list is passed to a hysteresis logic threshold mechanism, which is used in subsequent steps to generate a correction signal opposite to the error direction, thereby achieving error compensation. Actual results show that this method maintains stable direction determination even under load rate fluctuations, effectively reducing erroneous triggering caused by transient noise. S4.3: When the error direction consistency of a battery cluster meets 90% and its cumulative error exceeds the preset error threshold, the hysteresis logic threshold mechanism is triggered. The error threshold is set jointly based on the rated capacity of the battery cluster and the standard deviation of the SOC estimation error to avoid false triggering caused by noise interference, while ensuring a sensitive response to systematic deviations. A joint threshold determination process is performed on the cumulative error value of each battery cluster in the error observation matrix. The input data includes the rated capacity parameter of the battery cluster and the standard deviation of the SOC estimation error. A joint threshold setting algorithm is adopted (parameter: rated capacity). Standard deviation of SOC estimation error This allows for the dynamic calculation of error trigger thresholds and the acquisition of initial estimates of error thresholds. Furthermore, the scale-matching correction of the initial estimate is achieved by using the normalized scaling correction method (parameter: normalization coefficient of the standard deviation of the rated capacity and SOC estimation error), and a sequence of normalized error thresholds is generated. Furthermore, a noise protection filtering algorithm (parameters: sliding window length M and filtering coefficient α) is used to suppress high-frequency noise on the unitized error threshold and obtain a smoothed threshold vector. Furthermore, a directional consistency triggering logic (parameters: directional consistency ratio threshold 0.9 and number of cycles N=3) is adopted to jointly determine the cumulative error and the directional consistency condition, and generate a hysteresis logic triggering signal; By using a hysteresis logic threshold mechanism, the trigger signal from the previous step is logically ANDed with the smoothed threshold vector to transform it into a systematic deviation trigger event, thereby achieving the dual goals of noise suppression and deviation-sensitive response. For example, in an energy storage system containing 12 battery clusters, the rated capacity is... Take 240Ah, and the standard deviation of the SOC estimation error. The initial estimate of the error threshold is calculated using a joint threshold setting algorithm. = Ah. Using the normalized proportional correction factor. Correcting this value yields... = The normalized threshold of Ah is used. This threshold is input into the noise protection filtering algorithm, with the sliding window length M=5 and the filtering coefficient α= The output smoothing threshold is approximately: Ah. Within three consecutive sampling periods, the consistency rate of the error direction of this battery cluster reached... It meets the directional consistency threshold of 0.9, and the cumulative error is... Ah, above the smoothing threshold Ah, the hysteresis logic threshold mechanism is triggered, generating a deviation correction trigger event and using it for the reverse generation and injection of subsequent correction signals, enabling the system to accurately correct the scheduling priority of the battery cluster in the next control cycle, thereby significantly improving the consistency and robustness of long-cycle current sharing control. S4.4: Based on the triggered hysteresis logic threshold event, a correction signal opposite to the error direction is generated. The amplitude of the correction signal is mapped to the cumulative error through a linear negative correlation function. That is, the larger the error, the stronger the amplitude of the correction signal. The slope of the function is set according to the statistical results of the historical error distribution of the system to achieve precise control of the error compensation intensity. Based on the triggering conditions of hysteresis logic threshold events, a direction reversal mapping method (parameters: error direction vector, hysteresis event identifier) ​​is adopted to set the direction of the correction signal to be opposite to the direction of the current accumulated error. By using direction reversal mapping to form an input pair between the accumulated error and the signal direction, it is ensured that the compensation effect and the system deviation direction form a counterbalancing effect, thereby improving the effectiveness and stability of the compensation. A linear negative correlation function mapping algorithm (parameters: cumulative error, function slope coefficient) is used to calculate the inverse relationship between the amplitude of the corrected signal and the cumulative error, and to obtain the initial value of the corrected signal amplitude. The function structure is as follows:

[0022] in, To correct the signal amplitude, As a reference amplitude benchmark, The linear slope coefficient, This refers to the accumulated error. The function of increasing the accumulated error leads to an increase in the amplitude of the correction signal, ensuring that the compensation strength is dynamically adjusted according to the degree of deviation. By employing the historical error distribution statistical analysis method (parameters: battery cluster error history vector, sampling period window length), the adaptive setting of the function slope coefficient is achieved, and the slope correction coefficient is obtained; Furthermore, the central tendency and dispersion of the historical error distribution are quantified by mean drift and variance calculation, generating parameter correction values ​​for the linear negative correlation function to ensure that the mapping relationship conforms to the current system operating characteristics; By employing amplitude clipping and saturation protection (parameters: maximum allowable amplitude, minimum allowable amplitude), the amplitude of the correction signal is limited within its safe operating range, generating a protected amplitude output. This process suppresses amplitude peaks while preserving dynamic response capabilities, preventing systemic oscillations caused by sudden amplitude changes. By using a direction and amplitude binding encapsulation method, the inverted direction information and the amplitude information corrected by linear negative correlation mapping and historical distribution are integrated into a standardized correction signal data structure, forming a complete compensation signal output and providing a consistent input for the subsequent slow channel weight optimization module; By using the binding output processing method of the linear negative correlation function, the slope adaptive setting and amplitude clipping results of the previous step are transformed into a correction signal with the characteristics of directional hedging and intensity adjustment as needed, so as to achieve a precise compensation effect for systematic deviations. For example, in the DC combiner cabinet of a large energy storage power station, the cumulative error of a certain battery cluster is selected. Ah, the direction indicator is positive. The signal generated by the direction reversal mapping is in the negative direction. Reference amplitude standard. A. The slope coefficient is obtained from historical error statistics. A / Ah. Substitute into the formula to calculate the amplitude of the corrected signal:

[0023] get The initial value of amplitude A. After amplitude clipping and protection, it is limited to... A to Within range A, maintain output. Amplitude. Ultimately, the signal direction and amplitude are linked, forming a negative direction. The standardized correction signal of A is injected into the slow channel weight optimization. Operational verification shows that it significantly improves the long-term current sharing consistency of the system under this condition and effectively suppresses the circulating current phenomenon caused by the accumulation of deviation. S4.5: The generated correction signal is encapsulated into an error compensation vector, and a timestamp and battery cluster identification information are added to form a standardized error compensation data structure, which serves as the input parameter for the subsequent slow channel weight optimization module to support the dynamic adjustment of the schedulable priority of battery clusters.

[0024] Step S5: The correction signal is injected in reverse into the weight optimization process of the slow channel, dynamically adjusting the schedulable priority coefficient of the corresponding battery cluster in the next sliding window period. The adjustment magnitude of the priority coefficient is exponentially positively correlated with the amplitude of the correction signal, and the adjustment range is limited to the interval [0.8, 1.2]. Specifically, this includes: S5.1: Based on the systematic deviation components extracted from the error observation matrix, identify the battery cluster objects that need to be adjusted first, and extract their corresponding correction signal amplitude and direction information to determine the input basis for slow channel weight optimization; S5.2: Based on the linear negative correlation between the amplitude of the correction signal and the cumulative error, calculate the error compensation coefficient corresponding to each battery cluster. This error compensation coefficient serves as the input parameter for the slow channel schedulable priority adjustment. S5.3: Utilize the mapping relationship between the error compensation coefficient and the exponential function to generate an adjustment factor that is proportional to the adjustment range of the battery cluster priority coefficient. The range of variation of the adjustment factor is limited to the interval [0.8, 1.2] to prevent system oscillation caused by sudden changes in the control quantity. S5.4: The adjustment factor is weighted and fused with the baseline schedulable priority coefficient of the battery cluster in the current sliding window period to obtain the updated priority coefficient matrix, which serves as the initial scheduling weight for the next sliding window period of the slow channel. The input consists of the adjustment factor vector generated by S5.3 and the baseline schedulable priority coefficient matrix of each battery cluster in the current sliding window period; A matrix weighted fusion method (parameters: adjustment factor vector, benchmark priority coefficient matrix) is adopted to realize the cluster-weighted update process, and the adjustment factor is directly multiplied with the benchmark coefficient of the corresponding cluster to obtain the weighted coefficient matrix; Furthermore, by using a cluster-by-cluster normalization method (parameter: weighted coefficient matrix), the coefficient range is limited to ensure that all updated coefficients are strictly kept within the range of [0.8, 1.2], and to eliminate the risk of numerical overflow during the calculation process. Furthermore, by using an exponential mapping function fusion method (parameters: adjustment factor, benchmark coefficient), an exponential positive correlation between the adjustment magnitude and the adjustment factor value is achieved. Specifically, the following exponential mapping formula is applied based on the multiplication result:

[0025] in, As the baseline coefficient value, The offset after subtracting 1 from the adjustment factor is controlled by... The range of values ​​is used to ensure the interval restriction of the coefficient; Furthermore, by concatenating column vectors (parameter: cluster-by-cluster update coefficients), a complete updated priority coefficient matrix is ​​generated. The rows of the matrix correspond to different sampling periods, and the columns correspond to the cell cluster numbers, thereby achieving a unified update of the scheduling weights for the entire cluster. By using matrix weighted fusion and exponential mapping, the adjustment factor vector from the previous step is transformed into an updated priority coefficient matrix, thereby achieving dynamic adjustment and range control of the weights. For example, in an energy storage system containing four battery clusters, the baseline schedulable priority coefficient matrix for the current sliding window period is [1.00, 0.95, 1.05, 0.90], and the adjustment factor vector is [1.10, 0.85, 1.20, 0.95]. Matrix weighted fusion is performed to obtain a preliminary weighted coefficient matrix [1.10, 0.8075, 1.26, 0.855]. The result is normalized and the range is limited, truncating the element 1.26 (exceeding 1.2) to 1.2 and raising the element 0.8075 (below 0.8) to 0.8, resulting in a limited matrix [1.10, 0.80, 1.20, 0.855]. According to the exponential mapping formula, let α be the offset after subtracting 1 from the adjustment factor, taking the α value range [-0.2, 0.2], and calculate the corrected coefficients: for example, the coefficient of cluster 1 = 1.10 × =1.10× The value was approximately 1.2157, which was truncated to 1.20. The final updated priority coefficient matrix is ​​[1.20, 0.80, 1.20, 0.85]. This matrix is ​​used as the initial scheduling weight for the slow channel in the next sliding window cycle. System operation verification shows that the current sharing consistency is significantly improved, and the current distribution performance is stable without oscillation when switching between high and low loads. S5.5: Based on the updated priority coefficient matrix, the baseline current sharing weight of the slow channel output is recalculated to achieve dynamic optimization of the long-term equilibrium control strategy and improve the current sharing consistency and scheduling robustness of the system under long-term operation.

[0026] Step S6: Perform weighted fusion of fast and slow channel outputs based on the current system load rate: when the load rate is higher than 80%, a linear combination of 0.7 fast channel weight + 0.3 slow channel weight is used; when the load rate is lower than 30%, a combination of 0.3 fast channel weight + 0.7 slow channel weight is used. For intermediate load rates, the weight allocation ratio is determined using linear interpolation. Specifically, this includes: S6.1: Based on real-time system load rate data collected from the industrial control field, perform load rate zoning judgment processing to determine the current load range. Specifically, perform threshold comparison calculation on the load rate signal. If the load rate is greater than 80%, it is marked as a high load range; if it is less than 30%, it is marked as a low load range; otherwise, it is marked as a medium load range. Based on the real-time load rate signal of the energy storage system collected by the industrial control system, a multi-source signal fusion algorithm (parameters: sampling frequency ≥ 10Hz, fusion window length = 1s) is adopted to achieve stable acquisition and anti-interference processing of the original load rate signal. Furthermore, by using a bandpass filtering method (parameters: passband range 0.1Hz~2Hz), transient spikes and low-frequency drift components are suppressed, and a smoothed effective load rate sequence is obtained. Furthermore, using a threshold comparison algorithm (parameters: high load threshold = 80%, low load threshold = 30%), the effective values ​​of the load rate are classified into intervals, and corresponding interval label codes are generated. Furthermore, a hysteresis protection mechanism (parameters: hysteresis threshold = ±2% load rate, judgment period = 5 seconds) is introduced into the interval classification results to suppress short-term fluctuations in the load rate signal and obtain a stable load interval state confirmed by the protection mechanism. Furthermore, through interval coding mapping, the stable load interval state is transformed into a standardized interval identifier signal, which serves as the trigger input for subsequent fast and slow channel weight setting; Through the threshold comparison and hysteresis protection algorithm described above, the effective value of the load rate in the previous step is transformed into a high load, medium load or low load range identifier, so as to achieve precise triggering of the fast and slow channel weight allocation strategy. For example, in the test operation of a 500kW energy storage system, the load rate signal is output by the power acquisition module at a sampling frequency of 20Hz, and then weighted and averaged through a 1-second fusion window to obtain a smoothed load rate value. During the test period, the original load rate signal fluctuated between 78% and 85%, and the effective value stabilized at around 82% after bandpass filtering. The threshold comparison result classified this state as a high load range. The high load determination was confirmed by a hysteresis protection mechanism, with the hysteresis threshold set at ±2%. If the load rate remained between 80% and 84% for 5 consecutive seconds, the high load range identification signal "01" was finally output. This identification signal directly triggered the allocation strategy of fast channel weight 0.7 and slow channel weight 0.3, which significantly shortened the load response time and significantly improved the system's current sharing consistency under dynamic load fluctuations. S6.2: For high-load intervals, based on the system's dynamic response requirements, a fast-channel weighting coefficient setting process is performed to generate a weighting factor of 0.7. Specifically, a fixed-ratio allocation algorithm is used to multiply the fast-channel output by 0.7 and the slow-channel output by 0.3 to form an initial weighted allocation strategy under high-load conditions; For setting the fast channel weight coefficient under high load conditions, the input objects are the high load tag signal output by the load rate partition judgment module and the short-term current command and long-term current sharing weight data generated by the dual-channel controller. A fixed-ratio allocation algorithm (parameters: fast channel ratio set to 0.7, slow channel ratio set to 0.3) is adopted to implement a weight allocation strategy that prioritizes short-term dynamic response under high load conditions. Furthermore, by using a scaling factor multiplication operation (parameters: fast channel coefficient 0.7, slow channel coefficient 0.3), the short-term current command vector is scaled, and the weighted fast channel current vector result in the high load range is obtained. Furthermore, by using a proportional coefficient multiplication operation (parameters: fast channel coefficient 0.7, slow channel coefficient 0.3), the scaling process of the slow channel baseline current sharing weight matrix is ​​realized, and a weighted slow channel weight matrix for the high load range is generated. Furthermore, through element-wise fusion operations of weighted vectors and weighted matrices, the initial synthesis output of the two channels in the high-load range is achieved, and high-frequency response fusion data for load matching is generated. By using a fixed-ratio allocation algorithm, the scaling results of the previous step are transformed into dual-channel weighted data under high load, achieving the technical effect of prioritizing dynamic response performance under instantaneous power demand conditions. For example, in a certain energy storage system, the real-time load rate signal in the high load range is 0.92. The system determines that it has entered the high load mode. The fast channel outputs a short-term current command vector of [120, 118, 121]A, and the slow channel outputs a long-term reference weight matrix of [[1.0, 0.95, 1.02]].

[0027] Perform scaling on the fast channel vector: The weighted result is [84.0, 82.6, 84.7]A.

[0028] Perform scaling on the slow channel matrix: The weighted result is [0.3, 0.285, 0.306].

[0029] The weighted fast-channel current and the weighted slow-channel weight are synthesized according to the element-by-element fusion rule to obtain the high-load range synthesis result [84.3, 82.885, 85.006]. This result is used as the input for subsequent weighted fusion processing, which effectively maintains the current response speed under high power conditions, while taking into account some balance indicators, and significantly improves the dynamic control stability of the system under instantaneous large load impact. S6.3: For the low-load range, based on the battery cluster capacity consistency optimization objective, a slow channel weight coefficient setting process is performed to generate a weight factor of 0.7. Specifically, a fixed ratio allocation algorithm is adopted, multiplying the fast channel output by 0.3 and the slow channel output by 0.7 to form an initial weighted allocation strategy under low-load conditions; Based on the condition that the load rate determination result is in the low load range, the slow channel output is used as the core signal for capacity consistency optimization. The input data includes the benchmark current sharing weight matrix of each battery cluster after optimization in step S5 and the short-term current command sequence of the fast channel. A fixed-ratio allocation algorithm is used (parameter: low-load fast channel weight). Low-load slow channel weight This enables coefficient-based processing of the two-channel outputs; Furthermore, by using vector scalar multiplication (parameters: the coefficient matrix corresponds one-to-one with the elements of each channel's output vector), the slow channel output is multiplied by... The weighting factor and the fast channel output multiplied by The weighting factors are used to obtain the weighted channel result vectors; Furthermore, by weighted superposition operation (parameter: the fast channel weight value and the slow channel weight value of the corresponding branch are added one by one), an initial fused current allocation vector under low load condition is generated, where each element represents the target allocation current of the corresponding branch; The results are normalized (parameter: all vector elements divided by the total target current) to ensure that the sum of the weighted allocation ratios of each branch is 1. To maintain the feasibility and numerical stability of the current distribution strategy; By using a fixed-ratio allocation and normalization coupling algorithm, the original output of the two channels in the previous step is transformed into initial weighted current allocation data under low load conditions, so as to achieve the expected technical effect of prioritizing the consistency of capacity of each battery cluster under low load conditions. For example, during the monitoring period of a certain energy storage system, the real-time load rate signal is collected and determined as follows: (Unit: %), enter the low load interval decision branch. The system input slow channel baseline current sharing weight matrix is ​​[ , , , The fast-channel short-term current command sequence is [ , , , Ampere. Applying a fixed-ratio allocation algorithm, multiply each element of the fast-channel sequence by... get[ , , , Multiply each element of the slow channel weight matrix by the total target current of 200 amperes, then multiply by... get[ , , , The two are added together term by term to obtain the fusion vector. , , , After normalization, each term is divided by the sum of 200 to form the final proportional vector. , , , This example demonstrates that under low load conditions, increasing the slow channel weight can significantly improve the capacity balance among battery clusters and avoid long-term health state deviations caused by short-term power fluctuations. S6.4: For medium load ranges, dynamic weight allocation of fast and slow channels is performed based on a linear interpolation function of load rate to obtain a fusion ratio that matches the current load rate. Specifically, a linear interpolation algorithm is used to normalize the values ​​of load rates between 30% and 80%, and the corresponding weight coefficients of fast and slow channels are calculated to ensure that the weight allocation changes continuously and smoothly with the load rate. S6.5: Based on the above weighted allocation strategy, perform linear weighted fusion processing of the fast and slow channel output signals to generate the final current allocation command. Specifically, multiply the short-term current command of the fast channel and the reference current sharing weight of the slow channel by their respective weighting coefficients, then add them one by one to output the fused current command signal as the control input of the power regulation unit.

[0030] Step S7: The fused current distribution command is sent to the power regulation unit, and the deviation between the actual output of each branch and the command value is monitored in real time. When the deviation exceeds 5% for 10 seconds, the feedback optimization mechanism is activated: the model predictive control parameters of the fast channel are tuned online, and the tuning step size is positively correlated with the integral value of the deviation. Specifically, this includes: S7.1: Transmit the fused current distribution command output by the multi-timescale coupled controller to the power conditioning unit to drive the DC / DC converter of each battery cluster branch to perform the corresponding current output operation, so as to realize the physical execution of the command current; S7.2: Based on the actual output current data of each branch fed back by the power regulation unit, the deviation between the command value and the actual value is calculated in real time to obtain the current error sequence of each branch, which serves as the trigger input condition for the feedback optimization mechanism. S7.3: Perform time window sliding integration processing on the current error sequence. When the deviation of a certain branch continues to exceed 5% for 10 seconds, it is determined to be a valid deviation event, and a trigger signal is generated to activate the execution process of the feedback optimization mechanism. Based on the difference vector between the real-time output current and the command current of each branch fed back by the power regulation unit, a sliding window integration method (window length parameter: 10 seconds, sampling period parameter: 1 second) is used to realize the time domain integration and accumulation function of the current error sequence. Furthermore, by using an amplitude threshold comparison algorithm (parameter: set threshold to 5% of rated current), the algorithm determines whether the error amplitude at each time sampling point within the integration window exceeds the limit, and obtains the Boolean sequence data for exceeding the limit judgment. Furthermore, by using the interval continuity detection method (parameter: number of consecutive out-of-limit periods ≥ window length), the continuity analysis of the Boolean sequence of out-of-limit judgment is realized, and a persistent out-of-limit identifier vector is generated as an event triggering condition. Furthermore, an event-triggered generation logic is adopted (parameter: persistent over-limit flag is true) to generate a trigger signal from the continuous analysis results, and the trigger signal is encapsulated into a standardized control event structure, which includes a trigger timestamp, branch number and event type code; By using the above-mentioned sliding window integration and persistent over-limit event triggering method, the error detection result of the previous step is transformed into the activation condition of the feedback optimization mechanism, thereby realizing the precise activation of the online tuning process of fast-channel model predictive control. For example, in the operation scenario of a DC combiner cabinet in an energy storage system, the sampling period is set to 1 second, the sliding window length is set to 10 seconds, the rated current of each branch is 200A, and the error threshold is set to 5% of the rated current, i.e., 10A. During a certain operation, the instantaneous error value of the No. 3 battery cluster branch was in the range of [10.5A, 12.3A] within 12 consecutive second-level sampling periods. Using sliding window integration, the window integration result was 118.6A·s, and the average deviation value was 11.86A, which was higher than the threshold of 10A. After continuous detection, the proportion of exceeding the limit for all sampling periods was 100%, which met the triggering condition and generated a control event structure containing branch number = 3, event type code = FEEDBACK_TRIGGER, and trigger timestamp. After this event was input to the feedback optimization module, the fast-channel MPC controller performed online tuning of the gain coefficient from 1.00 to 1.12, which significantly accelerated the dynamic load response of the system and significantly improved the current sharing effect in subsequent operation. S7.4: Based on the integral value of the deviation, the online tuning operation of the model predictive control parameters of the fast channel is performed. The proportional-integral adjustment algorithm is used to dynamically adjust the control gain and weight coefficient of the predictive model, so that the output of the fast channel is closer to the actual load demand. S7.5: The tuned model predictive control parameters are updated to the fast-channel controller after closed-loop verification to improve the system's response accuracy and control robustness under dynamic load changes, thereby achieving closed-loop compensation and performance optimization for current distribution errors.

[0031] Step S8: Periodically detect the singular value distribution of the error observation matrix. When the ratio of the largest singular value to the second largest singular value exceeds 10:1, it is determined that there is a structural mismatch in the system. At this time, the parameter self-update process of the multi-timescale coupled controller is initiated. The update content includes the sliding window length, Kalman filter covariance matrix, and hysteresis threshold. Specifically, it includes: S8.1: Perform singular value decomposition (SVD) operation based on the error observation matrix to obtain its singular value sequence, where the singular value sequence is arranged in descending order, with the largest singular value corresponding to the first principal component, the second largest singular value corresponding to the second principal component, and so on. S8.2: Calculate the ratio of the largest singular value to the second largest singular value in the singular value sequence to obtain the system structural mismatch criterion. The criterion is defined as the ratio of the largest singular value to the second largest singular value. The higher the ratio, the closer the error observation matrix is ​​to the rank-one matrix, and the greater the possibility of system structural mismatch. S8.3: Determine whether the ratio of the largest singular value to the second largest singular value exceeds the preset threshold of 10:1. If it does, it is determined that there is a structural mismatch in the system and the parameter self-update process is triggered. Otherwise, the current controller parameters remain unchanged. The input conditions include the maximum and second largest singular values ​​obtained by the S8.2 sub-step. These values ​​are derived from the singular value decomposition results of the error observation matrix and are expressed as dimensionless ratios. A ratio calculation method (parameters: maximum singular value, second largest singular value) is used to realize the numerical expression of the structural mismatch criterion and generate ratio results; Furthermore, by using a threshold comparison algorithm (parameter: threshold=10), the interval determination function of the comparison value results is realized, and the determination signal is obtained; Furthermore, by using a logic gating method (parameter: judgment signal type = Boolean true / false), the matching and verification between the judgment signal and the controller parameter update triggering conditions are realized, and the triggering decision result is generated; Furthermore, through conditional branch control processing (parameter: trigger decision result), the selection between the controller parameter update process and the hold process is realized, and the process identification signal is output; The ratio of the largest singular value to the second largest singular value is calculated using the following formula:

[0032] in, For the maximum singular value, It is the second largest singular value. This is the ratio result. A threshold comparison algorithm is used to... and If a comparison is made, ≥ If the structural mismatch criterion is met, then the criterion is valid; otherwise, the criterion is invalid. The above algorithm transforms the singular value decomposition result of the previous step into a mismatch judgment signal, thereby enabling accurate determination of the trigger condition for the self-update of controller parameters. For example, during the operating cycle of a certain energy storage system, the maximum singular value obtained by the singular value decomposition of the error observation matrix is... The second largest singular value The ratio calculation formula is used. , can be obtained By using a threshold comparison algorithm, With threshold Compare and satisfy ≥ Under the given conditions, the logic gating method outputs a Boolean "true" signal. The conditional branch control processing unit selects to execute the controller parameter update process based on this signal, passing process identification signals to subsequent sub-steps S8.4, S8.5, and S8.6. In this scenario, the initiation of the parameter self-update process improves the system's adaptability to structural mismatches in this cycle, allowing for dynamic adjustment of the sliding window length, Kalman filter covariance matrix, and hysteresis threshold, significantly improving the stability of current sharing control across multiple time scales. S8.4: Based on the structural mismatch state of the system, the parameter self-update process of the multi-timescale coupled controller is initiated. First, the window length of the sliding window regression model is dynamically adjusted. The adjustment is based on the variance characteristics of the current error distribution. The larger the variance, the longer the window length is to enhance the trend recognition capability. S8.5: Perform online correction on the covariance matrix in the extended Kalman filter. The correction is based on the historical statistical characteristics of the current error residual sequence, including mean drift and variance changes. The corrected covariance matrix is ​​used to improve the robustness of error estimation. S8.6: Dynamically adjust the hysteresis logic threshold based on the degree of structural mismatch in the system. The higher the degree of mismatch, the higher the threshold value, in order to prevent false triggering of compensation actions and ensure the stability and controllability of the parameter update process.

[0033] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0034] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and rules of the present invention should be included within the scope of protection of the present invention.

Claims

1. A smart current sharing control method for a DC combiner cabinet in an energy storage system, characterized in that, Includes the following steps: S1: Real-time acquisition of output current, voltage and circuit breaker status parameters of each battery cluster branch in the DC combiner cabinet of the energy storage system, and synchronous acquisition of SOC and temperature data provided by the battery management system to form a multi-dimensional operating status dataset. S2: Based on the multi-dimensional operating status data, a dual-channel status prediction structure including a fast channel and a slow channel is constructed. The fast channel generates a short-term current command, and the slow channel outputs a benchmark current sharing weight. S3: Calculate the error residual sequence between the actual discharge depth and the predicted discharge depth of each battery cluster, use extended Kalman filtering to recursively estimate the residual dynamics, extract the systematic deviation components and generate the error observation matrix; S4: When the cumulative error of a battery cluster in the error observation matrix exceeds a preset threshold, the hysteresis logic threshold mechanism is triggered to generate a correction signal opposite to the error direction; S5: Inject the correction signal in reverse into the weight optimization process of the slow channel to dynamically adjust the schedulable priority coefficient of the corresponding battery cluster in the next sliding window period. S6: Based on the current system load rate, perform weighted fusion of fast and slow channel outputs to generate a fused current distribution command; S7: The fused current distribution command is sent to the power regulation unit, and the deviation between the actual output of each branch and the command value is monitored in real time. When the deviation exceeds the preset deviation threshold and the duration reaches the preset time threshold, the feedback optimization mechanism is activated to adjust the model prediction control parameters of the fast channel online.

2. The intelligent current sharing control method for a DC combiner cabinet in an energy storage system according to claim 1, characterized in that, The process following step S7 also includes: S8: Periodically detect the distribution of singular values ​​in the error observation matrix. When the ratio of the largest singular value to the second largest singular value exceeds the preset singular value ratio threshold, it is determined that there is a structural mismatch in the system, and the parameter self-update process of the multi-timescale coupled controller is initiated.

3. The intelligent current sharing control method for a DC combiner cabinet in an energy storage system according to claim 1, characterized in that, Step S1 specifically includes: Based on a high-precision current sensor and voltage sampling module, the output current and voltage of each battery cluster branch are synchronously acquired to obtain the original current and voltage data sequence. The on / off status of each branch circuit breaker in the DC combiner cabinet is monitored in real time using a digital input module to obtain the circuit breaker status signal; A communication connection is established with the battery management system through the CAN bus and Modbus protocol interface, and the SOC and temperature data of each battery cluster are periodically acquired based on a predefined data frame format. The original current and voltage data sequence, the circuit breaker status signal, and the SOC and temperature data are timestamped to construct a multi-source heterogeneous data set under a unified time reference. The validity of the multi-source heterogeneous data set is verified based on the data integrity verification mechanism. If the data of a certain battery cluster branch is found to be missing or abnormal, the data repair process is triggered to form a multi-dimensional operating status dataset with complete structure and controllable quality.

4. The intelligent current sharing control method for a DC combiner cabinet of an energy storage system according to claim 3, characterized in that, The high-precision current sensor is a high-precision Hall effect sensor with a range of ±500A and linearity better than 0.5%FS; the voltage sampling module is a precision voltage divider type voltage sampling module with an insulation withstand voltage greater than or equal to 1500V and a sampling error less than or equal to 0.5%FS.

5. The intelligent current sharing control method for a DC combiner cabinet in an energy storage system according to claim 1, characterized in that, Step S2 specifically includes: The real-time collected output current and voltage data of each battery cluster are normalized to obtain a standardized current and voltage vector with unified dimensions. Based on the standardized current and voltage vector, a model prediction control framework with a sampling frequency of 1 second is constructed. A rolling optimization strategy is used to generate short-term current commands, wherein the optimization objective function includes minimizing the rate of change of current commands and power tracking error, and outputs a fast-channel current prediction sequence. Feature extraction is performed on the SOC degradation trend and temperature accumulation effect parameters provided by the battery management system. The long-term capacity degradation behavior is modeled using a sliding window regression model, and the health status trend feature vector of the battery cluster is extracted. Based on the trend feature vector of the battery cluster health status, a sliding window regression model is performed, the weight parameters are optimized using the least squares method, and the slow channel baseline average current weight matrix is ​​output and updated on a minute-by-minute basis. The fast-channel current prediction sequence and the slow-channel reference current sharing weight matrix are time-aligned to construct a fusion interface for the dual-channel state prediction structure, generating a joint state prediction vector that includes short-term dynamics and long-term trends.

6. The intelligent current sharing control method for a DC combiner cabinet in an energy storage system according to claim 1, characterized in that, Step S3 specifically includes: Based on the real-time discharge current and voltage data of each battery cluster, combined with the SOC status information, an ampere-hour integration and open-circuit voltage correction fusion algorithm is executed to obtain a high-precision actual depth of discharge value. Based on the short-term current command output by the fast-channel model predictive control framework and the benchmark current sharing weight output by the slow-channel sliding window regression model, the theoretical depth of discharge value of each battery cluster at the current time scale is calculated. The actual discharge depth value and the theoretical discharge depth value are calculated point by point to generate an error residual sequence; Based on the error residual sequence, a state-space model of the extended Kalman filter is constructed; The error residual sequence is recursively updated and covariance corrected using the state-space model of the extended Kalman filter, and the systematic deviation components of each battery cluster are extracted to form a deviation estimation vector. The deviation estimation vectors of each battery cluster are arranged according to the branch number to construct the error observation matrix.

7. The intelligent current sharing control method for a DC combiner cabinet in an energy storage system according to claim 6, characterized in that, Step S3 further includes using the ampere-hour integration method to calculate the depth of discharge, with a sampling frequency of not less than 10Hz, the integration window synchronized with the fast channel, and correction performed in conjunction with the open-circuit voltage-capacity curve, and the actual integration result and the open-circuit voltage derivation result fused together by weighting coefficients.

8. The intelligent current sharing control method for a DC combiner cabinet in an energy storage system according to claim 1, characterized in that, Step S4 specifically includes: For each row of data in the error observation matrix, cluster-by-cluster error accumulation is performed. The cumulative error of each battery cluster in three consecutive sampling periods is extracted. The error sequence is weighted and summed using the sliding window integration method to obtain the cumulative error vector. Based on the error accumulation vector, an error direction consistency analysis is performed. The error direction within each sampling period is extracted by the sign function, and the percentage of error direction consistency within three consecutive periods is calculated. The consistency of direction is judged by whether the output value of the sign function is consistent. If the percentage of error direction consistency reaches 90% or more in three consecutive periods, it is determined that the battery cluster has a persistent prediction deviation. When the error direction consistency of a battery cluster meets 90% and its cumulative error exceeds the preset error threshold, the hysteresis logic threshold mechanism is triggered. The error threshold is set jointly based on the rated capacity of the battery cluster and the standard deviation of the SOC estimation error. Based on the triggered hysteresis logic threshold event, a correction signal opposite to the error direction is generated; The correction signal is encapsulated into an error compensation vector, and a timestamp and battery cluster identification information are added to form a standardized error compensation data structure.

9. The intelligent current sharing control method for a DC combiner cabinet of an energy storage system according to claim 8, characterized in that, The amplitude of the correction signal is mapped to the cumulative error through a linear negative correlation function. The larger the error, the stronger the amplitude of the correction signal. The slope of the linear negative correlation function is set according to the statistical results of the historical error distribution of the system.

10. The intelligent current sharing control method for a DC combiner cabinet in an energy storage system according to claim 1, characterized in that, Step S5 specifically includes: Based on the systematic deviation components extracted from the error observation matrix, the battery cluster objects that need to be adjusted first are identified, and their corresponding correction signal amplitude and direction information are extracted. Based on the linear negative correlation between the amplitude of the correction signal and the cumulative error, the error compensation coefficient corresponding to each battery cluster is calculated. Using the mapping relationship between the error compensation coefficient and the exponential function, an adjustment factor proportional to the adjustment range of the battery cluster priority coefficient is generated; The adjustment factor is weighted and fused with the baseline schedulable priority coefficient of the battery cluster within the current sliding window period to obtain the updated priority coefficient matrix; Based on the updated priority coefficient matrix, the baseline current sharing weight of the slow channel output is recalculated.

Images