Method for estimating SOC (State of Charge) of high-capacity energy storage battery based on data-driven coulomb method

Abstract

The invention discloses a high-capacity energy storage battery SOC estimation method based on a data-driven coulomb method, and the method comprises the following steps: S101, constructing an incremental capacity curve based on an SOC axis through employing the voltage and capacity data collected in the charging and discharging processes of a battery; s102, respectively extracting initial SOC calibration points in the charging and discharging processes, and determining a calibration SOC value by using a specific dQ / dV value; s103, aiming at the error problem of the actual capacity, extracting health indexes from the voltage-time curve by adopting a maximum information coefficient method; and S104, constructing a Gaussian process regression model by taking the health indexes in the S103 as input. According to the method for estimating the SOC of the high-capacity energy storage battery based on the data-driven coulomb method, physical feature extraction and statistical learning are combined: intrinsic differential features of a battery charging and discharging curve are analyzed by utilizing incremental capacity analysis (ICA), a voltage platform is converted into a stable calibration point in an SOC domain, and rapid self-correction of an initial error is realized; and meanwhile, the implicit relevance between the voltage time sequence data and the capacity fading is mined through a maximum information coefficient (MIC).

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of high-energy-density energy storage lithium-ion battery management, in particular to a method for SOC estimation of large-capacity energy storage battery based on data-driven Coulomb method. BACKGROUND

[0002] Robust estimation of the state of charge of energy storage batteries is a core technology to ensure the safe and efficient operation of battery systems, and the difficulty lies in the strong nonlinearity of battery characteristics, parameter time variation (such as aging and temperature influence) and measurement noise interference.

[0003] In the prior art, a lithium-ion energy storage battery SOC online estimation method is disclosed in Chinese patent No. CN202010544392.3, which comprises obtaining the rated parameters of the lithium-ion energy storage battery to be detected, and establishing an equivalent circuit model of the lithium-ion energy storage battery to be detected; online identification of the model parameters of the established equivalent circuit model; according to the established equivalent circuit model and the model parameters obtained by online identification, an improved adaptive extended Kalman filter for lithium-ion energy storage battery is established; the model parameters obtained by online identification are input into the improved adaptive extended Kalman filter for lithium-ion energy storage battery, and the online estimation of the SOC of the lithium-ion energy storage battery is carried out.

[0004] In the combined technology, a first-order RC equivalent circuit model can be used, but in the actual use process, multiple parameter dynamic identification needs to be carried out through the forgetting factor recursive least square method and the simulated annealing algorithm, which leads to a significant increase in model complexity. The first-order RC model needs to identify the ohmic internal resistance, polarization capacitance / resistance parameters at the same time, which leads to an increase in calculation dimension. Although the forgetting factor recursive least square method introduces a forgetting factor, the forgetting characteristic further increases the time consumption of calculation. In addition, the adaptive extended Kalman filter needs to iterate the covariance matrix multiple times in state estimation, which challenges the computing power of the embedded BMS and is difficult to meet the real-time working condition requirements.

[0005] Moreover, the accuracy of the first-order RC model is highly dependent on the parameter identification results, and battery aging will lead to nonlinear evolution of polarization characteristics. The traditional FFRLS algorithm is difficult to track parameter mutations, and there is a parameter convergence lag problem. In addition, when the temperature fluctuates sharply, the RC model parameters and the measured dynamic response are mismatched, and a large number of calibration experiments need to be relied on for compensation, which has poor engineering adaptability. SUMMARY

[0006] The purpose of the present application is to provide a method for SOC estimation of large-capacity energy storage battery based on data-driven Coulomb method to solve the problem of large computing power demand and poor adaptability in the background technology.

[0007] To achieve the above-mentioned purpose, the present application provides the following technical scheme: a method for SOC estimation of large-capacity energy storage battery based on data-driven Coulomb method, comprising the following steps:

[0008] S101, constructing an incremental capacity curve based on the SOC axis using voltage and capacity data collected during battery charging and discharging;

[0009] S102, extracting initial SOC calibration points during the charging and discharging processes, respectively, and determining a calibration SOC value using a specific dQ / dV value;

[0010] S103. To address the error problem of actual capacity, the maximum information coefficient method is used to extract health indicators from the voltage-time curve;

[0011] S104. Using the health indicator described in S103 as input, a Gaussian process regression model is constructed to perform online estimation of the actual capacity. The parameters that have been calibrated with both the initial SOC and the actual capacity are applied to the coulomb counting method to achieve accurate estimation of the battery SOC.

[0012] Preferably, in the step of extracting initial SOC calibration points during the charging and discharging processes and calibrating the initial SOC value based on the incremental capacity analysis of the SOC axis in S102, this step reconstructs the coordinate system of the incremental capacity analysis, establishes a mapping relationship between SOC and dQ / dV, and implements rapid calibration of the initial SOC based on preset feature points. The specific implementation process is as follows:

[0013] S201, data acquisition: During the battery charge / discharge process, the battery terminal voltage V(t), current I(t) and timestamp t are collected in real time at a fixed sampling frequency, and the cumulative capacity is calculated by integration

[0014] S202, SOC calculation: Preliminary estimate of SOC value based on the coulomb counting method, the formula is:

[0015]

[0016] Where C is the nominal capacity of the battery, and SOC initial is the uncalibrated initial value;

[0017] S203, dQ / dV calculation: Use the differential method to calculate the incremental capacity value. Extract discrete data pairs of capacity and voltage from the charge and discharge process, and use the five-point central difference method to calculate the instantaneous dQ / dV value. The formula is:

[0018]

[0019] S204. Coordinate transformation: Reconstruct the traditional dQ / dV curve with voltage as the horizontal axis into a curve with SOC as the horizontal axis. The specific method is to replace the voltage axis with the SOC axis, calculate the SOC in real time by integrating the current-time data, generate the dQ / dSOC curve, and form a two-dimensional data set of SOC(t) and dVdQ(t).

[0020] Preferably, in the step of extracting multiple health indicators in step S103, the optimal health indicators are screened by quantifying the nonlinear correlation between the voltage curve segment and the battery capacity, and in the charging process, the voltage values at the initial stage of charging are collected from the battery voltage-time curve, a total of 9 sampling points are obtained, which are marked as HI1 to HI9, and the cumulative charging capacity at the corresponding time is recorded; in the discharging process, the voltage data is collected in a suitable time region to form another set of health indicator data, and the specific implementation process is as follows:

[0021] S301, formulate the extraction rule of candidate HIs, the time window selection rule of His in the charging stage is to extract 9 voltage values within 5-45 seconds after the start of constant current charging with an interval of 5 seconds, marked as HI1 to HI9, and the voltage sequence is standardized, the formula is:

[0022]

[0023] Wherein μV is the mean value, and σV is the standard deviation;

[0024] S302, MIC correlation analysis, the MIC calculation principle is to form a two-dimensional data set {(HIi,Cj)} by combining HI in the battery historical data with the capacity C;

[0025] S303, robustness enhancement design, verify by sliding time window, randomly select multiple time windows in the historical data, verify the MIC value stability of HIs, if the MIC value difference of two HIs is less than 5% and the time is close, then the latter is removed to reduce the dimension.

[0026] Preferably, in step S104, the Gaussian process regression model is constructed to estimate the actual capacity online, the mapping relationship between HIs and battery capacity is established by a non-parametric probability model, and the specific implementation process is as follows:

[0027] S401, model construction, first define the input and output of the model, solve the covariance function, and finally optimize the hyperparameters and solve;

[0028] S402, online capacity estimation, after the initial SOC and the actual capacity are corrected by the above steps, the traditional coulomb counting method is modified to realize accurate tracking of SOC;

[0029] S403, online SOC estimation result verification and feedback, to verify the effectiveness of the modified coulomb counting method in actual application.

[0030] Preferably, the specific steps of step S403 are as follows:

[0031] S403-1, charging process verification, set initial conditions: the actual initial SOC is 0%, but intentionally set an error initial SOC in the system to simulate possible sensor errors in practice; when charging starts, the system captures the calibration point dQ / dV = 2.03 within 3 seconds, automatically corrects the SOC to 0.13%; during the subsequent charging process, the system calculates the SOC based on the correction parameters, the results show that it is very close to the actual SOC, the error is controlled within 0.05%; at the same time, compared with the uncorrected SOC estimation, it is found that the cumulative error increases continuously and finally exceeds 125%, which proves the necessity of the correction method;

[0032] S403-2, discharge process verification, the actual discharge initial SOC is 100%, but the system pre-sets an error initial SOC; in the early stage of discharge, when the system detects dQ / dV = 6.08, it automatically corrects the SOC to 99.5%; during the subsequent discharge process, the actual capacity value is updated in real time for SOC calculation, the results show that the SOC decline trajectory is basically consistent with the true curve, the error is controlled within 0.25%; compared with the uncorrected SOC, it is found that it may eventually appear negative;

[0033] S403-3, feedback correction mechanism, the system has a self-checking module, which periodically detects the error of the SOC curve; when the error exceeds the preset range for consecutive multiple sampling periods, the secondary correction process is automatically triggered to recalculate the calibration point and the actual capacity; the feedback result will be used for fine-tuning of the subsequent SOC calculation parameters, to ensure that the error accumulation is always at a low level in long-term operation.

[0034] Preferably, in the step S101 of collecting voltage and capacity data of the battery charging and discharging process, two lithium ion batteries with the same nominal capacity are used; a high-precision current and voltage acquisition module and a data recording system are provided, which have online sampling and real-time calculation functions; in a constant temperature room, the CC-CV mode is used for charging, the charging current is 1C, and when the battery terminal voltage reaches 4.2V, it is converted to constant voltage charging until the charging current decreases to 0.1C; the discharge process uses CC mode, the discharge current is 3C, and the cutoff voltage is set to 2.75V; 30 minutes of standing recovery time is set after each cycle to ensure stable battery state, and the matrix D = [V, I, t] is constructed, wherein V ∈ RN×1 is the voltage sequence, I ∈ RN×1 is the current sequence, t ∈ RN×1 is the time stamp, and N is the number of data points.

[0035] Preferably, the target dQ / dV value of the calibration point in the charging stage in step S203 is set to 2.03, corresponding to the starting point of the phase change platform of the graphite negative electrode of the lithium-ion battery. In the early stage of charging, the dQ / dSOC curve fluctuates violently, but the specific dQ / dSOC value corresponds to the SOC with high repeatability. The SOC matching logic is to traverse the charging sub-interval, extract the SOC value that satisfies dVdQ∈[2.03-∈,2.03+∈], where ∈ is the tolerance threshold, and calculate the arithmetic mean thereof as the calibration point SOC calibration=0.13%. The target dQ / dV value of the calibration point in the discharge stage is set to 6.08, corresponding to the end point of the delithiation phase change platform of the positive electrode material. The SOC matching logic is to extract the SOC value that satisfies dVdQ∈[6.08-∈,6.08+∈] within the discharge sub-interval, and calculate the arithmetic mean thereof as the calibration point SOC calibration=99.5%;

[0036] A data-driven approach is used to calibrate the actual capacity, and the changes in the voltage-time curve during the charge / discharge process are used to extract health indicators reflecting the battery degradation state. The maximum information coefficient method is used to analyze the correlation between each health indicator and the actual capacity, and the indicators with higher MIC values ​​are selected. Then, using the selected health indicators as input, a Gaussian process regression model is constructed to achieve online estimation of the actual capacity. The capacity parameters used in the coulomb counting method are corrected to ensure the accuracy of the SOC calculation.

[0037] In the initial SOC correction mechanism, the trigger condition is when the BMS detects that the current dQ / dV value enters the preset range, activating the calibration process. The dynamic update uses a sliding window algorithm to filter the SOC calibration point to eliminate noise interference. The formula is:

[0038]

[0039] Preferably, in step S204, a Savitzky-Golay filter is used, the default window width is set to 11 points, the default polynomial order is set to 3, the dQ / dSOC curve is smoothed to eliminate high-frequency noise, and the phase change peak in the curve is identified by a peak detection algorithm. The formula is:

[0040]

[0041] Preferably, the time window selection rule of the discharge stage HIs in step S301 is to extract 9 voltage values ​​at intervals of 25 seconds within 425-625 seconds before the end of the constant current discharge, and record them as HI1 to HI9.

[0042] Preferably, the correlation coefficient between each health indicator and the actual capacity is calculated in step S302 using the MIC algorithm; for the charging stage, the MIC values of the battery health indicators are obtained by calculating the data respectively, and the average of the two is taken; according to the average MIC value, the four indicators with the highest MIC values are selected as the input features for subsequent actual capacity estimation; for the discharging stage, the four indicators with the highest MIC values are also selected as the input according to the MIC value sorting; the dynamic grid division algorithm is used to calculate the normalized mutual information value:

[0043]

[0044] Where n G ,m G are the row and column numbers of the grid G, I is the mutual information function, and the screening rule is to calculate the MIC value of each HI, arrange in descending order, and select the top 4 HIs with the largest MIC values as the model input.

[0045] Compared with the prior art, the method for estimating the SOC of a large-capacity energy storage battery based on a data-driven Coulomb method has the beneficial effects that the specific content is as follows:

[0046] 1. Combining physical feature extraction with statistical learning: using incremental capacity analysis (ICA) to analyze the intrinsic differential features of the battery charge-discharge curve, converting the voltage platform into a stable calibration point in the SOC domain, and realizing fast initial error self-correction; at the same time, the maximum information coefficient (MIC) is used to mine the implicit correlation between the voltage time series data and the capacity decay, and a Gaussian process regression model is constructed to directly map the health state, which structurally decouples the strong coupling chain of parameter identification and state estimation in traditional methods. This de-modeling design greatly simplifies the calculation path, reduces the algorithm complexity from O(n 3 ) to O(n), and functionally realizes millisecond-level real-time response, providing essential optimization for high dynamic conditions.

[0047] 2. Establishing a dynamic and adaptive SOC estimation system through a data-driven mechanism: the differential feature points (such as dQ / dV extreme values) extracted by ICA essentially correspond to the phase transition critical state of the electrode material, and their SOC coordinates are extremely sensitive to aging, which guarantees the cross-period stability of the initial calibration from the physical mechanism level; MIC analysis captures health factors that are strongly related to capacity decay from the voltage relaxation process, and a probabilistic prediction model is established through Gaussian process regression, which can autonomously learn the nonlinear relationship of the aging trajectory. This double-layer architecture of "feature extraction + statistical inference" breaks through the rigid constraints of fixed models, maintains the consistency of estimation accuracy throughout the entire life cycle of the battery, and solves the problem of error accumulation caused by model mismatch in traditional methods.

[0048] 3. Structural fault tolerance is improved through a dual dynamic calibration mechanism. In the initial SOC calibration stage, the differential characteristics of the ICA feature points give it a natural filtering effect on voltage measurement noise. Its mathematical essence is to lock the SOC benchmark by maximizing the local capacity change rate, and the noise is significantly suppressed in the differential operation. In the capacity prediction stage, the health factor screened by MIC is constructed based on statistical correlation, and Gaussian process regression autonomously weights effective features through the covariance function, and has probabilistic robust processing capabilities for outliers. Compared with the traditional method that relies on the post-compensation strategy of H∞ robust control theory, this solution functionally eliminates the error transmission path from the source through the collaborative design of data preprocessing and model structure, realizes the full-process anti-interference optimization of the "noise-feature-estimation" link, and provides systematic protection for complex working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 Schematic diagram of a flow chart of a method for estimating SOC of a large-capacity energy storage battery based on a data-driven coulomb counting method according to the present invention;

[0050] Figure 2 The fast calibration process and key method diagram of the battery SOC based on the data-driven coulomb counting method of the present invention;

[0051] Figure 3 A schematic diagram of the health indicator analysis process of the data-driven coulomb counting method of the present invention;

[0052] Figure 4 Schematic diagram of Gaussian process regression model prediction for the data-driven coulomb counting method of the present invention. DETAILED DESCRIPTION

[0053] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0054] Example 1: Please refer to Figure 1 , this embodiment includes the following experiments:

[0055] Experimental conditions:

[0056] Experimental environment and data collection

[0057] Test subjects: Two commercial ternary lithium-ion batteries (numbers #1 and #2), each with a nominal capacity of 2500mAh.

[0058] Device configuration:

[0059] High-precision battery testing system (current accuracy ±0.05% FS, voltage accuracy ±0.02% FS);

[0060] Constant temperature box (temperature controlled at 25±0.5℃);

[0061] Real-time data recording system (sampling frequency 1Hz);

[0062] Charge and discharge protocol:

[0063] Charge:

[0064] Constant current (CC) stage: 1C (2500mA) charging until the voltage reaches 4.2V;

[0065] Constant voltage (CV) stage: maintain 4.2V until the current drops to 0.1C (250mA);

[0066] Discharge: Constant current (CC) 3C (7500mA) discharge to cut-off voltage 2.75V

[0067] Rest: Rest for 30 minutes after each cycle.

[0068] Experimental steps:

[0069] S101, constructing an incremental capacity (IC) curve based on the SOC axis using voltage and capacity data collected during battery charging and discharging;

[0070] S102, extracting initial SOC calibration points during the charging and discharging processes, respectively, and determining a calibration SOC value using a specific dQ / dV value;

[0071] S103. To address the error problem of actual capacity, the maximum information coefficient (MIC) method is used to extract multiple health indicators (HIs) from the voltage-time curve;

[0072] S104 uses the health indicators from S103 as input to construct a Gaussian process regression (GPR) model to estimate the actual capacity online. The parameters, which have been calibrated using both the initial SOC and the actual capacity, are then applied to the coulomb counting method to accurately estimate the battery SOC.

[0073] To collect voltage and capacity data during the battery charge and discharge process, the S101 uses two lithium-ion batteries with the same nominal capacity (hereinafter referred to as Battery #1 and Battery #2). The system is equipped with a high-precision current and voltage acquisition module and data logging system, providing online sampling and real-time calculation capabilities. In a constant-temperature chamber, charging is performed using CC-CV (constant current, constant voltage) mode with a charge current of 1C. When the battery terminal voltage reaches 4.2V, constant voltage charging is switched until the charge current drops to 0.1C. Discharge is performed using CC (constant current) mode with a discharge current of 3C and a cutoff voltage set at 2.75V. A 30-minute rest recovery period is provided after each cycle to ensure battery stability. A matrix D = [V, I, t] is constructed, where V∈RN×1 represents the voltage sequence, I∈RN×1 represents the current sequence, t∈RN×1 represents the timestamp, and N represents the number of data points.

[0074] Example 2: Please refer to Figure 2 , this embodiment includes the following experiments:

[0075] Experimental environment and data collection

[0076] Test object: ternary lithium-ion battery (nominal capacity 50Ah, NMC532 / graphite system)

[0077] Environmental control: constant temperature of 25°C, sampling frequency of 1Hz;

[0078] Charge and discharge protocol:

[0079] Charging: 1C constant current to 4.2V → constant voltage to current ≤ 0.05C

[0080] Discharge: 1C constant current to 2.8V (simulating electric vehicle driving conditions)

[0081] In S102, initial SOC calibration points are extracted during the charging and discharging processes, and the initial SOC value is calibrated based on the incremental capacity analysis of the SOC axis. This step reconstructs the coordinate system of the incremental capacity analysis (ICA), establishes the mapping relationship between SOC and dQ / dV, and realizes rapid calibration of the initial SOC based on the preset feature points. Figure 2 The specific implementation process is as follows:

[0082] S201, data acquisition: During the battery charge / discharge process, the battery terminal voltage V(t), current I(t) and timestamp t are collected in real time at a fixed sampling frequency (such as 1Hz), and the cumulative capacity is calculated by integration

[0083] S202, SOC calculation: Preliminary estimate of SOC value based on the coulomb counting method, the formula is:

[0084]

[0085] Where C is the nominal capacity of the battery, and SOC is the initial value that is not calibrated (default 0% or 100%).

[0086] S203, dQ / dV calculation: Use the differential method to calculate the incremental capacity value. Extract discrete data pairs of capacity (Q) and voltage (V) from the charge and discharge process, and use the five-point central difference method to calculate the instantaneous dQ / dV value. The formula is:

[0087]

[0088] S204, coordinate transformation: Reconstruct the traditional dQ / dV curve with voltage as the horizontal axis into a curve with SOC as the horizontal axis. The specific method is to replace the voltage axis with the SOC axis, calculate the SOC in real time by integrating the current-time data, and generate the dQ / dSOC curve. A two-dimensional data set of SOC(t) and dVdQ(t) is formed. A Savitzky-Golay filter (the default window width is set to 11 points, and the default polynomial order is set to 3) is used to smooth the dQ / dSOC curve and eliminate high-frequency noise. Subsequently, the phase change peak in the curve is identified by a peak detection algorithm (local maximum threshold method). The formula is:

[0089]

[0090] In the calibration point extraction rule in S203, the target dQ / dV value for the charging phase calibration point is set to 2.03, corresponding to the starting point of the phase change plateau in the graphite negative electrode of lithium-ion batteries. During the initial charging period (SOC 0-5%), the dQ / dSOC curve exhibits significant fluctuations, but the correlation between a specific dQ / dSOC value and SOC is highly repeatable.

[0091] The SOC matching logic traverses the charge subrange (e.g., SOC 0-5%), extracts SOC values ​​that satisfy dVdQ∈[2.03-∈, 2.03+∈] (∈ is a tolerance threshold, e.g., 0.01), and calculates their arithmetic mean as the calibration point SOCCalibration = 0.13%. The target dQ / dV value for the discharge calibration point is set to 6.08, corresponding to the endpoint of the cathode material delithiation phase transition plateau. The SOC matching logic extracts SOC values ​​that satisfy dVdQ∈[6.08-∈, 6.08+∈] within the discharge subrange (e.g., SOC 95-100%), and calculates their arithmetic mean as the calibration point SOCCalibration = 99.5%.

[0092] Actual battery capacity C actual Due to manufacturing errors and aging, the nominal capacity C nominalTo improve the calculation accuracy of coulomb counting method, the data-driven method is used to calibrate the actual capacity in the embodiment. The health indicators (HIs) reflecting the battery degradation state are extracted from the voltage-time curve changes during the charging / discharging process. The maximum information coefficient (MIC) method is used to analyze the correlation between each health indicator and the actual capacity, and the indicators with higher MIC values are selected. Then, the selected health indicators are used as input to build a Gaussian process regression (GPR) model, which realizes online estimation of the actual capacity. The capacity parameters used in the coulomb counting method are corrected to ensure the accuracy of SOC calculation.

[0093] In the initial SOC correction mechanism, the trigger condition is when the BMS detects that the current dQ / dV value enters the preset interval (charging phase [2.02, 2.04] or discharging phase [6.07, 6.09]), the calibration process is activated. The sliding window algorithm (window length N = 10 sampling points) is used to filter the SOC calibration points, eliminate noise interference, and the formula is:

[0094]

[0095] The calibrated SOC is replaced by the initial value SOCinitial in the coulomb counting method, and the cumulative capacity integrator is reset. The calibration point tolerance is dynamically adjusted according to the battery aging degree. The initial tolerance is ±0.1%, and it is relaxed to ±0.15% after 50 cycles to compensate for the curve shift caused by the loss of electrode active material. When the BMS detects dQ / dSOC = 6.08, the temperature compensation algorithm (compensation coefficient 0.05% / ℃) is used to correct the SOC to 99.5%, ensuring the calibration robustness under high-rate discharge.

[0096] Example Three: Please refer to Figure 3 The embodiment includes the following experiments:

[0097] Experimental environment and data collection

[0098] Test object: 4 groups of 50 Ah ternary batteries (cycled to 20% capacity attenuation);

[0099] Data source: voltage-time data set accumulated for 2000 times of charge / discharge cycles;

[0100] Key parameters:

[0101]

[0102] In the step of extracting multiple health indicators (HIs) in S103, the optimal health indicators (HIs) are screened by quantifying the nonlinear correlation between the voltage curve segment and the battery capacity. In the charging process, the voltage values at the initial stage of charging (for example, from the 5th second to the 45th second, and one data point is taken every 5 seconds) are collected from the battery voltage-time curve, and a total of 9 sampling points are obtained, which are marked as HI1 to HI9, and the cumulative charging capacity at the corresponding time is recorded; in the discharging process, the voltage data are collected by selecting a suitable time region (for example, from the 425th second to the 625th second, and one data point is taken every 25 seconds), to form another group of health indicator data, such as Figure 3 The specific implementation process is as follows:

[0103] S301 formulates the extraction rule of candidate HIs. The time window selection rule of HIs in the charging stage is that 9 voltage values (V5, V10,..., V45) are extracted at intervals of 5 seconds within 5-45 seconds after the start of constant current charging (in this stage, the battery polarization effect is significant, and the capacity attenuation is sensitive), which are marked as HI1 to HI9. The voltage sequence is subjected to standardization processing, and the formula is:

[0104]

[0105] Wherein μV is the mean value, and σV is the standard deviation.

[0106] The time window selection rule of HIs in the discharging stage is that 9 voltage values (V425, V450,..., V625) are extracted at intervals of 25 seconds within 425-625 seconds before the end of constant current discharging (in this stage, the battery terminal voltage drop rate is strongly related to the capacity attenuation), which are marked as HI1 to HI9.

[0107] S302 performs MIC correlation analysis. The principle of MIC calculation is to form a two-dimensional data set {(HIi, Cj)} by combining the HI in the battery historical data with the actual capacity C. The correlation coefficient between each health indicator (HI1-HI9) and the actual capacity is calculated by using the MIC algorithm; for the charging stage, the MIC values of the health indicators are obtained by calculating the data of battery #1 and battery #2 respectively, and the average value of the two is taken; according to the average MIC value, the four indicators with the highest MIC values (HI3, HI4, HI5 and HI7 in this embodiment) are selected as the input features for subsequent actual capacity estimation; for the discharging stage, the four indicators with the highest MIC values (for example, HI6, HI7, HI8 and HI9) are also selected as the input according to the MIC value sorting; the dynamic grid division algorithm is used to calculate the normalized mutual information value:

[0108]

[0109] Wherein n G ,m Gis the number of rows and columns in the grid G, and I(·) is the mutual information function. The screening rule is to calculate the MIC value of each HI and sort them in descending order. The first four HIs with the largest MIC values ​​are selected as the model input (e.g., HI3, HI4, HI5, and HI7 in the charging stage).

[0110] S303 performs robustness enhancement design. Using a sliding time window, multiple time windows (e.g., 50 cycles per window) are randomly selected from the historical data to verify the stability of the MIC values ​​of the HIs. If the MIC values ​​of two HIs differ by less than 5% and are close in time (e.g., HI3 and HI4 are 5 seconds apart), the latter is removed to reduce dimensionality.

[0111] Example 4: Please refer to Figure 4 , this embodiment includes the following experiments:

[0112] Experimental environment and data collection

[0113] Test object: 4 sets of 60Ah NMC811 / graphite system power batteries (cycle aging to 80% SOH);

[0114] Hardware configuration:

[0115] High-precision BMS development board (STM32H7, 512KB RAM)

[0116] 16-bit ADC voltage acquisition (±0.5mV accuracy)

[0117] Hall effect current sensor (±0.1% accuracy)

[0118] Software environment:

[0119] Embedded GPR model inference framework (TensorFlow Lite Micro)

[0120] Real-time operating system (FreeRTOS, 1ms task cycle)

[0121] In S104, a Gaussian process regression (GPR) model is constructed to perform online estimation of actual capacity. A mapping relationship between HIs and battery capacity is established through a non-parametric probability model. The specific implementation process is as follows:

[0122] S401 Model Construction. First, define the input and output of the model and solve the covariance function. Finally, optimize and solve the hyperparameters. Using the training set data, determine the hyperparameters of the GPR model by the maximum likelihood estimation method. f ,l,σ n]; Use the conjugate gradient method to solve the hyperparameters so that the log-likelihood function reaches its maximum value; continuously evaluate the model's fitting effect on the validation set during training to ensure overfitting; use the trained GPR model to predict the validation set and output the predicted capacity C estimated ; Calculate the relative error between the predicted value and the actual measured capacity, requiring the error to be within 1.5%; for the charging stage, the error is generally controlled within 1.5%; for the discharging stage, the error is controlled within 2.5%; after the model training is completed, the GPR model is embedded in the BMS system or diagnostic platform to achieve real-time online capacity calibration. Figure 4 As shown, the specific steps can be broken down into:

[0123] S401-1. Input and output definition. Gaussian process regression (GPR) is selected as the capacity estimation model; the four health indicators selected in S102 form the input vector x, and the actual charge or discharge capacity is the output y; the input is the feature vector x = [HI1, HI2, HI3, HI4] composed of the four selected HIs, and the output is the actual battery capacity C 实际 .

[0124] S401-2. Solve the covariance function: use the squared exponential (SE) kernel function:

[0125]

[0126] Where σf is the signal variance, l is the characteristic length scale, is the noise variance, δ ij is the Kroneckerdelta function.

[0127] S401-3, hyperparameter optimization, the method is to solve the optimal hyperparameter θ=(σ f ,l,σ n ):

[0128]

[0129] And solve it iteratively using the conjugate gradient method:

[0130]

[0131] S402, online capacity estimation: After the initial SOC and actual capacity are calibrated through the above steps, the traditional coulomb counting method is modified to achieve accurate tracking of the SOC.

[0132] S402-1. New data prediction: For new input x*, the posterior distribution is:

[0133]

[0134] in:

[0135]

[0136] S402-2, Capacity Update Strategy: When σ*>0.1C nominal, it is determined that the current estimation uncertainty is too high, triggering the HIs re-screening process. In the application of the coulomb counting method in S203 to accurately estimate the battery SOC, the calibration parameters and the capacity estimation value are integrated to achieve a closed-loop correction of the SOC. The specific parameter injection mechanism is to first replace the initial SOC, and replace the SOC calibration value output in step S102 with the SOC initial value in the original coulomb counting method. Then perform a dynamic capacity update, and replace the C output in step S104 with the SOC calibration value output in step S104. 实际 Replacement nominal capacity C 标称 , the calculation formula is updated as follows:

[0137]

[0138] At the same time, the coulomb counting method has a fault tolerance and recovery mechanism. If the current fluctuation of 10 consecutive sampling points exceeds ±20% (excluding pulse conditions), the integration is suspended and switched to the open circuit voltage (OCV) correction mode. The above formula is executed for the charging and discharging stages respectively to ensure that the SOC curve is smooth and has no cumulative error; the system monitors the slope of the SOC curve in real time. If abnormal fluctuations occur, the secondary correction process is started to further reduce the error. In addition, the calibration validity period T is set. valid After a timeout (e.g., 300 seconds), the nominal capacity is automatically restored until the next calibration is triggered. The two mechanisms work together to ensure timely calibration while avoiding error accumulation caused by a single failure event.

[0139] S403, online SOC estimation result verification and feedback. To verify the effectiveness of the modified coulomb counting method in practical applications, this embodiment performs the following verification steps:

[0140] S403-1. Charging process verification: Initial conditions are set: the actual initial SOC is 0%, but an incorrect initial SOC (e.g., 25%) is intentionally set in the system to simulate possible sensor errors in reality. When charging begins, the system captures the calibration point of dQ / dV = 2.03 within the third second and automatically corrects the SOC to 0.13%. During subsequent charging, the system calculates the SOC based on the corrected parameters, and the result shows that it is very close to the actual SOC, with an error within 0.05%. At the same time, compared with the uncorrected SOC estimate, it is found that the cumulative error is increasing and may eventually exceed 125%, proving the necessity of the correction method.

[0141] S403-2, discharge process verification: set initial conditions, actual discharge initial SOC is 100%, but the system is preset with an error initial SOC (such as 80%); in the early stage of discharge, when the system detects dQ / dV = 6.08, the SOC is automatically corrected to 99.5%; in the subsequent discharge process, the real-time updated actual capacity value is used for SOC calculation, the result shows that the SOC descending trajectory is basically consistent with the true curve, and the error is controlled within 0.25%; compared with the uncorrected SOC, it is found that the final value may appear negative (for example, -20%), which further verifies the effectiveness of the method of the application;

[0142] S403-3, feedback correction mechanism: the system is built-in with a self-checking module, which periodically detects the error of the SOC curve; when the error exceeds the preset range for continuous multiple sampling periods, a secondary correction process is automatically triggered to recalculate the calibration point and the actual capacity; the feedback result will be used for fine-tuning of the subsequent SOC calculation parameters, to ensure that the error accumulation is always at a low level in long-term operation.

[0143] Although embodiments of the application have been shown and described, it is to be understood that various modifications, substitutions, replacements and changes can be made to these embodiments without departing from the principles and spirit of the application, and the scope of the application is defined by the appended claims and their equivalents.

Claims

1. A method for estimating SOC of a large-capacity energy storage battery based on a data-driven coulomb method, characterized in that: The following steps are involved: S101, constructing an incremental capacity curve based on the SOC axis using voltage and capacity data collected during battery charging and discharging; S102, extracting initial SOC calibration points during the charging and discharging processes, respectively, and determining a calibration SOC value using a specific dQ / dV value; S103. To address the error problem of actual capacity, the maximum information coefficient method is used to extract health indicators from the voltage-time curve; S104. A Gaussian process regression model is constructed using the health indicator of S103 as input to perform online estimation of the actual capacity. The parameters after dual calibration of the initial SOC and actual capacity are applied to the coulomb counting method to achieve accurate estimation of the battery SOC.

2. The method for estimating SOC of a large-capacity energy storage battery based on a data-driven coulomb method according to claim 1, characterized in that: In the extraction of initial SOC calibration points during the charging and discharging processes and calibration of the initial SOC value based on the incremental capacity analysis of the SOC axis described in S102, this step reconstructs the coordinate system of the incremental capacity analysis, establishes a mapping relationship between SOC and dQ / dV, and implements rapid calibration of the initial SOC based on preset feature points. The specific implementation process is as follows: S201, data acquisition: During the battery charge / discharge process, the battery terminal voltage V(t), current I(t) and timestamp t are collected in real time at a fixed sampling frequency, and the cumulative capacity is calculated by integration S202, SOC calculation: Preliminary estimate of SOC value based on the coulomb counting method, the formula is: Where C is the nominal battery capacity, and SOC initial is the uncalibrated initial value; S203, dQ / dV calculation: Use the differential method to calculate the incremental capacity value. Extract discrete data pairs of capacity and voltage from the charge and discharge process, and use the five-point central difference method to calculate the instantaneous dQ / dV value. The formula is: S204. Coordinate transformation: Reconstruct the traditional dQ / dV curve with voltage as the horizontal axis into a curve with SOC as the horizontal axis. The specific method is to replace the voltage axis with the SOC axis, calculate the SOC in real time by integrating the current-time data, generate the dQ / dSOC curve, and form a two-dimensional data set of SOC(t) and dVdQ(t).

3. The method for estimating SOC of a large-capacity energy storage battery based on a data-driven coulomb method according to claim 1, characterized in that: In step S103, extracting multiple health indicators, the optimal health indicator is selected by quantifying the nonlinear correlation between the voltage curve segments and the battery capacity. During the charging process, the voltage values ​​at the initial charging stage are collected from the battery voltage-time curve, resulting in nine sampling points, labeled HI1 to HI9, and the cumulative charge capacity at the corresponding moments is recorded. During the discharging process, appropriate time zones are selected to collect voltage data to form another set of health indicator data. The specific implementation process is as follows: S301. Formulate a rule for extracting candidate HIs. The time window selection rule for His in the charging phase is to extract 9 voltage values ​​at 5-second intervals within 5-45 seconds after the start of constant current charging, and record them as HI1 to HI9. Standardize the voltage sequence using the formula: Where μV is the mean and σV is the standard deviation; S302, perform MIC correlation analysis. The MIC calculation principle is to combine the HI and capacity C in the battery history data into a two-dimensional data set {(HIi, Cj)}; S303: Perform robustness enhancement design and verify the stability of the MIC value of HIs by sliding time window verification. Randomly select multiple time windows in the historical data to verify the stability of the MIC value of HIs. If the difference in the MIC values ​​of two HIs is less than 5% and the time is close, the latter is eliminated to reduce the dimension.

4. The method for estimating SOC of a large-capacity energy storage battery based on a data-driven coulomb method according to claim 1, characterized in that: In step S104, a Gaussian process regression model is constructed to perform online estimation of the actual capacity. A mapping relationship between HIs and battery capacity is established through a non-parametric probability model. The specific implementation process is as follows: S401: Model construction. First, define the model's input and output, solve the covariance function, and finally optimize and solve the hyperparameters. S402, online capacity estimation: After the initial SOC and actual capacity are corrected in the above steps, the traditional coulomb counting method is modified to achieve accurate tracking of SOC; S403, online SOC estimation result verification and feedback, is to verify the effectiveness of the modified Coulomb counting method in practical applications.

5. The method for estimating SOC of a large-capacity energy storage battery based on a data-driven coulomb method according to claim 4, characterized in that: The specific steps of step S403 are as follows: S403-1. Charging process verification: Set initial conditions: The actual initial SOC is 0%, but an incorrect initial SOC is intentionally set in the system to simulate possible sensor errors in practice. When charging begins, the system captures the calibration point of dQ / dV = 2.03 within the third second and automatically corrects the SOC to 0.13%. During subsequent charging, the system calculated the SOC based on the corrected parameters, and the result was very close to the actual SOC, with an error within 0.05%. Furthermore, when compared to the uncorrected SOC estimate, the cumulative error increased, eventually exceeding 125%, demonstrating the necessity of the correction method. S403-2, discharge process verification, the actual initial discharge SOC is 100%, but the system presets an incorrect initial SOC; at the beginning of discharge, when the system detects dQ / dV = 6.08, it automatically corrects the SOC to 99.5%; During subsequent discharge, the SOC was calculated using the real-time updated actual capacity value. The results showed that the SOC decline trajectory basically coincided with the true curve, with the error controlled within 0.25%. Compared with the uncorrected SOC, it was found that it could eventually become negative. S403-3, feedback correction mechanism, the system has a built-in self-test module to periodically perform error detection on the SOC curve; When the error exceeds the preset range for multiple consecutive sampling cycles, the secondary calibration process is automatically triggered to recalculate the calibration point and actual capacity; The feedback results will be used to fine-tune the subsequent SOC calculation parameters to ensure that the error accumulation remains at a low level during long-term operation.

6. The method for estimating SOC of a large-capacity energy storage battery based on a data-driven coulomb method according to claim 1, characterized in that: In step S101, the voltage and capacity data of the battery charge and discharge process are collected, and two lithium-ion batteries with the same nominal capacity are used; a high-precision current and voltage acquisition module and a data recording system are equipped with online sampling and real-time calculation functions; in a constant temperature room, CC-CV mode charging is adopted with a charging current of 1C. When the battery terminal voltage reaches 4.2V, constant voltage charging is switched until the charging current drops to 0.1C; the discharge process adopts CC mode with a discharge current of 3C and a cut-off voltage set to 2.75V; a 30-minute static recovery time is set after each cycle to ensure that the battery state is stable, and a matrix D = [V, I, t] is constructed, where V∈RN×1 is the voltage sequence, I∈RN×1 is the current sequence, t∈RN×1 is the timestamp, and N is the number of data points.

7. The method for estimating SOC of a large-capacity energy storage battery based on a data-driven coulomb method according to claim 1, characterized in that: In step S203, the target dQ / dV value of the calibration point in the charging phase is set to 2.03, corresponding to the starting point of the phase change platform of the graphite negative electrode of the lithium-ion battery. In the early stage of charging, the dQ / dSOC curve fluctuates violently, but the specific dQ / dSOC value corresponds to the SOC with high repeatability. The SOC matching logic traverses the charging sub-interval and extracts the SOC value that satisfies dVdQ∈[2.03-∈,2.03+∈], where ∈ is the tolerance threshold, and calculates the arithmetic mean thereof as the calibration point SOC calibration = 0.13%. The target dQ / dV value of the calibration point in the discharge phase is set to 6.08, corresponding to the end point of the delithiation phase change platform of the positive electrode material. The SOC matching logic extracts the SOC value that satisfies dVdQ∈[6.08-∈,6.08+∈] within the discharge sub-interval and calculates the arithmetic mean thereof as the calibration point SOC calibration = 99.5%; A data-driven approach is used to calibrate the actual capacity, and the changes in the voltage-time curve during the charge / discharge process are used to extract health indicators reflecting the battery degradation state. The maximum information coefficient method is used to analyze the correlation between each health indicator and the actual capacity, and the indicators with higher MIC values ​​are selected. Then, using the selected health indicators as input, a Gaussian process regression model is constructed to achieve online estimation of the actual capacity. The capacity parameters used in the coulomb counting method are corrected to ensure the accuracy of the SOC calculation. In the initial SOC correction mechanism, the trigger condition is when the BMS detects that the current dQ / dV value enters the preset range, activating the calibration process. The dynamic update uses a sliding window algorithm to filter the SOC calibration point to eliminate noise interference. The formula is:

8. The method for estimating SOC of a large-capacity energy storage battery based on a data-driven coulomb method according to claim 1, characterized in that: In step S204, a Savitzky-Golay filter is used, with a default window width of 11 points and a default polynomial order of 3, to smooth the dQ / dSOC curve and eliminate high-frequency noise. The phase transition peak in the curve is identified using a peak detection algorithm. The formula is:

9. The method for estimating SOC of a large-capacity energy storage battery based on a data-driven coulomb method according to claim 3, characterized in that: The time window selection rule of the discharge stage HIs in step S301 is to extract 9 voltage values ​​at intervals of 25 seconds within 425-625 seconds before the end of the constant current discharge, and record them as HI1 to HI9.

10. The method for estimating SOC of a large-capacity energy storage battery based on a data-driven coulomb method according to claim 1, characterized in that: In step S302, the correlation coefficient between each health indicator and the actual capacity is calculated using the MIC algorithm; For the charging stage, the MIC value of each battery health indicator is obtained by calculating the data separately, and the average of the two is taken; the four indicators with the highest MIC values ​​are selected as the input features for the subsequent actual capacity estimation according to the sorting of the average MIC values; In the discharge phase, the four indicators with the highest MIC values ​​are selected as inputs based on the MIC value sorting. The dynamic grid partitioning algorithm is used to calculate the normalized mutual information value: where n G ,m G is the number of rows and columns of the grid G, I is the mutual information function, and the screening rule is to calculate the MIC value of each HI, sort them in descending order, and select the first four HIs with the largest MIC values ​​as the model input.

Images