Lithium ion battery charge state detection method based on adaptive optimization

Abstract

The invention discloses a lithium ion battery state-of-charge detection method based on adaptive optimization, and particularly relates to the technical field of lithium ion battery state-of-charge detection, and the method comprises the steps: collecting the operation data of a new station to extract the fingerprint vector and state credibility evidence of the new station; performing similarity retrieval and weighted interpolation from a pre-established parameter and strategy library based on the fingerprint, generating a localized model parameter prior and decision threshold, and evaluating a migration risk; under an equivalent circuit state space model framework, constructing a dual-objective optimization problem fusing voltage fitting and charge conservation, dynamically deploying the weights of the two according to real-time evidence and risk, and solving the optimal charge state under the constraint of parameter update amplitude; according to the method, multi-source residual errors and evidences are accumulated to drive a gating state machine, a parameter learning strategy is intelligently decided, minimum-cost parameter self-calibration is started when a favorable working condition is detected, and the problems that in a large-scale energy storage power station, a battery model is difficult to universally adapt, and state estimation is prone to data interference are effectively solved.

Description

Technical Field

[0001] This invention relates to the field of lithium-ion battery state-of-charge detection technology, and more specifically, to a lithium-ion battery state-of-charge detection method based on adaptive optimization. Background Technology

[0002] Highly reliable online estimation of the state of charge (SOC) of lithium-ion batteries is crucial for the safe and efficient operation of energy storage power stations. Model-based adaptive estimation methods face a fundamental contradiction in large-scale deployments: the difficulty in reconciling "personalized adaptation" with "continuous robust learning." On one hand, the inherent differences in battery aging, operating environment, and scheduling modes among different energy storage sites lead to severe initial mismatches when using fixed or universal model parameters, resulting in low estimation accuracy and slow convergence. While online learning can provide adjustments, on the other hand, many energy storage power stations operate under relatively stable conditions, and the excitation signals are insufficient to support complete and reliable identification of model parameters. Furthermore, noise, anomalies, and asynchrony are unavoidable in the data acquisition process. If adaptive algorithms indiscriminately update the model based on such low-quality or insufficiently stimulated data, parameter drift, overfitting, and even estimator divergence can easily occur, leading to the risk of "negative transference."

[0003] Existing technologies often address initialization or online robustness issues in isolation, lacking a closed-loop mechanism that integrates prior knowledge guidance, online risk perception, and long-term experience accumulation.

[0004] Therefore, developing a SOC estimation method that can provide a safe and reliable personalized starting point for new sites and make intelligent decisions on learning timing and priorities in complex operating environments, thereby achieving stable estimation and autonomous knowledge evolution, has become the key to promoting the intelligent operation and maintenance of energy storage systems. Summary of the Invention

[0005] To overcome the aforementioned deficiencies in existing technologies, this invention provides a lithium-ion battery state-of-charge detection method based on adaptive optimization. This method constructs a dual-track perception system that separates the fingerprint vector of a new site from state credibility evidence, establishes a knowledge transfer initialization mechanism based on fingerprint similarity retrieval and risk perception, designs an evidence-driven dual-objective adaptive optimization estimator, and forms a closed-loop gated self-calibration and knowledge evolution process to address the problems mentioned in the background technology.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a lithium-ion battery state-of-charge detection method based on adaptive optimization, comprising collecting current, voltage, and temperature data of batteries at new sites; establishing an equivalent circuit state-space model of the battery, wherein the state variables of the model include the state of charge and polarization voltage, and the model parameters include ohmic internal resistance and polarization parameters, the model parameters varying with temperature, specifically including the following steps: Step S1: Collect multi-source runtime sequence data of the new site within the set sliding window, extract and store fingerprint vectors that characterize the long-term operating mode of the new site, as well as state credibility evidence used to evaluate real-time data quality and model observability. Step S2: Based on the fingerprint vector, perform similarity retrieval from the pre-built parameter and strategy library. The parameter and strategy library records the fingerprint summary vector of historical sites, battery model parameter priors, filter noise setting priors, and gate threshold priors. Based on the retrieval results, generate local battery model parameter priors, local noise setting priors, and local gate threshold priors for new sites through weighted interpolation, and simultaneously generate negative migration risk levels. Step S3: Within the equivalent circuit state-space model framework, construct an innovation consistency constraint with voltage fitting as the objective and a conservation consistency constraint with charge conservation as the objective; dynamically adjust the weights of the two constraints based on state credibility evidence and negative migration risk level to form a dual-objective adaptive optimization problem; solve the optimization problem, while applying a continuity constraint on the update magnitude of model parameters, output the optimal state of charge estimate, and simultaneously output the voltage innovation residual sequence and conservation residual sequence. Step S4: Multi-timescale fusion and accumulation of voltage innovation residual sequence, conservation residual sequence, state credibility evidence, and negative migration risk level to drive the gated state machine to make state decisions; the gated state machine performs parameter updates, parameter freezing, or parameter rollback under different gated states based on the accumulated evidence; when a natural step or observable excitation window is detected, minimum cost self-calibration is initiated, updating only a parameter subset pre-selected based on parameter sensitivity and prior uncertainty; the parameter update results obtained from self-calibration are associated with the current operating condition fingerprint and written back to the parameter and strategy library.

[0007] The explanation is as follows: Step S1 constructs a fingerprint vector database for the new site and a state credibility buffer that is updated in real time, which forms the data foundation for all subsequent intelligent decisions of this invention; Step S2 will use the fingerprint vectors of the new site to perform similarity retrieval and prior knowledge interpolation; Step S3 will dynamically balance the dual constraints inside the estimator based on the state credibility; Step S4 will integrate the lifecycle of the evidence management model parameter update.

[0008] Preferably, the process of extracting the fingerprint vector includes: extracting temperature distribution morphology parameters, fundamental statistical features and fluctuation features of power commands, and the proportion of current rest periods from the multi-source runtime sequence data; and extracting polarization time constant features from the voltage relaxation curve of the rest periods; combining the features to form the fingerprint vector; and iteratively updating the fingerprint vector using an exponential smoothing strategy, specifically by weighting and summing the currently stored fingerprint vector with the temporary fingerprint vector extracted based on the latest sliding window using a preset smoothing factor to generate and store the updated fingerprint vector.

[0009] Preferably, in step S2, the similarity retrieval and weighted interpolation process based on fingerprint vectors includes: calculating the weighted Euclidean distance between the fingerprint summary vector and parameters of the new site and the fingerprint summary vectors of each record in the strategy library; selecting a candidate record set based on the distance and calculating the initial weights; reading the evidence adjustment factor calculated based on the state credibility evidence, adjusting and normalizing the initial weights to obtain the final weights used for interpolation; the process of generating local model parameter priors through fingerprint distance weighted interpolation is limited by a set of dynamic credibility envelopes; the width of the dynamic credibility envelopes is jointly determined by the historical volatility of each parameter and the parameter identifiability calculated based on the short-term observation data of the current site. The explanation explains that the identifiability parameter is approximated by calculating the condition number of the Fisher information matrix corresponding to the current short-term observation data; a larger condition number indicates poorer identifiability. The width of the dynamic confidence envelope is proportional to the historical volatility and inversely proportional to the current identifiability. This mechanism makes the confidence interval for knowledge transfer no longer static, but dynamically shrinks or expands with the current data quality of the target site. When the data quality is high (high identifiability), more accurate interpolation is allowed even with large historical fluctuations; when the data quality is poor, a more conservative historical fluctuation range is relied upon.

[0010] It should be noted that the dynamic reliable envelope refers to a dynamically changing reliable interval applied to each parameter when generating local model parameter priors through weighted interpolation of candidate records. The width of this envelope is determined by two factors: first, the historical volatility of the parameter across all records in the parameter and policy library; and second, the parameter identifiability calculated based on short-term observation data of the current site. Specifically, for any model parameter θ, its dynamic reliable envelope width... The calculation formula is:

[0011] in, This is the historical volatility of the parameter in the candidate records (i.e., the width standard deviation of the parameter's prior range across all candidate records). The identifiability of parameters corresponding to short-term observation data at the current site (approximated by the reciprocal of the condition number of the Fisher information matrix; the smaller the condition number, the better). The larger the value, the greater the value; k is a preset scaling factor (e.g., 2); during the weighted interpolation process, the prior range of the finally generated local parameters must fall within the dynamic reliable envelope, that is, the interpolation result must not exceed the area centered on the weighted average value with a width of 1. The range.

[0012] Preferably, in step S2, the process of generating local battery model parameter priors adopts a parameter-specific fusion strategy: for battery model parameters, the weighted statistics of each parameter are calculated according to the final weights, and a local prior is generated in terms of value range based on the intersection of the weighted statistics and the original prior range of the candidate records; for filter noise setting priors and gate threshold priors, the corresponding values ​​of the candidate records are weighted and averaged according to the final weights to generate local priors in terms of specific numerical values.

[0013] Preferably, in step S3, the comprehensive cost function of the constructed bi-objective adaptive optimization problem is composed of a weighted sum of a voltage fitting consistency cost term and a charge conservation consistency cost term; the voltage fitting consistency cost term includes a penalty for the sum of squares of the normalized voltage prediction residuals within the sliding window, and an additional penalty for the absolute value of the first-order autocorrelation coefficient of the residual sequence; the charge conservation consistency cost term includes a penalty for the sum of squares of the difference between the predicted state of charge increment and the Coulomb integral increment within the sliding window, and an additional penalty for the absolute value of the difference between their cumulative sums.

[0014] Preferably, in step S3, the process of dynamically allocating the weights of the two constraints specifically involves: linearly combining the state credibility evidence with the negative migration risk level, and inputting the combination result into a Sigmoid mapping function to output an adaptive weight value between zero and one, which is used to perform a weighted summation of the two cost terms in the comprehensive cost function.

[0015] Preferably, in step S4, the gating state machine includes an active learning state, an observation freeze state, and a safety rollback state; the state transition is driven by the comparison result of the comprehensive confidence index and the local gating threshold; when in the observation freeze state, the constraint on the model parameter update magnitude in step S3 is tightened or the update of some model parameters is frozen; when in the safety rollback state, the parameter update in step S3 is prohibited.

[0016] Preferably, in step S4, the process of initiating the minimum cost self-calibration to update the identifiable parameter subset includes: sorting the parameters according to the product of their sensitivity and prior uncertainty, and selecting the top few parameters with the largest product to form the subset to be calibrated; within the calibration time window corresponding to the natural step or observable excitation condition window, optimizing the parameters in the subset to be calibrated with the goal of minimizing voltage prediction error and applying parameter boundary and update magnitude constraints; and writing the update result back to the parameter and policy library by associating the self-calibrated updated local battery model parameter prior, the adjusted noise setting prior, and the latest fingerprint summary of the current new site as a new record to update the parameter and policy library.

[0017] Preferably, the observation method for the observable excitation condition window is as follows: Calculate the operating condition excitation dispersion index (OCEDI) and the parameter stress sensitivity index (PSSI); based on the positions of the operating condition excitation dispersion index and the parameter stress sensitivity index in the preset decision matrix, determine whether the current operating condition is an observable excitation window, and decide whether to trigger and to what degree of conservatism to trigger the self-calibration. The operating condition excitation dispersion index is obtained by performing a frequency domain transformation on the current sequence within a preset evaluation window and calculating the distribution ratio of its energy in the key dynamic frequency band of the battery model. The parameter stress sensitivity index is obtained by calculating the sum of the absolute values ​​of the partial derivatives of the output voltage of the equivalent circuit model with respect to a preset subset of key model parameters based on the current state of charge and temperature at the current operating point. Based on the calculated operating condition excitation dispersion index and parameter stress sensitivity index, a preset two-dimensional decision matrix is ​​queried, with the horizontal and vertical axes of the decision matrix corresponding to the discretization levels of the two indices, respectively. Based on the units mapped by the indices in the matrix, it is determined whether the observable excitation operating condition window conditions are met, and the level of conservatism to be adopted if self-calibration is triggered is simultaneously determined. The level of conservatism is associated with a set of self-calibration parameters, including the calibration time window length, parameter update regularization strength, and write-back smoothing factor.

[0018] The explanation is as follows: the key dynamic frequency band refers to the characteristic frequency band corresponding to the polarization time constant of the first-order RC equivalent circuit of the battery, and its upper and lower limits are determined by the range of polarization time constant in the prior parameters of the local battery model; the key model parameter subset refers to the subset to be calibrated that is pre-selected according to the sensitivity-uncertainty product sorting; the two-dimensional decision matrix is ​​constructed through offline simulation and historical data learning: using the operating condition excitation dispersion index (OCEDI) and parameter stress sensitivity index as features, and using whether effective calibration is successfully completed and the degree of improvement in short-term estimation error after calibration as labels, a classification and regression model is trained, and the decision boundary is discretized to form the matrix.

[0019] The explanation is as follows: the preset two-dimensional decision matrix is ​​pre-constructed through offline simulation and learning from historical data; the construction process is as follows: First, the Operating Condition Excitation Dispersion Index (OCEDI) and the Parametric Stress Sensitivity Index (PSSI) are discretized into several levels, for example, low, medium, and high, forming a 3×3 matrix unit. Each unit corresponds to an operating condition type and is pre-associated with a set of self-calibration strategy parameters (including calibration time window length, parameter update regularization strength, and write-back smoothing factor). The matrix is ​​constructed based on historical data or simulation experiments: for each (OCEDI, PSSI) combination, the success rate of parameter self-calibration and the degree of improvement in estimation error under that operating condition are evaluated, and the strategy parameters that optimize the calibration effect are selected and filled into the corresponding unit. In actual operation, the level is determined based on the real-time calculated OCEDI and PSSI values, and the corresponding unit in the matrix is ​​queried to determine whether the observable excitation operating condition window condition is met and to obtain the corresponding self-calibration conservatism parameters.

[0020] Preferably, the evidence accumulation includes: performing transient filtering and trend extraction on the innovation residual and conservation residual sequences respectively; analyzing the correlation patterns between the two residual trend components and matching them with predefined model mismatch failure modes to generate a physical-oriented comprehensive model mismatch confidence level, which is used to drive the gated state machine; The explanation states that the trend correlation pattern analysis of the innovative residual and the conserved residual refers to examining the consistency of the signs and phases of their trends. For example, if both trends are positive, it indicates a mismatch dominated by capacity decay, while if the trends are opposite, it indicates a mismatch dominated by internal resistance change. The predefined model mismatch failure modes are several typical mismatch feature templates extracted by clustering analysis of historical failure case data.

[0021] The technical effects and advantages of this invention are as follows: The solution provided by this invention effectively overcomes the difficulties in model initialization caused by differences in battery characteristics at new energy storage sites, as well as the estimation inaccuracies and parameter drift caused by fluctuations in operating conditions and insufficient excitation during online operation. It generates personalized prior configurations for each site through knowledge transfer driven by site fingerprints, achieving rapid and reliable initialization without manual parameter tuning. Furthermore, by leveraging a dual-objective optimization mechanism guided by evidence and risk, and a gated learning mechanism, it ensures robustness of full-condition estimation while continuously tracking battery aging and autonomously accumulating system knowledge, forming a closed-loop estimation system with progressive learning capabilities. Attached Figure Description Figure 1 This is a simplified flowchart of the lithium-ion battery state-of-charge detection method of the present invention. Detailed Implementation

[0022] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0023] At the same time, it should be understood that, for ease of description, the dimensions of the various parts shown in the accompanying drawings are not drawn according to actual scale.

[0024] The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the scope of this application and its application or use.

[0025] Techniques, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and equipment should be considered part of the specification.

[0026] See Figure 1 A simplified flowchart of a lithium-ion battery state-of-charge detection method is provided in this invention. Figure 1 The adaptive optimization-based lithium-ion battery state-of-charge detection method shown includes: Collect operational data from new sites to extract their fingerprint vectors and state credibility evidence; Based on fingerprint vectors, similarity retrieval and weighted interpolation are performed from pre-built parameters and policy libraries to generate localized model parameter priors and decision thresholds, and to assess migration risk. Within the framework of the equivalent circuit state-space model, a dual-objective optimization problem integrating voltage fitting and charge conservation is constructed. The weights of the two are dynamically adjusted based on real-time evidence and risk, and the optimal state of charge is solved under the constraint of parameter update magnitude. Accumulate multi-source residuals and evidence to drive a gated state machine, implement an intelligent decision parameter learning strategy, and initiate minimum-cost parameter self-calibration when a favorable operating condition is detected.

[0027] Specifically, the method includes the following steps: Step S1: Collect the current, voltage, temperature, power command and actual power, equalization action record and sampling anomaly flag of the new site. Extract and store the fingerprint vector and state credibility evidence of the new site in the sliding window. The fingerprint vector of the new site is used to characterize the long-term operation mode of the new site, and the state credibility evidence is used to characterize the short-term observability and data reliability. Background: Step S1 establishes a dual-track observation system of fingerprint vectors and state credibility evidence for new sites, creating conditions for cross-site knowledge utilization. Step S2, based on the operational fingerprints of new sites, performs intelligent retrieval and interpolation operations from a pre-built parameter and strategy library covering prior knowledge and strategies of multiple new sites. This automatically generates localized prior parameters, filter configurations, and decision thresholds for the battery model, while quantitatively assessing the potential risks of knowledge transfer. This replaces the traditional, inefficient, and expert-dependent manual parameter tuning process, providing a high-starting-point and robust initial configuration for subsequent online adaptive optimization.

[0028] Technical Motivation: Configuring battery model parameters and state estimator parameters for each new (energy storage) power station is a tedious and error-prone task. If a set of default parameters is used directly, it cannot adapt to the inherent differences between new stations due to different battery aging states, operating environments, and scheduling modes. If it relies entirely on online learning after deployment, the algorithm may converge slowly or even converge to an incorrect parameter space in the early stages of learning due to the lack of reasonable prior guidance, resulting in long-term deviations in SOC estimation. Therefore, an initialization mechanism is needed that can automatically match the characteristics of new stations, learn from experience from similar historical new stations, safely integrate the experience of multiple similar new stations, and judge the reliability of the fusion result.

[0029] Furthermore, the specific implementation process of step S1 includes: Step 101 (Multi-source data acquisition and sliding window management): Set a fixed-duration sliding window, such as 72 hours, which is sufficient to cover typical daily charge-discharge cycles and multi-day operation modes. Within this window, continuously collect raw operating sequence data of the new site, including: the terminal voltage of the battery cluster, the total current (including the positive or negative sign indicating the direction of charging or discharging), the temperature of multiple representative measurement points, the power command from the upper-level energy management system and its actual executed power value; the status flags of the balancing action automatically triggered within the new site, and the sampling abnormality status flags reported by the self-check. The explanation is as follows: After a new site is put into operation, its upper-level energy management system acts as the scheduling center, issuing specific charging and discharging power commands to the battery management system of the new site. After receiving the command, the new site executes the corresponding operation and feeds back the actual executed power value to the upper-level energy management system in real time, thus forming a closed-loop interaction of command, execution, and feedback. The power commands and actual power timing data generated by this interaction are the key information sources for subsequently extracting the operating fingerprint of the new site and evaluating its scheduling mode and execution consistency. Step 102 (Extraction and smooth update of fingerprint vectors for the new site): A set of features characterizing the long-term inherent operating pattern of the new site, sensitive to differences between new sites, and robust to short-term fluctuations are extracted to form the fingerprint vector of the new site. This includes: calculating the empirical cumulative distribution of the temperature sequence within the calculation window and fitting it to a Beta distribution using the method of moments (MoM); recording the two shape parameters α and β (temperature spectrum shape parameters, calculated by fitting the empirical distribution of the temperature sequence within the sliding window to a Beta distribution using MoM); and calculating the average daily difference between the highest and lowest temperatures, denoted as the daily temperature difference mean A; for the power command sequence, processing it through a Butterworth low-pass filter with a cutoff frequency of 0.01 Hz to separate the slowly varying fundamental component, and calculating the mean and standard deviation of this fundamental component. The process involves calculating the zero-crossing rate (power micro-cycle intensity) of the difference between the original command and the fundamental component (i.e., the fluctuation component), and counting the number of battery charge / discharge state transitions within the entire window. Next, it identifies all time periods (rest periods) where the continuous current amplitude is below a set threshold, and calculates the proportion of the total rest period duration to the window duration. It also identifies rest periods where the current of the energy storage battery continuously falls below a set threshold during operation. For each rest period, it extracts terminal voltage data within a preset time period after its start. The voltage data is then fitted to a first-order exponential decay curve, and a time constant characterizing the voltage relaxation rate is extracted from the fitting result. Statistical processing is performed on the time constants successfully fitted for all rest periods within the same analysis window, and the median value is used as a characteristic representing the dynamic response speed of battery polarization at that site. The threshold is defined as the absolute value threshold of the current change rate, used to identify significant current transient events during battery operation. The specific value is determined based on the battery type, rated capacity, and sampling noise level: For large energy storage cells with a rated capacity of 100Ah or more, the threshold is typically set at 0.1C / s (i.e., 0.1 times the rated current per second), corresponding to an absolute current change rate of approximately 10A / s; for smaller power cells, the threshold can be appropriately increased to 0.2C / s. In practical engineering applications, the principle for setting this threshold is: it should be greater than the maximum current fluctuation caused by noise during normal operation (usually three times the noise standard deviation) to ensure effective capture of real current surges caused by load switching or command jumps, while avoiding false noise triggering.

[0030] In one possible embodiment, the relaxation time constant is obtained as follows: For each rest period, voltage data within 30 minutes of the start were selected, and the data were fitted in the form of... The first-order exponential decay curve is used to obtain the relaxation time constant τ (polarization time constant during the rest period), and the median of all effective τ values ​​is taken as the characteristic. Here, V∞ represents the steady-state voltage after the relaxation process ends, that is, the terminal voltage that the battery eventually stabilizes after a sufficiently long rest period (current close to zero); ΔV represents the difference between the initial voltage and the steady-state voltage, reflecting the voltage jump caused by the instantaneous removal of the ohmic voltage drop and polarization voltage at the beginning of the rest period; the time constant of the relaxation process is the time required for the curve to decay to 1 / e (approximately 36.8%) of the initial difference (ΔV), which directly characterizes the dynamic response speed of the battery polarization process and is related to the product of the battery's polarization resistance and polarization capacitance; t represents the time variable starting from the rest period.

[0031] The system calculates the total number of triggers, the proportion of active balancing events to the total number of events, and the average gradient of voltage changes during balancing from the balancing action flags. It also captures all moments when the current rate of change exceeds a set threshold, calculates the voltage drop values ​​within a short time window before and after that moment, and averages these drop values. Finally, it combines all these features in sequence to form a new site fingerprint vector F (i.e., combining all features in sequence to form a temporary fingerprint vector based on the current window). ; In this invention, the battery model parameters specifically refer to the key physical characteristic parameters used to construct the temperature-affected equivalent circuit state-space model (such as first-order RC, second-order RC, etc.). These parameters directly describe the internal electrical dynamics of the battery and mainly include: ohmic internal resistance (characterizing the voltage drop generated when current instantaneously passes through), polarization resistance and polarization capacitance (combined to characterize the dynamic characteristics of the battery relaxation process, their product being the polarization time constant), and rated capacity (used to convert the current integral into a basic scale for state-of-charge change). These parameters are not fixed and will change with battery aging, temperature, and usage history. In this invention, through the knowledge transfer mechanism in step S2, initial values ​​or prior ranges are extracted and fused from the operating experience of similar historical sites to generate local battery model parameters, which serve as the personalized starting point for state estimation at new sites. These parameters are the core inputs of the state estimation algorithm, and their accuracy directly determines the accuracy of state-of-charge estimation.

[0032] In this invention, the state estimator parameters specifically refer to the algorithm configuration and adjustment parameters required when performing the bi-objective adaptive optimization using filters (such as extended Kalman filters and their variants). These parameters do not directly describe the battery's physical characteristics but are used to configure the convergence, robustness, and sensitivity of the estimation process. They mainly include: the process noise covariance matrix (used to model the uncertainty of the state transition equation), the observation noise covariance matrix (used to model the statistical characteristics of voltage measurement noise), and possibly filter gain adjustment coefficients or attenuation factors. In this invention, the initial settings of these parameters are also retrieved from the parameter and strategy library and interpolated in step S2 to ensure that the state estimator at the new site can start and run with a pace and confidence level that matches the characteristics of the local battery model.

[0033] The explanation is as follows: the temperature spectrum morphology parameter is used to quantify the long-term climate environment differences between new sites, providing a key dimension for S2 similarity retrieval; the power microcycle intensity is used to characterize the scheduling mode characteristics of new sites; and the polarization time constant during rest periods is used to reflect the inherent dynamic response characteristics of the battery.

[0034] In one possible embodiment, the fingerprint vector is updated using an exponential smoothing strategy, which adjusts the currently stored fingerprint vector... With the new temporary fingerprint vector Through formula The process is iterated, with the smoothing factor λ taking a value of 0.8, generating and storing new... This smoothing process aims to maintain the stability of the fingerprint vector.

[0035] Step 103 (Dynamic calculation of state credibility evidence vector): (Within a short evaluation window, such as 15 minutes), a set of metrics for evaluating real-time data quality and battery model observability is dynamically calculated, forming a state credibility evidence vector, including: Calculate data quality indicators: directly count the proportion of sampling anomaly flags that are set within the short window; calculate the root mean square error of the difference between voltage and current sampling timestamps; The richness of the input excitation is evaluated through two complementary metrics. Frequency domain effectiveness is assessed by performing a Fast Fourier Transform on the current sequence within the window and calculating the distribution ratio of its energy within the critical frequency band of battery polarization dynamics from 0.001 Hz to 0.1 Hz. This metric directly reflects the excitability of the current excitation signal for the core dynamic parameters of the battery model. Temporal activity is evaluated by calculating the average value of the absolute value sequence of the current rate of change within the window, denoted as μ_|dI|. This metric reflects the overall intensity of the input signal change. To assess the statistical consistency between the model and the data: obtain the voltage prediction residual sequence (i.e., voltage innovation residual) output by the state estimator at the previous time step, and calculate the following three statistics of the residual sequence within the current short window: the absolute value of the mean, used to assess the systematic bias; the absolute value of the first-order sample autocorrelation coefficient, used to assess the white noise characteristics of the residual; and the absolute value of the cumulative sum, used to assess the trend accumulation of error.

[0036] All the above indicators are normalized to the range of 0 to 1, and then combined in order to form a state credibility evidence vector C, which is stored in a circular buffer.

[0037] In this invention, the model observability index is an evaluative concept, not the model or parameters themselves. Specifically, it refers to a series of quantitative evidence (such as the energy percentage of a specific frequency band) calculated in step S1 by analyzing short-term data (such as the frequency domain characteristics of current excitation) to quantitatively determine whether the current operating conditions are sufficient to effectively excite battery dynamics and reliably identify the battery model parameters. In short, the battery model parameters are the unknowns to be solved, while the model observability index is a conditional criterion for evaluating whether the unknowns can be solved under the current conditions.

[0038] Step S2: Establish a parameter and strategy library, which records the prior range of equivalent circuit parameters, filter noise setting prior, gating threshold prior, and fingerprint vector summary of new sites; perform fingerprint similarity retrieval on the fingerprint vector of new sites to obtain candidate library records, and generate local battery model parameter prior, local noise setting prior, and local gating threshold prior based on fingerprint distance weighted interpolation, while generating negative migration risk level; Background: To avoid risks such as significant deviations in state of charge estimation, continuous parameter drift, and even false protection triggers in the early stages of operation due to directly applying fixed or default battery model parameters to new sites, step S2 aims to quickly provide personalized initial configurations for each new site. Specifically, based on the fingerprints collected during the initial operation of the new site, intelligent retrieval and interpolation are performed from a pre-built parameter and strategy library covering historical experience from multiple sites. This automatically generates localized battery model parameter priors, filter configurations, and decision thresholds, and quantitatively assesses the potential risks of this knowledge transfer, thereby replacing the traditional, inefficient, and expert-experience-dependent manual parameter tuning.

[0039] Furthermore, the operation of step S2 includes the following steps: Step 201 (Building and maintaining the parameter and strategy library): A parameter and strategy library is constructed and maintained. Each record in the library is associated with and stores four types of information: the first type is the initial value range (represented by minimum and maximum values) of key parameters of the equivalent circuit model (ohmic internal resistance, polarization resistance, and polarization capacitance in the battery model); the second type is the initial setting value of the process noise and observation noise covariance used by the state estimation filter; the third type is the gating threshold (such as parameter freezing threshold) of the control parameter update strategy; and the fourth type is the fingerprint summary vector of the new energy storage site corresponding to the record. The fingerprint summary vector is composed of key features selected from the complete long-term operating fingerprint of the new site, such as temperature distribution shape parameters, fundamental mean of power command, power fluctuation zero-crossing rate, current rest duration ratio, and median voltage relaxation time constant. Step 202 (Weight calculation for retrieval and evidence moderation based on weighted fingerprint distance) For a new site, based on the fingerprint vector summary vector generated in step S1, the weighted Euclidean distance between it and the fingerprint summary vectors corresponding to all records in the parameter and policy library is calculated. The weighted Euclidean distance is calculated as follows: First, the fingerprint summary vector of the new site and the fingerprint summary vectors of the records in the library are normalized in each feature dimension. Then, the difference between each normalized dimension is multiplied by the weight coefficient of the corresponding dimension in a pre-set feature importance weight vector. Finally, the sum of squares of the weighted differences of all dimensions is calculated and the square root is taken to obtain the final weighted distance value. According to the calculated weighted distance value, the K library records with the smallest distance value are selected as the candidate record set, where K is a preset positive integer (e.g., 3 or 5). The initial weight of each record in the candidate set is calculated. The initial weight is calculated by using a Gaussian kernel function based on its weighted distance value. The smaller the distance value, the larger the initial weight. The evidence adjustment factor η is read, and the initial weight of each candidate record is multiplied by the evidence adjustment factor η and then normalized to obtain the final weight. In one possible embodiment, the evidence adjustment factor is obtained by reading key quality indicators (sampling anomaly ratio) from the current state credibility evidence vector of the new site. and root mean square value of sampling synchronization error According to the formula The evidence moderating factor was calculated. ,in and The positive real number adjustment coefficient is preset based on experience and is used to control the sensitivity of different quality indicators; the value of the evidence adjustment factor ranges between (0,1], and the lower the value, the worse the current data quality. Among them, and These are pre-set positive real-valued adjustment coefficients based on experience, used to control the sensitivity of the sampling anomaly ratio and sampling synchronization error to the evidence adjustment factor; typically, and The value ranges from 1 to 10, and the specific value can be determined through offline simulation or optimization using historical data; for example, it can be set to... =5、 =3, so that when the sampling anomaly ratio reaches 0.1 or the root mean square of the synchronization error reaches 0.01, The weight of low-quality data is reduced to around 0.6. In engineering practice, the weight of low-quality data can be reduced appropriately. However, it is generally sufficient to keep the weight of low-quality data at a fixed value.

[0040] In one possible embodiment, the feature importance weight vector is obtained through the following steps: selecting a set of historical records from the parameter and policy library to form a training set; for each feature of the fingerprint summary vector, calculating the absolute value of the Spearman rank correlation coefficient between the sequence of the feature value on the training set and the nominal value of the key battery model parameter (such as ohmic internal resistance) in the corresponding record; and normalizing the absolute values ​​of the correlation coefficients of all features (making their sum equal to 1) to obtain the weight vector.

[0041] Step 203 (Parameter-Specific Fusion and Migration Risk Assessment): Using the final weights and candidate record set output in step 202 as input, parameter-specific weighted fusion is performed to generate local prior configurations and simultaneously calculate the negative migration risk level. The parameter-specific weighted fusion divides the parameters to be fused into two categories: The first category consists of battery model parameters that are sensitive to the operating environment and have significant differences between sites (such as ohmic internal resistance and polarization resistance). A conservative, quantifiable divergence range fusion strategy is adopted, that is, firstly, the weighted average and weighted standard deviation of each parameter are calculated to measure consensus divergence, then a preliminary confidence interval is constructed with the weighted average as the center and the weighted standard deviation multiplied by the expansion coefficient as the half-width, and finally, the intersection of this interval and the original prior range of all candidate records is taken to generate a local battery model parameter prior in the form of a value range; The second category consists of algorithm configuration parameters, including filter noise parameters and various gating thresholds. A direct weighted averaging strategy is adopted, and the corresponding set values ​​in the candidate records are weighted and averaged according to the final weights to generate local noise setting priors and local gating threshold priors in the form of specific values.

[0042] It should be noted that the expansion coefficient is a positive real number used to adjust the width of the confidence interval, typically ranging from 1 to 3. The specific value can be determined based on the number of candidate records, the historical fluctuation characteristics of the parameters, and the desired confidence level. For example, when the number of candidate records is large and the parameter distribution is approximately normal, an expansion coefficient of 2 can be used. In this case, the constructed interval corresponds to approximately 95% confidence level, which can better balance the accuracy and conservatism of the prior. In practice, this coefficient can also be optimized through offline cross-validation to minimize the initial estimation error of the local prior generated based on this confidence interval on the new site.

[0043] To quantify the reliability of this knowledge transfer, the negative transfer risk level is calculated simultaneously: Key parameters from the first category of parameters are selected, and the ratio of their weighted standard deviation to their weighted average is calculated to obtain the relative divergence of each parameter. The average relative divergence is then calculated. Simultaneously, the average weighted fingerprint distance between the candidate record set and the new site is calculated. Finally, the average relative divergence is multiplied by the average weighted fingerprint distance to obtain the negative transfer risk level. A higher level indicates greater potential uncertainty and risk in parameter transfer.

[0044] Step S3: Construct a bi-objective adaptive optimization state-of-charge estimator within the equivalent circuit state-space framework with temperature influence. Adaptively adjust the weights of innovation consistency constraints and conservation consistency constraints based on state credibility evidence and negative migration risk level. Apply constraints to the model parameter update magnitude to output the state of charge and the corresponding voltage innovation residual and conservation residual. Background: Steps S1 and S2 output structured state credibility evidence, localized battery model parameter priors, and negative migration risk levels. Step S3, based on the battery equivalent circuit state-space model with temperature influence, extends the traditional single-objective state estimation into a dual-objective adaptive optimization problem, simultaneously pursuing innovative sequence consistency between voltage observations and model predictions, and conservation consistency between state of charge changes and charge input. The weights of the state credibility evidence and negative migration risk levels are dynamically adjusted according to the state credibility evidence and negative migration risk levels. Under strict parameter change continuity constraints, the optimal state of charge estimate and corresponding residual sequence are output. The technical motivation is that SOC estimation methods based on extended Kalman filtering (EKF) or its variants simply pursue optimal voltage fitting. When the quality of observation data deteriorates or there are unmodeled biases in the battery model parameters, it may lead to over-adjustment or even divergence of the filter. Although the ampere-hour integration method can guarantee charge conservation, it cannot correct accumulated errors and is sensitive to the initial SOC. Therefore, a mechanism is needed that can use voltage information to correct when the observation information is reliable, rely on physical conservation laws to stabilize the estimation results when the observation information is unreliable or the model credibility is low, and ensure smooth online parameter updates to avoid drastic changes in model characteristics.

[0045] The innovation consistency constraint refers to the optimization objective of ensuring consistency between the predicted terminal voltage of the battery equivalent circuit model and the actual measured voltage during state estimation. This constraint is achieved by minimizing the statistical measure (such as the normalized sum of squares) and its autocorrelation penalty term of the voltage prediction residual sequence within a sliding time window. Essentially, it aims for the best fit between the model output and external observation data. The conservation consistency constraint refers to the optimization objective of ensuring consistency between the change in state of charge predicted by the battery model's state equation and the theoretical change calculated according to the law of charge conservation (Coulomb integral). This constraint is achieved by minimizing the statistical measure (such as the normalized sum of squares) and its cumulative deviation penalty term of the difference between the two within a sliding time window. Essentially, it aims for a strict fit between the model's internal state calculations and physical laws. The weights between the two are dynamically adjusted based on the reliability of the real-time evaluation data, the sufficiency of the incentives, and the risk of prior knowledge transfer.

[0046] In this invention, the equivalent circuit state-space framework and the state-space model are the same technical concept. The latter refers to a battery modeling method based on an equivalent circuit as the physical foundation and state-space equations as the mathematical expression. It defines battery state of charge, polarization voltage, etc., as state variables, and establishes quantitative relationships between these variables and external current, voltage, and model parameters through state equations and observation equations. The state-space model adopts a first-order RC equivalent circuit structure, and its state-space equations are defined as follows:

[0047] in, , These are the state of charge and polarization voltage (state variables) at time k, respectively. Δt is the measured current at time k (gated state machine input); Δt is the sampling period; η is the coulomb efficiency; This refers to the battery's rated capacity. , These are the polarization resistance and polarization time constant at temperature Tk, respectively. ; , These are the estimated values ​​of the state of charge and the estimated values ​​of the polarization voltage state predicted by the model at time k+1, respectively. It is an exponential decay factor in the state-space equation, and its physical meaning is: in one sampling period Internally, the battery polarization voltage is based on the polarization time constant at the current temperature. The completed natural relaxation (degradation) ratio, this factor transforms the continuous-time battery dynamics into discrete-time state transition relationships. The observation equation (voltage prediction equation) is as follows: ; in, The model terminal voltage prediction value at time k; This is the battery open-circuit voltage-state-of-charge mapping function; For temperature The internal resistance of the ohm below.

[0048] Furthermore, the battery open-circuit voltage-state-of-charge mapping function This function characterizes the one-to-one nonlinear relationship between the battery's state of charge (SOC) and its static open-circuit voltage (OCV) under specific temperature and aging conditions. This function is an inherent property of the battery and, for a given battery type, can be obtained through a standard experimental procedure: the battery undergoes a complete charge-discharge cycle under controlled temperature conditions, with sufficient rest periods interspersed, and a series of open-circuit voltage values ​​and corresponding SOC values ​​at equilibrium states are measured and recorded, thus obtaining discrete OCV-SOC data point pairs. Based on these data points, conventional mathematical methods, such as constructing lookup tables or using polynomial fitting, spline interpolation, etc., can be used to establish a continuous or piecewise function mapping relationship. Once calibrated, this function can be used as a known component of the observation equations in the state-space model, transforming the state variable SOCk into the corresponding predicted open-circuit voltage value. Therefore, obtaining or calibrating this mapping function is a fundamental and routine preliminary step in implementing the model described in this invention.

[0049] Furthermore, the state of charge estimator receives the prior local battery model parameters and local noise setting generated in step S2, as well as the state credibility evidence and negative migration risk level provided in step S1 as inputs. Based on these inputs, within the state-space model, it solves a bi-objective optimization problem with dynamically weighted evidence and risk, while simultaneously applying physical constraints on parameter range and update magnitude. Finally, it calculates and outputs the optimal state of charge estimate in real time, along with voltage innovation residual sequences and conservation residual sequences for performance self-evaluation. Furthermore, the specific implementation process of step S3 includes: Step 301 (State-space model construction and prior information injection): An equivalent circuit state-space model of the battery is established. The state variables of this model include the state of charge and polarization voltage. The battery model parameters include ohmic internal resistance, polarization parameters, etc., and the values ​​of these parameters are compensated and corrected based on the real-time measured temperature. The prior local battery model parameters output in step S2 are used as the initial estimated values ​​of the battery model parameters, and the range of parameter values ​​defined therein is used as the boundary constraint for subsequent parameter updates. The prior local noise settings output in step S2 are used as the initial settings of the process noise and observation noise co-matrix in the state estimation. Step 302 (Construction of an evidence- and risk-driven bi-objective optimization function): Construct a comprehensive cost function J for simultaneously optimizing voltage fitting consistency (innovation consistency) and charge conservation consistency (conservation consistency), which has the form: ,in, Calculate the sum of squares of the normalized voltage prediction residuals within the sliding window, and add a penalty term proportional to the absolute value of the first-order autocorrelation coefficient of the residual sequence. The sum of squares of the difference between the normalized SOC increment and the current integral within the sliding window is calculated, and a penalty term proportional to the absolute value of the difference between their cumulative sums is added. The adaptive weight α is obtained through a Sigmoid function, whose input is a linear combination of the state credibility evidence provided in step S1 (including sampling failure rate, current excitation band energy, voltage residual bias, etc.) and the negative migration risk level provided in step S2. The combination coefficients are determined through learning from historical data, so that α approaches 1 when the data quality is high, the excitation is sufficient, and the risk is low, and approaches 0 otherwise. The specific method for dynamically adjusting weights is as follows: Indicators such as sampling failure rate, current excitation band energy, and voltage residual bias are selected from the state credibility evidence and combined linearly with the negative migration risk level to obtain a comprehensive decision variable. The weighting coefficients of each indicator can be determined through optimization using historical data or preset based on expert experience. Subsequently, this comprehensive variable is input into an S-shaped function for nonlinear mapping, outputting an adaptive weight value between 0 and 1. When data quality is high, excitation is sufficient, and risk is low, the weight approaches 1, and optimization focuses on voltage fitting; conversely, the weight approaches 0, and optimization focuses on charge conservation, thereby achieving dynamic weight adjustment based on real-time evidence and risk.

[0050] Furthermore, the penalty item Specifically defined as the voltage innovation residual sequence within the sliding window. The absolute value of the first-order sample autocorrelation coefficient, i.e.:

[0051] Where N represents the total number of data points (or sampling times) contained within the sliding window; i is the time-series index of the data point within the window. The penalty term is the mean of the residuals within the window. Specifically defined as SOC increment sequence With current integral sequence The absolute value of the difference between the cumulative sums within the sliding window, i.e.:

[0052] The explanation explains that the voltage innovation residual sequence refers to the voltage prediction residual values ​​arranged in chronological order within a sliding window. Each residual value is the difference between the actual measured voltage and the voltage prediction value obtained based on the equivalent circuit model and its state estimation. The absolute value of the first-order autocorrelation coefficient of this sequence can effectively reflect whether there are systematic biases or trend errors in the residuals (e.g., caused by unmodeled temperature transients or sensor drift). If this value is large, it indicates that there may be unmodeled errors in the voltage observation equation. In this case, the dependence on the voltage fitting target should be reduced, and the confidence in the physical conservation target should be increased. The SOC increment sequence and current integral sequence are used to construct the physical conservation target (…). The basis of this (charge conservation consistency) is as follows: each value in the SOC increment sequence is the predicted change in state of charge from the previous moment to the current moment, based on the state-space equations and using the model parameters to be optimized; while each value in the current integral sequence is the theoretical change in state of charge directly calculated from the measured current through coulomb counting, based on the law of charge conservation. The conservation consistency cost function minimizes the difference between these two sequences (i.e., the deviation between the change predicted by the model and the change required by the external physical laws), forcing the state estimation process to always conform to the principle of charge conservation, thus providing a stable physical benchmark for SOC estimation when voltage observation information is unreliable.

[0053] Step 303 (Constraint optimization solution and state and residual output): At each estimated time step, the optimal state of charge (SOC) value and updated battery model parameters are solved with the goal of minimizing the comprehensive cost function J. The solution process must satisfy two constraints: first, the updated battery model parameters must not exceed their local prior range; second, the norm of the change of the battery model parameter vector at adjacent time steps must not exceed a preset small threshold (e.g., 5% of the width of its prior range). This constrained optimization problem is solved using a sequential quadratic programming algorithm. After solving, the estimated SOC value at the current time step is output, and the voltage innovation residual sequence and charge conservation residual sequence within the current sliding window are recorded as the output of this step.

[0054] Step S4: Accumulate evidence for the voltage innovation residual, conservation residual, state credibility evidence and negative migration risk level, drive the gating state machine to perform parameter freezing or parameter rollback under different gating states, and start minimum cost self-calibration to update the identifiable parameter subset when a natural step or observable excitation condition window is detected, and write the update result back to the parameter and strategy library.

[0055] Background: Steps S1 to S3 complete the entire process from recognizing new site characteristics and generating initial strategies to online adaptive estimation. Given that a truly scalable gated state machine must possess self-monitoring, management, and evolution capabilities, step S4, based on long-term accumulated multi-source evidence, employs a gated state machine intelligent decision-making parameter update strategy. When evidence is insufficient or risk is too high, parameters are frozen or rolled back to ensure estimation safety. When a specific favorable condition is detected, low-cost, highly targeted parameter self-calibration is initiated, and the calibration results are smoothly written back to the parameter and strategy library, achieving new site-level knowledge recognition. The continuous accumulation and closed-loop evolution of knowledge, step S4 is used to solve the fundamental contradiction faced by the online adaptive estimator, namely, overuse of the latest data can easily lead to noise overfitting and parameter drift, while excessive conservatism makes it difficult to track the actual aging and changes of the battery. Moreover, the complete identifiability of battery model parameters depends on sufficient input stimulation, which is often not met during the smooth operation of energy storage power stations. Therefore, a decision-making mechanism that exceeds fixed thresholds and has state memory is needed to determine the learning opportunity, the pause opportunity, and the learning focus, so that the gated state machine has the ability to accumulate experience and performs better when facing similar operating conditions or migrating to a new site.

[0056] Furthermore, the natural step refers to a current step change caused by load or scheduling command, which exceeds a threshold (e.g., 0.2C) and lasts for a certain period of time; Furthermore, the observable excitation condition window refers to the current fluctuation sequence with a certain signal-to-noise ratio and duration generated by natural load fluctuations or small grid adjustments during the steady-state operation of the gated state machine (for example, the current fluctuates continuously for more than 1 minute within the range of ±0.05C). Although this condition is insufficient to identify all parameters, it can be combined with prior information to meaningfully constrain and calibrate some highly sensitive parameters.

[0057] Furthermore, the specific implementation process of step S4 includes: Step 401 (Spatiotemporal fusion and comprehensive confidence calculation of multi-source evidence): The system continuously receives the voltage innovation and conservation residual sequences output from step S3, the state credibility evidence vector output from step S1, and the negative migration risk level output from step S2. It analyzes the two types of residuals separately: calculating the absolute value of their statistical bias within the most recent 1-hour time window to assess immediate model mismatch; simultaneously, calculating the absolute value of their trend slope obtained through linear regression within the most recent 24-hour time window to capture slow parameter drift; and based on the current current excitation frequency band energy... The level of the fusion weight is dynamically determined based on its magnitude. This is used to weight and combine short-term biases with long-term trends to form a comprehensive residual measure. Finally, this comprehensive residual measure, the negative migration risk level, and credibility evidence such as data failure rate are substituted into a predefined linear combination formula and mapped using the Sigmoid function to generate a comprehensive confidence index between 0 and 1. Comprehensive confidence index The lower the value, the worse the overall reliability. The fusion weight W is dynamically determined based on the energy level of the current excitation frequency band. Specifically, the energy value is first normalized to the zero-to-one range, and then the weight is obtained through a piecewise linear function mapping. For example, a low-energy threshold and a high-energy threshold are set. When the energy is below the low threshold, a smaller weight (e.g., 0.2) is used, and when it is above the high threshold, a larger weight (e.g., 0.8) is used, with linear interpolation in the intermediate range. A larger weight indicates a higher proportion of short-term residuals in the overall metric; conversely, a smaller weight indicates greater reliance on long-term trends. The thresholds can be preset based on the battery's dynamic characteristics and sampling frequency, such as a low threshold of 0.1 and a high threshold of 0.5, or optimized and adjusted using historical data.

[0058] The explanation explains that the comprehensive confidence index CIT integrates instantaneous, short-term, and long-term evidence, and its calculation itself has low-pass filtering characteristics, avoiding drastic fluctuations in the index caused by single-point noise. Importantly, the state transition design of the gated state machine has hysteresis and state memory: for example, the transition from a safe fallback state to an active learning state must satisfy… The condition is satisfied for at least T sampling periods (e.g., T=60), rather than instantaneously. This decision-making mechanism based on accumulated evidence and state history memory ensures the robustness and security of the state transition of the gated state machine, which is fundamentally different from simple comparison based on instantaneous thresholds.

[0059] Step 402 (Gated State Decision and Execution Based on Confidence and Evidence): A gated state machine is pre-defined, comprising three core states: active learning, observation freeze, and safe rollback; state transitions are determined by a comprehensive confidence index. Compared with the local gating threshold generated in step S2 ( The comparison results are driven by the comparison, and the evidence is taken into account, such as the input stimulus level. in, It refers to the active learning trigger threshold, which is the minimum threshold value that must be met to drive the gated state machine into or maintain the active learning state, and which characterizes the overall confidence level of the system. This refers to the observation freeze trigger threshold, which is the critical value that drives the gated state machine to transition from the active learning state to the observation freeze state, representing a decrease in the overall confidence level. This refers to the safety rollback trigger threshold, which is the bottom line value that drives the gated state machine into the safety rollback state and represents the overall confidence level deteriorating to a level of safety risk. Among them, the gate threshold These correspond to the confidence boundaries for active learning, observation freeze, and safe rollback states, respectively; typical value ranges are: Between 0.7 and 0.9, Between 0.4 and 0.6, The value is between 0.2 and 0.4; the value is determined by the system's trade-off between estimation accuracy and safety, and can be obtained through offline simulation optimization: with the goal of minimizing estimation error and controlling parameter drift, the optimal combination is searched on historical data in a grid. In practical applications, it can be fine-tuned according to the characteristics of the site to ensure the robustness of state switching.

[0060] when Furthermore, when the current excitation is sufficient, it is in an active learning state, allowing step S3 to perform a normal update of the entire parameter set; when Descending to If the interval or excitation remains insufficient, transition to the observation freeze state. At this point, tighten the parameter update magnitude constraint in step S3 (e.g., halve the maximum rate of change coefficient), or directly freeze the polarization parameters that are insensitive to the excitation. Simultaneously, the SOC output includes a parameter related to... A confidence interval that is directly proportional; when If a sharp deterioration in the residual is detected in the short term, a safe rollback state is entered, completely prohibiting parameter updates in step S3, and the option to restore to a historical reliable parameter snapshot is available. The SOC output turns to a high uncertainty state and triggers an alarm. Step 403 (Natural step perception and minimum cost parameter self-calibration and knowledge write-back): During the operation of the gated state machine, the current sequence is monitored in parallel. When the absolute change in current exceeds a threshold (e.g., 0.5C) and remains stable at the new value for more than 30 seconds, a natural step excitation window is identified. Then, minimum-cost self-calibration is initiated: within a short calibration window (e.g., 2 minutes), only the top two parameters (e.g., ohmic internal resistance and main polarization resistance) pre-selected based on the product of parameter sensitivity and prior uncertainty are calibrated. Optimization is performed; the optimization objective is to minimize the sum of squares of the differences between the observed voltage values ​​and the model predictions within the window, while satisfying the parameter boundary and maximum variation constraints, and to add a regularization term with the prior central value as a reference; the optimization solution yields new parameter estimates. Then, with a smaller smoothing factor (e.g., 0.2) according to the formula Update the local prior; finally, associate the updated local prior and the adjusted noise setting with the latest fingerprint digest of the new site, and write it back to the parameter and policy library as a new record to complete this learning loop. in, The prior values ​​of the local battery model parameters stored before this self-calibration operation can be the initial values ​​generated in step S2 or the values ​​that have been updated after previous self-calibrations. The smoothing factor η is a preset constant between 0 and 1, which controls the magnitude and speed of the update of the new estimate to the original prior. The smaller the value, the more conservative and smooth the update. This refers to the updated prior values ​​of the local battery model parameters calculated using the smooth update formula described above. These values ​​will be overwritten in storage for subsequent state estimation and associated with a fingerprint summary written back to the parameter and policy library.

[0061] In one possible embodiment, the parameter sensitivity is approximated by calculating the absolute value of the partial derivative of the equivalent circuit model output voltage with respect to the parameter at the current operating point; the prior uncertainty is characterized by quantifying the relative dispersion of the parameter in the prior parameters of its local battery model. (with Ohm internal resistance) For example, The maximum value set for the parameter of ohmic internal resistance. The minimum value is set for the ohmic internal resistance parameter; the larger the formula output, the more uncertain the prior knowledge of the parameter is. The sensitivity of each parameter is normalized and multiplied by this uncertainty quantification value to obtain its sensitivity-uncertainty product. Based on this, the parameters with the largest products are selected to form the subset to be calibrated.

[0062] The write-back operation follows these evolutionary rules: a) Record structure: Each record contains the updated parameter prior vector, the corresponding operating condition fingerprint (including average SOC, temperature, and current statistical features), and the evidence strength of the calibration window (such as excitation energy and fitting residuals). b) Fingerprint matching and fusion: When the Euclidean distance between the fingerprint of a new record and the fingerprint of a record in the database is less than the threshold, it is considered a similar working condition; the new and old parameter priors will be weighted and averaged according to the strength of evidence to generate a new prior after fusion and increase the confidence weight of the record. c) Add a new record: If no matching fingerprint is found, add a new record; This mechanism enables the knowledge base to continuously absorb site operation experience, and its output prior is a high-confidence experience value reinforced by long-term multi-source evidence, realizing a closed-loop evolution from online instantaneous learning to offline group knowledge accumulation.

[0063] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for detecting the state of charge (SOC) of a lithium-ion battery based on adaptive optimization, comprising collecting current, voltage, and temperature data of the battery at a new site; establishing an equivalent circuit state-space model of the battery, wherein the state variables of the model include the SOC and polarization voltage, and the model parameters include ohmic internal resistance and polarization parameters, wherein the model parameters vary with temperature, characterized in that... Includes the following steps: Step S1: Collect multi-source runtime sequence data of the new site, extract and store fingerprint vectors that characterize the long-term operating mode of the new site, as well as state credibility evidence used to evaluate real-time data quality and model observability. Step S2: Based on the fingerprint vector, perform similarity retrieval from the pre-built parameter and strategy library. Based on the retrieval results, generate local battery model parameter priors, local noise setting priors, and local gating threshold priors for the new site through weighted interpolation, and simultaneously generate negative migration risk levels. Step S3: Within the equivalent circuit state-space model framework, construct an innovation consistency constraint with voltage fitting as the objective and a conservation consistency constraint with charge conservation as the objective; dynamically adjust the weights of the two constraints based on state credibility evidence and negative migration risk level to form a dual-objective adaptive optimization problem; solve the optimization problem, apply a continuity constraint on the update magnitude of model parameters, output the optimal state of charge estimate, and output the voltage innovation residual sequence and conservation residual sequence. Step S4: Multi-timescale fusion and accumulation of voltage innovation residual sequence, conservation residual sequence, state credibility evidence and negative migration risk level to drive gating state machine to make state decisions; gating state machine performs parameter update, parameter freeze or parameter rollback based on accumulated evidence under different gating states. When a natural step or observable excitation window is detected, minimum cost self-calibration is initiated, updating only a subset of parameters pre-selected based on parameter sensitivity and prior uncertainty; the parameter update results obtained from self-calibration are then associated with the current operating condition fingerprint and written back to the parameter and policy library.

2. The method for detecting the state of charge of a lithium-ion battery based on adaptive optimization according to claim 1, characterized in that, The process of extracting the fingerprint vector includes: extracting temperature distribution morphology parameters, fundamental frequency statistical characteristics and fluctuation characteristics of power commands, and the proportion of current rest periods from the multi-source runtime timing data; and extracting polarization time constant characteristics from the voltage relaxation curve of the rest periods; and combining the features to form the fingerprint vector.

3. The method for detecting the state of charge of a lithium-ion battery based on adaptive optimization according to claim 2, characterized in that, In step S2, the fingerprint vector-based similarity retrieval and weighted interpolation process includes: calculating the weighted Euclidean distance between the fingerprint summary vector of the new site and the parameters and the fingerprint summary vectors of each record in the policy library; selecting a candidate record set based on the distance and calculating the initial weights; reading the evidence adjustment factor calculated based on the state credibility evidence, adjusting and normalizing the initial weights to obtain the final weights used for interpolation; the process of generating local model parameter priors through fingerprint distance weighted interpolation is limited by a set of dynamic credibility envelopes; the width of the dynamic credibility envelopes is jointly determined by the historical volatility of each parameter and the parameter identifiability calculated based on the short-term observation data of the current site.

4. The method for detecting the state of charge of a lithium-ion battery based on adaptive optimization according to claim 3, characterized in that, In step S2, the process of generating local battery model parameter priors adopts a parameter-specific fusion strategy: for battery model parameters, the weighted statistics of each parameter are calculated according to the final weights, and a local prior is generated in terms of value range based on the intersection of the weighted statistics and the original prior range of the candidate records; for filter noise setting priors and gate threshold priors, the corresponding values ​​of the candidate records are weighted and averaged according to the final weights to generate local priors in terms of specific numerical values.

5. The method for detecting the state of charge of a lithium-ion battery based on adaptive optimization according to claim 1, characterized in that, In step S3, the comprehensive cost function of the constructed bi-objective adaptive optimization problem is composed of a weighted sum of a voltage fitting consistency cost term and a charge conservation consistency cost term. The voltage fitting consistency cost term includes a penalty for the sum of squares of the normalized voltage prediction residuals within the sliding window, as well as an additional penalty for the absolute value of the first-order autocorrelation coefficient of the residual sequence. The charge conservation consistency cost term includes a penalty for the sum of squares of the difference between the predicted state of charge increment and the Coulomb integral increment within the sliding window, as well as an additional penalty for the absolute value of the difference between their cumulative sums.

6. The method for detecting the state of charge of a lithium-ion battery based on adaptive optimization according to claim 5, characterized in that, In step S3, the process of dynamically allocating the weights of the two constraints is as follows: the state credibility evidence and the negative migration risk level are linearly combined, and the combination result is input into a Sigmoid mapping function to output an adaptive weight value between zero and one, which is used to perform a weighted summation of the two cost terms in the comprehensive cost function.

7. The method for detecting the state of charge of a lithium-ion battery based on adaptive optimization according to claim 1, characterized in that, In step S4, the gating state machine includes an active learning state, an observation freeze state, and a safety rollback state; the state transition is driven by the comparison result of the comprehensive confidence index and the local gating threshold; when in the observation freeze state, the constraint on the model parameter update magnitude in step S3 is tightened or the update of some model parameters is frozen; when in the safety rollback state, the parameter update in step S3 is prohibited.

8. The method for detecting the state of charge of a lithium-ion battery based on adaptive optimization according to claim 1, characterized in that, In step S4, the process of initiating the minimum cost self-calibration to update the identifiable parameter subset includes: sorting the parameters according to the product of the sensitivity of each model parameter and its prior uncertainty, selecting the top few parameters with the largest product to form the subset to be calibrated; within the calibration time window corresponding to the natural step or observable excitation condition window, minimizing the voltage prediction error as the objective, and applying parameter boundary and update amplitude constraints to optimize the solution of the parameters in the subset to be calibrated.

9. The method for detecting the state of charge of a lithium-ion battery based on adaptive optimization according to claim 1, characterized in that, The observation method for the observable excitation condition window is as follows: Calculate the excitation dispersion index and the parameter stress sensitivity index; based on the positions of the excitation dispersion index and the parameter stress sensitivity index in the preset decision matrix, determine whether the current condition is an observable excitation window, and decide whether to trigger and to what degree of conservatism to trigger the self-calibration accordingly. The operating condition excitation dispersion index is obtained by performing frequency domain transformation on the current sequence within the preset evaluation window and calculating the distribution ratio of its energy in the key dynamic frequency band of the battery model; the parameter stress sensitivity index is obtained by calculating the sum of the absolute values ​​of the partial derivatives of the output voltage of the equivalent circuit model with respect to the preset key model parameter subset based on the state of charge and temperature of the current operating point. Based on the calculated operating condition excitation dispersion index and parameter stress sensitivity index, a preset two-dimensional decision matrix is ​​queried. The horizontal and vertical axes of the decision matrix correspond to the discretization levels of the two indices, respectively. Based on the cells mapped by the indices in the matrix, it is determined whether the observable excitation operating condition window conditions are met, and the level of conservatism to be adopted if self-calibration is triggered is determined simultaneously. The level of conservatism is associated with a set of self-calibration parameters, including the calibration time window length, parameter update regularization strength, and write-back smoothing factor.

10. The method for detecting the state of charge of a lithium-ion battery based on adaptive optimization according to claim 1, characterized in that, The evidence accumulation includes: performing transient filtering and trend extraction on the innovation residual and conservation residual sequences respectively; analyzing the correlation patterns between the two residual trend components and matching them with predefined model mismatch failure modes to generate a physical-oriented comprehensive model mismatch confidence level, which is used to drive the gated state machine.

Images