Self-adaptive repairing method of energy storage battery CCS board

Abstract

The invention discloses a self-adaptive repairing method for an energy storage battery CCS board, and relates to the technical field of energy storage battery management. According to the method, the problems of poor model adaptability and insufficient strategy rigidity and real-time performance in existing fault diagnosis and repair are solved. Comprising the following steps: constructing an exclusive digital twinborn body and a multi-level dynamic model library for a target battery pack; performing online calibration and incremental learning on model parameters during stable operation of the battery; calling multiple models in parallel to carry out state prediction, and realizing weighted fusion through confidence analysis to obtain an optimal state value and a confidence interval thereof; and according to the degree of confidence of fusion output, executing a hierarchical repair strategy from full-automatic regulation to redundant buffering until a conservative security mechanism is triggered. Individualized precise diagnosis and self-adaptive repair of the battery state are realized, and the reliability, the safety and the operation and maintenance efficiency of the energy storage system are improved.

Description

Technical Field

[0001] This invention relates to the field of energy storage battery management technology, specifically to a self-adaptive repair method for an energy storage battery CCS board. Background Technology

[0002] In energy storage battery systems, the CCS board, as a key component of the battery module, is responsible for functions such as electrical connection of the cells, voltage acquisition, temperature monitoring, and equalization management. Its working status directly affects the safety and performance of the battery system.

[0003] Currently, diagnostic and repair methods for CCS boards and their corresponding battery states mostly rely on rule-based judgments with fixed thresholds or single physical models, which have the following problems: Poor model adaptability: Battery systems are affected by various factors such as aging, environment, and usage habits during actual operation. Fixed models cannot accurately reflect individual differences and dynamic changes, leading to diagnostic biases; Rigid repair strategies: Traditional methods often adopt a "one-size-fits-all" repair strategy, which cannot provide graded responses based on the confidence level of the diagnostic results, easily resulting in over-repair or under-repair; Insufficient real-time and continuity: Most methods rely on historical data or offline training, making it difficult to achieve continuous model calibration and optimization throughout the battery's entire life cycle, causing diagnostic capabilities to degrade as the system changes.

[0004] Therefore, there is an urgent need for a method that can adapt to individual battery differences, optimize models in real time, and perform intelligent graded repair based on diagnostic confidence levels to improve the reliability, safety, and service life of energy storage systems. Summary of the Invention

[0005] The purpose of this invention is to address the problems of poor model adaptability, rigid repair strategies, and insufficient real-time performance in existing fault diagnosis and repair methods for energy storage battery CCS boards, and to propose an adaptive repair method for energy storage battery CCS boards.

[0006] The objective of this invention can be achieved through the following technical solution: a self-adaptive repair method for an energy storage battery CCS board, comprising: Step 1: Instantiate a dedicated digital twin model for the target battery pack and establish a multi-level dynamic model library corresponding to the digital twin model; Step 2: When the battery is in a stable range with small data fluctuations, the parameters of the digital twin model are calibrated in real time, and the model weights are updated based on the data-driven model. Step 3: Parallel call of the calibrated multiple models, perform consistency analysis on the prediction results of each model, construct confidence analysis based on error distribution, output dispersion and model confidence level, and obtain the optimal state value and its confidence interval after weighted fusion; Step 4: Based on the output status value and confidence interval, perform graded repair in conjunction with the current battery operating condition data.

[0007] As a preferred embodiment of the present invention, the process of constructing a dedicated digital twin model is as follows: A unique twin identifier is generated based on the identity information of the target battery pack; multi-dimensional characteristic data of the battery pack are collected from the factory and generated into cell-level feature vectors after preprocessing; an equipment-level electrical and thermal topology relationship matrix is ​​constructed according to the actual structure of the battery pack; based on the feature vectors and topology relationships, the physical model, data-driven model and empirical model are instantiated and encapsulated into a structured configuration file.

[0008] In a preferred embodiment of the present invention, the physical model is an improved equivalent circuit-thermal coupling model; the data-driven model includes a recurrent neural network subnetwork for time series prediction and an autoencoder subnetwork for anomaly detection; the empirical model library includes a general fault mode library, statistical models of similar products, and simplified diagnostic models based on physical rules.

[0009] In a preferred embodiment of the present invention, the conditions for real-time calibration are as follows: The system monitors the rate of change of current, voltage, and temperature. When the rate of change is lower than a preset threshold, it is marked as an identifiable interval. Within the stable interval, the internal resistance parameter is updated in real time using the least squares identification model. The thermal model parameters are inversely calculated using the thermal balance equation, and the dynamic parameter table is updated.

[0010] As a preferred embodiment of the present invention, the process of determining the optimal state value and its confidence interval after weighted fusion is as follows: Based on the calibrated and updated physical model, data-driven model, and empirical model, the prediction outputs of each model are obtained by parallel invocation; the dispersion between the prediction outputs of each model is calculated. Based on the historical prediction error distribution of each model and the preset model confidence level, the confidence level of each model is calculated. The predicted outputs of each model are weighted and fused based on their confidence levels to obtain the optimal state value; the confidence interval of the optimal state value is calculated based on the fusion residual distribution of the predicted outputs of each model.

[0011] In a preferred embodiment of the present invention, the process of performing graded repair is as follows: Based on the confidence interval of the optimal state value, the final confidence value is calculated; the final confidence value is then compared with multiple preset grading thresholds. Based on the comparison results, different levels of repair control strategies are implemented, including at least one of automated regulation, redundant buffer control, and conservative safety mechanisms.

[0012] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention improves the accuracy of state prediction and fault diagnosis by constructing a digital twin based on the unique identity of the battery and a multi-level dynamic model library, and by combining online calibration and incremental learning mechanisms to adapt to changes in the individual state of the battery in real time. 2. Based on the optimal state value and its confidence interval output by multi-model fusion, combined with confidence level threshold, this invention realizes a multi-level repair response from fully automatic adjustment to conservative safety mechanism, thereby improving the flexibility and safety of the repair strategy. 3. This invention reduces the risk of misjudgment by a single model through parallel model fusion and confidence analysis. It also supports automatic reporting of abnormal event data packets and human-computer interaction prompts, which facilitates rapid intervention by operation and maintenance personnel and improves the overall reliability and maintainability of energy storage. Attached Figure Description

[0013] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0014] Figure 1 This is a flowchart of the method steps of the present invention. Detailed Implementation

[0015] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0016] It should be understood that the terms “comprising” and “including” used in this disclosure and claims indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0017] It should also be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of this disclosure. As used in this disclosure and claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this disclosure and claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes such combinations.

[0018] Please see Figure 1 As shown, a self-adaptive repair method for an energy storage battery CCS panel includes the following steps: Step 1: Instantiate a dedicated digital twin model for the target battery pack and establish a multi-level dynamic model library corresponding to the digital twin model.

[0019] Step 2: When the battery is in a stable range with small data fluctuations, such as when it is idle or operating at low rate, the parameters of the digital twin model are calibrated in real time; the model weights are updated using an incremental learning strategy for the data-driven model.

[0020] Step 3: When analyzing the critical states of the battery pack, the calibrated multiple models are called in parallel, and the consistency analysis of the prediction results of each model is performed. Based on the error distribution, output dispersion, and model confidence level, a confidence analysis is constructed to obtain the optimal state value and its confidence interval after weighted fusion.

[0021] Step 4: Based on the output status value and confidence interval, and combined with the current operating condition data of the battery, perform graded repair.

[0022] The process of building a dedicated digital twin model and a multi-level dynamic model library is as follows: After obtaining the serial number, production batch number, and cell supplier code of the target battery pack, a unique digital twin identifier ID is constructed, where ID = f(SN, BN, SC), where SN is the serial number, BN is the batch number, and SC is the supplier code. The function f(·) is the uniqueness mapping function, which is implemented through a hash function.

[0023] Based on a unique digital twin model identifier (ID), multi-dimensional characteristic data of the battery pack are collected during the manufacturing stage, including electrical characteristic data, thermal characteristic data, mechanical structural characteristic data, and performance test data. To ensure the consistency and comparability of subsequent models, this embodiment adopts the 3σ principle to clean the above-mentioned multi-dimensional characteristic data, and uses a normalized mapping model x′=(x−μ) / σ to unify the data under different test environments to standard operating conditions (25°C, 50% SOC), where μ represents the mean, σ represents the standard deviation, and x′ is the normalized data point.

[0024] After data preprocessing, the cleaned multi-dimensional characteristic data generates cell-level feature vectors, which are represented as V. cell =[Cap, R0, C] poly (1–5), SDR, k T [SOH0, X, Y, Z], where Cap is the initial capacitance, R0 is the DC internal resistance, and C is the DC internal resistance. poly (1–5) are the 5th-order polynomial fitting coefficients of the OCV curve (obtained by least squares fitting), SDR is the self-discharge rate, and k is the self-discharge rate. T Here, SOH0 represents the temperature rise coefficient, SOH0 represents the initial healthy state, and X, Y, and Z represent the three-dimensional coordinates of the cell in the physical structure.

[0025] Based on the actual structural layout of the battery pack, a device-level topology is constructed, including the electrical connection matrix M. E With thermal coupling matrix M T M E (i, j) represents the electrical equivalent connection parameters between cells i and j, M T (i, j) represents its heat exchange equivalent parameters; Based on the above process, the instantiation stage of the digital twin model is entered, and physical models, data-driven models, and empirical model libraries are established according to model type. The physical model adopts an improved equivalent circuit-thermal coupling structure, including an electrochemical model and a thermal model. The basic calculation flow of the electrochemical model is expressed as V pred =OCV(SOC)−I×R0−Σ(R i ×I)−Σ(U i ), where V pred To predict the voltage, OCV(SOC) is the open-circuit voltage function, I is the current, and R is the resistance term. i With port voltage U i The parameters are derived from a second-order RC model, and all symbols are defined according to electrochemical meaning. The thermal model employs a finite element discretization process, using the node matrix H·T=Q to solve for thermal diffusion, where H is the thermal conductivity matrix, T is the node temperature vector, and Q is the heat generation power vector. The parameters of the above physical model are obtained by fitting HPPC data using a genetic algorithm, and each parameter is assigned an uncertainty in the form of a Gaussian distribution: θ ~ N(μ θ ,3σ θ 2 ), where θ is any physical parameter, μ θ To identify the mean, σ θ This represents the fitting error.

[0026] The data-driven model consists of two sub-networks, including a time-series prediction sub-network. This sub-network employs a Long Short-Term Memory (LSTM) network structure with three layers and 64, 32, and 16 hidden neurons respectively. The ReLU activation function is used. The input sequence S = {I(t), T(t)} represents current and temperature data collected within a continuous time window, with a window length of 60 seconds and a sampling interval of 1 second. This time-series prediction sub-network learns the dynamic response characteristics of the battery through training and outputs the predicted voltage V for the next time step. LSTM (t+1). The training data comes from the battery's historical operating data. The loss function is the mean squared error (MSE). The optimizer is Adam, and the initial learning rate is 0.001.

[0027] The anomaly detection model employs a stacked autoencoder with a symmetrical encoder-decoder structure, each containing three fully connected layers with 128, 64, and 32 neurons respectively. The sigmoid activation function is used. Input features X include real-time voltage, current, temperature, and their first-order differences, with a feature dimension of 6. The model learns the feature distribution under normal operating conditions through a combination of unsupervised pre-training and supervised fine-tuning. The reconstruction error E... rec Defined as the Euclidean distance between the input feature X and the reconstructed output X′, i.e., E rec =||X−X′||. When E rec When the error exceeds a preset threshold (such as 3 times the standard deviation of the reconstruction error of historical normal samples), it is judged as an abnormal state.

[0028] Uncertainty quantification of the data-driven model is performed using Monte Carlo Dropout technology. Dropout is kept active during the inference phase, and forward propagation is randomly performed N times (e.g., N=50) to obtain the probability distribution of the prediction results. By calculating the mean and standard deviation of this distribution, the predicted values ​​and their uncertainty analysis are obtained for subsequent model fusion.

[0029] The empirical model library is constructed using a three-layer structure, specifically including: General Fault Mode Model: Based on a library of typical fault cases, it includes feature rules and discrimination thresholds for common fault modes such as overvoltage, undervoltage, excessive temperature difference, and internal surge. Statistical model for similar products: Based on historical operational statistics of batteries from the same supplier, batch, or model, construct a model for parameter degradation trends and failure rate distribution; A simplified diagnostic model based on fundamental physical rules, including simplified physical relationships such as Ohm's law and thermal balance equations, is used to quickly verify the consistency of battery state.

[0030] Each model output includes a confidence level C. exp Through formula C exp =0.7M hist +0.3M rule Among them, M hist The historical matching degree is calculated based on the similarity between the current data and historical failure modes, with a value range of [0, 1]; M rule The completeness of the theory is scored based on the completeness and applicability of the physical rules used, with a value range of [0, 1].

[0031] After obtaining the physical model, data-driven model, and empirical model, a coupling mechanism between the models is further constructed; in this embodiment, the fused output takes the following form: Y final =W P ×Y P +W D ×YD +W E ×Y E , where Y P Y D Y E The output results of the three types of models are W respectively. P W D W E The initial weights for these weights are set to 0.5, 0.3, and 0.2.

[0032] After the model is built, three sub-processes are executed sequentially: cross-validation, twin file generation, and real-time interface activation. Using 10% of the reserved factory data as the validation set, voltage prediction errors and temperature prediction errors are verified, with errors required to be below 10mV and 1°C, respectively. The physical model, data-driven model, empirical model, topological relation matrix, and dynamic parameter table are uniformly encapsulated into a structured twin configuration file and uploaded to the cloud; ultimately, a multi-level dynamic model library is obtained.

[0033] The process of using an online identification algorithm to calibrate the parameters of a digital twin model in real time is as follows: The rates of change of real-time current I(t), terminal voltage V(t), and temperature T(t) are monitored, respectively. , , When the conditions are met <ε1, <ε2 and When the current resistance is less than ε3, the current window is marked as a identifiable range. ε1, ε2, and ε3 are the current, temperature, and voltage stability thresholds, respectively, set according to the battery type and operating environment. For example, ε1 = 0.01 A / s, ε2 = 0.001 V / s, and ε3 = 0.01°C / s. Within this stable range, the internal resistance parameters are analyzed in real time using a least-squares identification model. Its basic form is as follows: ,in, For online identification of internal resistance parameters, V model (k, R) represents the voltage prediction value of the physical model at sampling point k, N is the identification window length, t represents the sampling time, k represents the sampling point index, and R is the resistance parameter variable to be solved. For the heat capacity C in the thermal model... th With heat transfer coefficient K th Then the heat balance equation C is adopted. th · =I²R+K th (T env Perform the inverse parameter calculation for -T), where T env The ambient temperature.

[0034] For all the above physical parameters, write the new identification values ​​into the dynamic parameter table T.dyn In the middle, update the confidence interval of the parameters.

[0035] This embodiment employs an incremental learning strategy to adjust the network weights. The process involves constructing a mini-batch dataset B={x} from the most recently collected stable interval samples. i y i}, and use the incremental gradient descent formula W {t+1} =W t −η· Lightweight updates are performed on the model parameters, where W t Let L(B) be the current network weights, L(B) be the loss function applied to batch data B, and η be the learning rate, such as η = 0.001. i For the input sample, y i For the corresponding output labels; The specific process for analyzing the critical states of the battery pack is as follows: Based on the constructed physical model, data-driven model, and empirical model, and after calibration, the models are scheduled through a unified prediction interface to form a parallel output set Y={Y P Y D Y E}, where Y P For the output of the calibrated physical model, Y D Y is the output of the data-driven model after incremental learning. E The output of the empirical model; the consistency among these predictions is calculated by determining the dispersion index of each model output: D = Var(Y P Y D Y E ); Further, confidence intervals based on error distribution and model confidence level are constructed, using the confidence level calculation formula: C i =α·(1 / σ i )+β·L i The confidence level C of output model i i , where σ i L is the standard deviation of the residuals in historical predictions output by the model. i The model level is represented by the following values: physical model (highest level), data-driven model (lowest level), and empirical model (lowest level). For example, the physical model is set to 1.0, the data-driven model to 0.7, and the empirical model to 0.5. α and β are weighting coefficients with values ​​of α=0.6 and β=0.4.

[0036] Finally, the prediction results are weighted and fused based on the confidence scores of each model to obtain the optimal state value Y. final That is: Y final =ΣC i ·Y i / ΣCi And calculate the confidence interval CI, that is: CI=[Y final −k·σ fusion Y final +k·σ fusion ]; where σ fusion The standard deviation of the fused residuals is given, and k is the confidence interval coefficient. For example, a value of 1.96 is provided.

[0037] Based on the output status value and confidence interval, and combined with the current battery operating condition data, the graded repair process is as follows: Based on the confidence interval, the final confidence value C is obtained through multi-model weighted fusion. final That is: C final =ΣC i ·w i Among them, w i For model weight fusion, satisfy Σw i =1.

[0038] The final confidence value C final Compare with the set grading thresholds θ1 and θ2: when the final confidence value C final If the confidence level is ≥θ1, meaning it falls within the high confidence range, then the reliability of the prediction result meets the requirements for automatic adjustment, and the state value Y is directly adopted. final As a control input, it performs automated regulation operations on the heat dissipation system, voltage balancing module, or power dispatching unit. The generated control command includes: the target temperature adjustment amount ΔT. cmd Balanced current setting I bal and power adjustment ratio P adj Each control variable is calculated from the aforementioned predicted value and the current operating state of the battery, for example, ΔT. cmd =K p ·(T pred -T set ), K p This is the proportional control coefficient, used to convert temperature deviation into a control variable; T pred T represents the battery temperature predicted by the model. set The set target temperature value.

[0039] When θ2 ≤ final confidence value C final When the confidence level is less than θ1, the confidence level is in the medium confidence range. Based on the accepted predictive control results, additional buffer strategies such as derating, safety redundancy, and time redundancy are employed to improve the risk tolerance of the control. Specifically, this includes scaling the predictive control variable by a derating factor γ (0 < γ < 1, e.g., γ = 0.7), i.e., Control... output =γ×Y final And enable safety protection boundaries, such as reducing the power limit by ΔP.safe Temperature alarm advance ΔT pre Etc., to reduce the impact of potential misjudgments.

[0040] When the final confidence value C final When the confidence level is less than θ2, the confidence level is in the low confidence range, and the prediction reliability is insufficient to support normal automatic regulation. Therefore, conservative safety mechanisms are triggered, including forcibly entering a safe power mode (such as fixed derating of 50%), limiting the charge and discharge rate, and reducing the thermal management threshold.

[0041] It should be noted that the first threshold and the second threshold are custom settings by those skilled in the art. For example, the first threshold is 0.8 and the second threshold is 0.5.

[0042] In addition, this embodiment further displays a "prediction confidence is too low" prompt message through a human-computer interaction interface, and uploads an abnormal event data packet to the host computer or maintenance terminal. The data packet contains real-time data of the most recent period, model output sequence, confidence evolution curve and recommended handling measures, so as to enable quick human decision-making.

[0043] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A self-adaptive repair method for a CCS panel of an energy storage battery, characterized in that, include: Step 1: Instantiate a dedicated digital twin model for the target battery pack and establish a multi-level dynamic model library corresponding to the digital twin model; Step 2: When the battery is in a stable range with small data fluctuations, the parameters of the digital twin model are calibrated in real time, and the model weights are updated based on the data-driven model. Step 3: Parallel call of the calibrated multiple models, perform consistency analysis on the prediction results of each model, construct confidence analysis based on error distribution, output dispersion and model confidence level, and obtain the optimal state value and its confidence interval after weighted fusion; Step 4: Based on the output status value and confidence interval, perform graded repair in conjunction with the current battery operating condition data.

2. The self-adaptive repair method for a CCS panel of an energy storage battery according to claim 1, characterized in that, The process of building a custom digital twin model is as follows: Based on the identification information of the target battery pack, a unique twin identifier is generated; multi-dimensional characteristic data of the battery pack are collected from the factory and, after preprocessing, cell-level feature vectors are generated. Based on the actual structure of the battery pack, construct a device-level electrical and thermal topology matrix; Based on feature vectors and topological relationships, physical models, data-driven models, and empirical models are instantiated and encapsulated into structured configuration files.

3. The self-adaptive repair method for a CCS panel of an energy storage battery according to claim 2, characterized in that, The physical model is an improved equivalent circuit-thermal coupling model; the data-driven model includes a recurrent neural network subnetwork for time series prediction and an autoencoder subnetwork for anomaly detection; the empirical model library includes a general fault mode library, statistical models of similar products, and simplified diagnostic models based on physical rules.

4. The self-adaptive repair method for a CCS panel of an energy storage battery according to claim 1, characterized in that, The conditions for real-time calibration are: The system monitors the rate of change of current, voltage, and temperature. When the rate of change is lower than a preset threshold, it is marked as an identifiable interval. Within the stable interval, the internal resistance parameter is updated in real time using the least squares identification model. The thermal model parameters are inversely calculated using the thermal balance equation, and the dynamic parameter table is updated.

5. The self-adaptive repair method for a CCS panel of an energy storage battery according to claim 1, characterized in that, The process of weighted fusion for determining the optimal state value and its confidence interval is as follows: Based on the calibrated and updated physical model, data-driven model, and empirical model, the prediction outputs of each model are obtained by parallel invocation; the dispersion between the prediction outputs of each model is calculated. Based on the historical prediction error distribution of each model and the preset model confidence level, the confidence level of each model is calculated. The predicted outputs of each model are weighted and fused based on their confidence levels to obtain the optimal state value; the confidence interval of the optimal state value is calculated based on the fusion residual distribution of the predicted outputs of each model.

6. The self-adaptive repair method for a CCS panel of an energy storage battery according to claim 5, characterized in that, The process of performing graded repair is as follows: Based on the confidence interval of the optimal state value, the final confidence value is calculated; the final confidence value is then compared with multiple preset grading thresholds. Based on the comparison results, different levels of repair control strategies are implemented, including at least one of automated regulation, redundant buffer control, and conservative safety mechanisms.

Images