Large model enabling energy storage battery state estimation interpretable report generation method and system

Abstract

The invention discloses a large model enabling energy storage battery state estimation interpretable report generation method and system, and the method comprises the steps: obtaining experiment charge-discharge cycle data of a battery historical complete life cycle as offline training data, carrying out the preprocessing, and training a state estimation model, voltage section precision and aging process priori knowledge are extracted from historical data through a data-driven priori mining engine; a battery state is obtained in real time through a trained state estimation model, after each charge-discharge cycle is finished, a semantic text description is generated by a battery state semantic conversion module, and then the semantic text description and priori knowledge are input into a large language model; the system is driven to execute a standardized analysis process of internal consistency evaluation, prior conformity analysis and feature logicality test, and finally an interpretable report including a health state decision value, confidence, an analysis basis and operation and maintenance suggestions is output. According to the invention, the transparency and intelligent operation and maintenance level of battery state evaluation are obviously improved.

Description

Technical Field

[0001] This invention relates to the field of battery management system technology, specifically to the field of interpretable joint estimation technology of battery state of charge and health, and particularly to a method and system for generating interpretable reports on energy storage battery state estimation using a large model. Background Technology

[0002] With the widespread application of lithium-ion batteries in electric vehicles, large-scale energy storage, and other fields, the need for battery management systems to accurately and reliably perceive their internal state is becoming increasingly urgent. The state of charge (SOC) and state of health (SOH) of a battery are two key parameters characterizing its operating performance and remaining lifespan, and there is a significant dynamic coupling relationship between them: the decay of SOH changes the battery's usable capacity and internal resistance characteristics, directly affecting the accuracy of SOC estimation; while operating conditions such as the SOC's operating range and charge / discharge rate, in turn, affect the battery's aging rate and the evolution trajectory of SOH. This complex two-way interaction makes it difficult for traditional discrete estimation methods to comprehensively and collaboratively reflect the battery's true state.

[0003] Currently, data-driven methods have become an important research direction for battery state estimation, and some works have attempted to use deep learning models to jointly estimate SOC and SOH. However, these methods still face several challenges: First, the rapid dynamic response of SOC and the long-term gradual trend of SOH differ significantly on the time scale, and most models fail to explicitly separate the characteristics of these two different physical processes, resulting in insufficient feature extraction and significant interference between tasks; second, existing models are mostly "black boxes," with opaque internal decision-making logic, and the estimation results lack a comparable explanation with the battery electrochemical mechanism, resulting in insufficient credibility and acceptability in safety-critical practical applications; in addition, how to effectively integrate the prior knowledge accumulated during battery operation (such as higher estimation accuracy in specific voltage ranges, typical temperature rise and charging characteristics in different health stages, etc.) into the model to improve the reliability of its inference still lacks a systematic fusion mechanism.

[0004] In recent years, large language models have demonstrated powerful capabilities in natural language understanding and generation, and have begun to be used for interpretation and report generation in industrial systems. However, battery time-series data is characterized by its continuous nature, high noise levels, and multivariate coupling, which differs modally from the discrete text symbols that large language models excel at. Directly using it for numerical sequence prediction has limited effectiveness, and the physical consistency and decision credibility of the generated content are not guaranteed.

[0005] Therefore, in view of the problems of incomplete feature separation, weak model interpretability, insufficient fusion of prior knowledge and difficulty of large language models to directly understand complex time series data in the current joint estimation method of battery state, there is an urgent need for a new estimation framework that can collaboratively model the dynamic coupling relationship between SOC and SOH, integrate physical knowledge and have high-order semantic interpretation capabilities, so as to improve the accuracy, reliability and intelligent operation and maintenance level of battery state management. Summary of the Invention

[0006] The purpose of this invention is to address the problems of insufficient specific constraints and lack of interpretable text in existing battery state estimation technologies. It provides a method and system for generating interpretable reports on energy storage battery state estimation using a large model. This invention estimates the battery's state of charge and health status and generates an interpretable natural language report that clarifies the estimation basis, key supporting features, decision confidence levels, and may include maintenance recommendations.

[0007] The objective of this invention is achieved through the following technical solution: The first aspect of this invention provides a method for generating an interpretable report on the state estimation of energy storage batteries using a large model, comprising the following steps: (1) Obtain experimental charge-discharge cycle data of the battery's complete historical life cycle as offline training data, including the battery's current, voltage, temperature, state of charge time series data and the health status of each cycle; (2) Preprocess the offline training data to convert it into a standardized, consistent-length model input format; (3) Construct a state estimation model and train the state estimation model using preprocessed offline training data. With minimizing the total loss function as the optimization objective, adjust the parameters of the state estimation model to obtain a well-trained state estimation model. (4) Prior knowledge mining is performed on the preprocessed offline training data through a data-driven prior mining engine to construct prior knowledge prompt words; (5) When applying online, the online data is preprocessed in the same way as the offline training data, and the real-time state of charge and health are estimated by the trained state estimation model. (6) After each battery charge-discharge cycle, start interpretability enhancement reasoning, and based on the preliminary estimated state of charge and health status and prior knowledge prompts, correct the current health status and generate an interpretable report.

[0008] Furthermore, the preprocessing of the offline training data specifically includes: First, the offline training data is resampled using window sampling. Then, the length of the input time series data is standardized using zero-value padding. A complete charge-discharge cycle data is divided into multiple sliding windows of length L. All input data are then normalized.

[0009] Furthermore, the state estimation model includes a fusion spatiotemporal encoder, a dual-branch convolutional module, a charged state decoder, and a healthy state decoder. The fusion spatiotemporal encoder includes an embedding layer and a Transformer encoder layer. The dual-branch convolutional module includes a charged feature branch and a healthy feature branch. The charged feature branch includes a convolutional layer and a pooling layer. The healthy feature branch includes a convolutional layer and a pooling layer. The preprocessed current, voltage, and temperature time-series data are fused and encoded by a fusion spatiotemporal encoder to obtain battery fusion features. The battery fusion features are then processed by the charge feature branch and health feature branch of a dual-branch convolution module to obtain charge features and health features, respectively. The charge features are processed by a charge state decoder to obtain an estimate of the charge state, and the health features are processed by a health state decoder to obtain an estimate of the health state.

[0010] Furthermore, training the state estimation model using preprocessed offline training data specifically includes: Based on the preprocessed offline training data, the preprocessed current, voltage, and temperature time-series data are input into the state estimation model to obtain estimates of the state of charge (SOC) and health state. The supervision loss, inter-cycle loss, and consistency loss are calculated based on the estimated SOC and health state and their corresponding preprocessed SOCs. The sum of these losses is used as the total loss function of the state estimation model. With minimizing the total loss function as the optimization objective, gradient descent is used to adjust the parameters of the state estimation model until a preset number of training epochs are reached to obtain a well-trained state estimation model.

[0011] Furthermore, the formula for calculating the supervision loss is as follows:

[0012] In the formula, Indicates monitoring losses, Loss of monitoring indicates health status. This represents the monitoring loss of the state of charge. and Let these represent the true and estimated values ​​of the state of charge of the i-th sample, respectively. and Let represent the true value and the estimated value of the health status of the i-th sample, respectively, and N represent the total number of samples; The formula for calculating the inter-cycle loss is:

[0013] In the formula, Indicates the inter-cycle loss. Intercycle loss representing health status Represents the intercycle loss in the state of charge. and Let represent the estimated health state and state of charge of all samples in the k-th charge-discharge cycle, respectively, where K represents the total number of battery charge-discharge cycles. Represents the calculation of variance. and The data sets represent charging and discharging respectively. Represents the ReLU activation function. This represents the difference between the current time and the previous time. The formula for calculating the consistency loss is as follows:

[0014] In the formula, Indicates consistency loss. Represents logarithmic calculation; The formula for calculating the total loss function of the state estimation model is as follows:

[0015] In the formula, This represents the total loss function of the state estimation model. and These represent the weight parameters for inter-cycle loss and consistency loss, respectively.

[0016] Furthermore, step (4) specifically includes the following sub-steps: (4.1) Voltage-Precision Prior Knowledge Mining: Suppose that the preprocessed offline training data contains N samples, calculate the average voltage and absolute error of health status estimation for each sample; divide the voltage range into M voltage intervals equally, and calculate the average absolute error of each voltage interval based on the absolute error of health status estimation; calculate the global average error based on the absolute error of health status estimation for all samples, and calculate the corresponding precision factor for each voltage interval based on its average absolute error and the global average error; select voltage intervals with precision factors greater than 1 to form the prior knowledge set of estimation precision. (4.2) Knowledge mining of battery aging characteristics: The preprocessed offline training data is divided into three health state intervals: healthy, medium and aged. The constant current charging time, charging temperature difference and charging state of charge change rate of each health state interval are calculated as the aging prior knowledge set. (4.3) Construction of prior knowledge cue words: The prior knowledge set of estimation accuracy and the prior knowledge set of aging are combined into meta-information, and together with the text content of battery data description and model feature description, they form prior knowledge cue words.

[0017] Furthermore, step (6) specifically includes the following sub-steps: (6.1) Through the battery state semantic conversion module, the original time series data of the charge-discharge cycle, the estimated values ​​of the state of charge and health, as well as the charge characteristics and health characteristics, are jointly converted into a structured semantic text description. (6.2) Combine prior knowledge prompts, structured semantic text descriptions, and pre-set professional guidance prompt templates to construct complete prompts for large language models; (6.3) Input the complete prompt words into the pre-trained large language model and drive the large language model to perform multi-step reasoning analysis including internal consistency evaluation, prior conformity analysis and feature logic test; (6.4) Based on the results of multi-step reasoning analysis, the large language model outputs a formatted and interpretable report containing the final health status decision value of the current charge-discharge cycle, decision confidence, reasoning basis analysis, and targeted operation and maintenance suggestions.

[0018] Furthermore, the battery state semantic conversion module specifically includes: The original time-series data, estimated values ​​of state of charge and health, and charge and health features are converted into structured semantic text descriptions, including temperature statistics, charging temperature difference, health status estimation statistics for voltage range, charging state of charge change rate, constant current charging time, charge feature variance, and health feature variance.

[0019] Furthermore, in the formatted interpretable report, the final health status decision value is a corrected or confirmed value of the health status after a comprehensive evaluation based on consistency assessment, prior conformity analysis, and characteristic logic test of the preliminary estimates of multiple windows within the current charge-discharge cycle.

[0020] A second aspect of this invention provides a large-model-enabled interpretable report generation system for energy storage battery state estimation, comprising one or more processors and a memory, wherein the memory is coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement the above-described large-model-enabled interpretable report generation method for energy storage battery state estimation.

[0021] The beneficial effects of this invention are as follows: This invention, by designing a dual-branch convolutional network that separates charge characteristics and health characteristics, explicitly models different physical processes within the battery. Combined with physical consistency constraints for joint training, it significantly improves the joint estimation accuracy of SOC and SOH, enhancing the model's ability to identify battery dynamics and degradation characteristics. Furthermore, by introducing a large language model as an interpretable inference engine and designing prompt templates containing prior knowledge and professional guidance, this invention can generate natural language reports with logical chains that reflect the physical laws of the battery, achieving a leap from "black box numerical values" to "white box interpretation" of state estimation results. This invention significantly improves the transparency and credibility of the battery status assessment decision-making process, as well as the level of intelligent operation and maintenance. Through a data-driven prior knowledge mining and injection mechanism, it integrates the voltage segment accuracy patterns and aging stage characteristics hidden in the training data into the prompting system, guiding the large language model to perform reasoning consistent with domain knowledge. This effectively reduces the illusion and bias of the generated content and enhances the reliability of the conclusions. Compared with existing technologies, the analysis report generated by this invention, which integrates numerical results and semantic interpretations, can be directly used for battery health diagnosis, operation and maintenance strategy formulation, and system performance evaluation, improving the intelligent operation and maintenance level and decision-making efficiency of the battery management system. Attached Figure Description

[0022] Figure 1 A flowchart illustrating the method for generating interpretable reports for energy storage battery state estimation, which enables the large-scale model of this invention.

[0023] Figure 2 Example diagram of battery state of charge estimation verified by this invention;

[0024] Figure 3 Example diagram of battery health status estimation to verify the present invention;

[0025] Figure 4 This is a screenshot of an interpretable report under a healthy cycle, used to verify the present invention; wherein, Figure 4 (a) in the image is a screenshot of the summary section of the internal consistency assessment under the healthy cycle; Figure 4 (b) in the figure is a screenshot of the summary section of the prior conformity analysis under the healthy cycle; Figure 4 (c) in the image is a screenshot of the summary section of the characteristic logic test under the healthy cycle. Figure 4 (d) in the image is a screenshot of the report summary section under the health cycle;

[0026] Figure 5 This is a screenshot of an interpretable report under aging cycles used to verify the present invention; wherein, Figure 5 (a) in the image is a screenshot of the summary section of the internal consistency assessment under the aging cycle; Figure 5 (b) in the figure is a screenshot of the summary section of the prior conformity analysis under aging cycle; Figure 5(c) in the image is a screenshot of the summary section of the characteristic logic test under the aging cycle; Figure 5 (d) in the image is a screenshot of the report summary section under the aging cycle;

[0027] Figure 6 A schematic diagram of a structure for generating an interpretable report on the state estimation of energy storage batteries, which enables the large-scale model of this invention. Detailed Implementation

[0028] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the invention as detailed in the appended claims.

[0029] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The singular forms “a,” “the,” and “the” used in this invention and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0030] It should be understood that although the terms first, second, third, etc., may be used in this invention to describe various information, this information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first information may also be referred to as second information without departing from the scope of this invention, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0031] The present invention will now be described in detail with reference to the accompanying drawings. Unless otherwise specified, the features of the following embodiments and implementations can be combined with each other.

[0032] See Figure 1 The present invention provides a method for generating interpretable reports on energy storage battery state estimation using a Large Language Model (LLM), which specifically includes the following steps:

[0033] (1) Obtain experimental charge-discharge cycle data of the battery's complete historical life cycle as offline training data, including the battery's current, voltage, temperature, state of charge time series data and the health status of each cycle.

[0034] In this embodiment, the nominal capacity of the batteries in the offline training data is 3.4Ah. All batteries were subjected to cyclic charging and discharging at an operating temperature of 23°C, using a constant current followed by a constant voltage charging and discharging mode. The charging cut-off voltage and discharging cut-off voltage were 4.2V and 2.7V, respectively, and the data sampling interval was 20s.

[0035] (2) Preprocess the offline training data to convert it into a standardized, consistent-length model input format.

[0036] Specifically, the offline training data is first resampled using windowed sampling, and then the length of the input time series data is unified using zero-value padding. A complete charge-discharge cycle data is divided into multiple sliding windows of length L. All input data are then normalized, and the normalization methods include max-min normalization, Z-score standardization, etc., which can be selected according to actual needs.

[0037] In this embodiment, the window length L can be set to 80.

[0038] (3) Construct a state estimation model and use the preprocessed offline training data to train the state estimation model end-to-end. With minimizing the total loss function of the state estimation model as the optimization objective, adjust the parameters of the state estimation model to obtain a well-trained state estimation model.

[0039] In this embodiment, the state estimation model includes a fusion spatiotemporal encoder, a dual-branch convolutional module, a state of charge decoder, and a health state decoder. The fusion spatiotemporal encoder includes an embedding layer and a Transformer encoder layer. The dual-branch convolutional module includes a charge feature branch and a health feature branch. The charge feature branch includes convolutional layers and pooling layers, and the health feature branch also includes convolutional layers and pooling layers. Preprocessed current, voltage, and temperature time-series data are fused and encoded using the fusion spatiotemporal encoder to obtain battery fusion features. These battery fusion features are then processed by the charge feature branch and the health feature branch of the dual-branch convolutional module to obtain charge features and health features, respectively. The charge features are then processed by the state of charge decoder to obtain an estimate of the state of charge. The health features are then processed by the health state decoder to obtain an estimate of the health state.

[0040] In this embodiment, the specific network structure of the state estimation model is shown in Table 1, where N represents the number of samples input to the estimation model.

[0041] Table 1: Network Structure of the Estimated Model

[0042] In this embodiment, the state estimation model is trained end-to-end using preprocessed offline training data. Specifically, this includes: inputting preprocessed current, voltage, and temperature time-series data into the state estimation model based on the preprocessed offline training data to obtain estimates of the state of charge and health states; calculating the supervision loss, inter-cycle loss, and consistency loss based on the estimates of the state of charge and health states and their corresponding preprocessed states of charge and health states; and using the gradient descent method to adjust the parameters of the state estimation model with the goal of minimizing the total loss function of the state estimation model, until a preset number of training rounds are reached to obtain a well-trained state estimation model.

[0043] Furthermore, the formula for calculating the monitoring loss is:

[0044] In the formula, Indicates monitoring losses, Loss of monitoring indicates health status. This represents the monitoring loss of the state of charge. and Let these represent the true and estimated values ​​of the state of charge of the i-th sample, respectively. and Let represent the true value and the estimated value of the health status of the i-th sample, respectively, and N represent the total number of samples.

[0045] Furthermore, the formula for calculating inter-cycle loss is:

[0046] In the formula, Indicates the inter-cycle loss. Intercycle loss representing health status Represents the intercycle loss in the state of charge. and Let represent the estimated health state and state of charge of all samples in the k-th charge-discharge cycle, respectively, where K represents the total number of battery charge-discharge cycles. Represents the calculation of variance. and The data sets represent charging and discharging respectively. Represents the ReLU activation function. It represents the difference between the current time and the previous time.

[0047] Furthermore, the formula for calculating the consistency loss is as follows:

[0048] In the formula, Indicates consistency loss. This indicates logarithmic calculation.

[0049] Furthermore, the formula for calculating the total loss function of the state estimation model is as follows:

[0050] In the formula, This represents the total loss function of the state estimation model. and These represent the weight parameters for inter-cycle loss and consistency loss, respectively.

[0051] In this embodiment, a training batch consists of all window data from one charge-discharge cycle. Each training iteration inputs 64 batches, i.e., all window data from 64 charge-discharge cycles, and the learning rate is set to 1e. -3 The training rounds are set to 200 rounds. and The value is 0.1.

[0052] (4) Prior knowledge mining is performed on the preprocessed offline training data through a data-driven prior mining engine to construct prior knowledge prompt words.

[0053] (4.1) Voltage-precision prior knowledge mining: Suppose that the preprocessed offline training data contains N samples, calculate the average voltage of each sample. Absolute error in health status estimation The absolute error of health status estimation The calculation formula is:

[0054] In the formula, This represents the absolute error in estimating the health status of the nth sample. and Let these represent the true and estimated health status values ​​of the nth sample, respectively. The voltage range is averaged and discretized into M voltage intervals: For each voltage range Absolute error based on health status estimation Calculate its mean absolute error:

[0055] In the formula, Indicates voltage range The mean absolute error, This represents the number of samples in the m-th voltage interval. This represents the lower voltage limit of the m-th voltage range. This represents the upper voltage limit of the m-th voltage range. The absolute error is estimated based on the health status of all samples. Calculate the global average error , for each voltage range Mean absolute error Compared with global average error Compare and calculate the accuracy factor corresponding to the voltage range. :

[0056] In the formula, To avoid dividing by zero for small constants. This indicates that the estimation accuracy for this voltage range is above average, therefore it was selected. The voltage range is used as a high-precision voltage range to form the prior knowledge set for estimation accuracy. :

[0057] (4.2) Knowledge mining of battery aging characteristics: The preprocessed offline training data is divided into 3 health status intervals according to the health status: healthy (SOH>0.9) Medium (0.9 ≥ SOH > 0.8) Aging (SOH≤0.8) Calculate the constant current charging time for each health state interval. Charging temperature difference Rate of change of state of charge during charging As a set of prior knowledge about aging :

[0058] In the formula, H represents the three health status ranges: healthy, moderate, and aged. , and These represent the constant current charging time (in seconds), charging temperature difference (in °C), and rate of change of state of charge during the healthy state interval H, respectively.

[0059] (4.3) Construction of prior knowledge cue words: The prior knowledge set for estimating accuracy is constructed. and the set of prior knowledge about aging Merge into metadata Together with textual content such as battery data descriptions and model feature descriptions (e.g., describing battery materials, voltage range, charging protocols, and other conventional information), these form prior knowledge prompts. .

[0060] In this embodiment, prior knowledge prompts as follows:

[0061] Prior knowledge:

[0062] [battery_info]

[0063] Data and Model: The input data consists of time-series data of voltage, current, and temperature, with a sampling interval of 20 seconds, divided into fixed-length (L) windows. Each window outputs a SOC estimate, a SOH estimate, and two health characteristics reflecting short-term dynamics (related to SOC) and long-term state (related to SOH), calculated by a time-series model. These two characteristics are derived from convolutional networks of different scales. Therefore, the health characteristics within the same charge-discharge cycle should be more similar; significant changes in the health characteristics may indicate changes in the true SOH or measurement anomalies.

[0064] Based on empirical statistics of offline training datasets, the following priors were found: .

[0065] [battery_info] describes the battery data, specifically: "Battery data comes from an 18650 cylindrical lithium battery, with LiCoO2 and LiNi as the cathode materials." 0.4 Co 0.4 Mn 0.2 O2, graphite as the negative electrode material. The nominal capacity of the battery is 3.4Ah. All batteries are cycle-charged and discharged at an operating temperature of 23 degrees Celsius.

[0066] (5) When applying online, the online data is preprocessed in the same way as the offline training data, and the pre-trained state estimation model is used to make a preliminary estimate of the real-time state of charge and health.

[0067] (6) After each battery charge-discharge cycle, start interpretability enhancement reasoning, and based on the preliminary estimated state of charge and health status and prior knowledge prompts, correct the current health status and generate an interpretable report.

[0068] (6.1) Through the battery state semantic conversion module, the original time-series data of the charge-discharge cycle, the estimated values ​​of the state of charge and health generated by the state estimation model, and the charge features and health features generated within the state estimation model are jointly converted into a structured semantic text description. .

[0069] Specifically, the battery state semantic conversion module includes converting the original time-series data, the estimated values ​​of state of charge and health generated by the state estimation model, and the charge and health features generated within the state estimation model into structured semantic text descriptions, including temperature statistics, charging temperature difference, voltage range health state estimation statistics, charging state of charge change rate, constant current charging time, charge feature variance, and health feature variance. .

[0070] Table 2: Battery State Semantic Transition Table

[0071] In Table 2, The highest temperature, The lowest temperature, The average temperature. T represents the temperature difference during the charging phase, where T is the temperature time-series data. The length of the timing data for charging. The temperature data at the c-th moment of charging. and Voltage range The mean and variance, and Let and represent the charge characteristic and the mean of the charge characteristic of the j-th sample, respectively. and Let represent the health characteristics and the mean of the health characteristics of the j-th sample, respectively, and J represent the total number of samples.

[0072] (6.2) Prior knowledge prompts Structured semantic text description It combines with preset professional guidance prompt templates to construct complete prompt words for large language models.

[0073] (6.3) Input the complete prompt words into the pre-trained large language model and drive the large language model to perform multi-step reasoning analysis, including internal consistency assessment, prior conformity analysis and feature logic test.

[0074] The internal consistency assessment specifically involves: using variance analysis based on health characteristics to determine the dispersion of health state estimates within charge-discharge cycles, and assessing the reliability of the preliminary estimation results. The prior conformity analysis specifically involves: comparing the estimation results for different voltage ranges with the mined prior accuracy, and analyzing whether the estimated values ​​are within the model's empirical error range. The feature logic verification specifically involves: verifying the degree of matching between the dynamics of charge characteristics and the rate of change of state of charge, as well as the physical rationality of the health characteristic levels and health state estimates.

[0075] Furthermore, large language models include, but are not limited to, Qwen, DeepSeek, and GPT.

[0076] (6.4) Based on the results of multi-step reasoning analysis, the large language model outputs a formatted and interpretable report containing the final health status decision value of the current charge-discharge cycle, decision confidence, reasoning basis analysis, and targeted operation and maintenance suggestions.

[0077] In the formatted interpretable report, the final health status decision value is a corrected or confirmed value of the health status after a comprehensive evaluation based on consistency assessment, prior conformity analysis, and characteristic logic test of the preliminary estimates of multiple windows within the current charge-discharge cycle.

[0078] In this embodiment, the specific conversion characteristics of the battery state semantic conversion module are shown in Table 2. Qwen3-Plus is used as the reasoning language model, and the complete prompt words are as follows:

[0079] "You are a professional AI for battery system state estimation and health diagnosis."

[0080]

[0081] Based strictly on the objective data and characteristics provided below, please perform step-by-step reasoning and comprehensive evaluation of the current battery cycle health status (SOH) and generate a structured technical analysis report.

[0082] Phase 1: Model Inference Data Confirmation

[0083]

[0084] Phase Two: Step-by-Step Reasoning Process

[0085] Please consider the following questions in order:

[0086] 1. Internal Consistency Assessment: Is the variance of the health characteristics in this cycle at a low level? This directly reflects the inherent consistency of the model's SOH estimation for this cycle. Observe the distribution of the SOH estimates. Are there any obvious outliers? Are these outliers associated with specific voltage ranges or temperature intervals? In summary, is the overall SOH window estimation result for this cycle reliable? Are there any estimation discrepancies caused by data quality, noise, or drastic changes in operating conditions?

[0087] 2. Prior Consistency Analysis: Compare the average SOH (low, medium, high) across different voltage ranges. Are the differences within the error range exhibited by the model in the training set? The voltage range is typically the most accurate. Do the SOH estimates for this cycle exhibit higher concentration (smaller dispersion) near this voltage range?

[0088] 3. Characteristic Logicality Test: Does the variance of the charge characteristic match the rate of change of SOC? This can indirectly verify whether the dynamic process data is normal. Considering the overall health characteristic level and the estimated SOH, is it reasonable from a physical perspective?

[0089] 4. Aging status test: Compare the data of different aging levels in the prior knowledge to determine whether an unhealthy state has occurred in this cycle.

[0090] Phase Three: Final Diagnosis and Report Generation

[0091] Based on the above reasoning, please output the following:

[0092] 1. Final SOH decision value: Provides a clear SOH percentage value based on the most reliable analysis at present.

[0093] 2. Decision confidence: Evaluate the degree of certainty of the decision using "high / medium / low".

[0094] 3. Key technical recommendations: Based on the current situation, propose 1-2 specific monitoring or maintenance recommendations.

[0095] Please ensure your report is professional and objective, and that all conclusions are supported by data. Now, begin your analysis:".

[0096] For example, four batteries from open-source data were selected for experimental verification (three batteries as offline training data and one as test data) to illustrate the steps and verify the effectiveness of the method described in this invention. Furthermore, to more clearly demonstrate the superiority of the method described in this invention in battery state estimation, five neural network methods—Deep Neural Networks (DNN), Long Short-Term Memory (LSTM), Convolutional Neural Networks (CNN), Temporal Convolutional Networks (TCN), and Transformer—were compared with the method described in this invention. The experimental results are shown in Table 3. The evaluation criteria used were Mean Square Error (RMSE) and Mean Absolute Error (MAE). The decoder structure of the comparison methods was the same as that of the method described in this invention, and other training parameters were also the same.

[0097] Table 3: Comparison of quantification results from different methods

[0098] As shown in Table 3, the method described in this invention is effective when compared with the other five methods. Figure 2 The state-of-charge estimation for the last two cycles is shown. Figure 3 This presents an estimate of the health status (calculated by averaging the SOH values ​​estimated over multiple windows in each cycle), provided by... Figure 2 and Figure 3 As can be seen, compared with methods without relevant constraints and multiple scales, the method described in this invention achieves optimal results on all indices of the two state variables.

[0099] The interpretable report generation section uses the 20th cycle when the test battery is relatively healthy and the 153rd cycle when abnormal capacity decay occurs as interpretable reasoning examples to verify the effectiveness of the interpretability enhancement reasoning section.

[0100] Screenshot of the generated interpretable report, as shown below Figure 4 and Figure 5 As shown, Figure 4 (a) in the image is a screenshot of the summary section of the internal consistency assessment under a healthy cycle. Figure 4 (b) in the image is a screenshot of the summary section of the prior conformity analysis under a healthy cycle. Figure 4 (c) in the image is a screenshot summarizing the characteristic logicality test under the healthy cycle. Figure 4 (d) in the image is a screenshot of the report summary section under the healthy cycle. Figure 5 (a) in the image is a screenshot of the summary section of the internal consistency assessment under the aging cycle. Figure 5 (b) in the figure is a screenshot of the summary section of the prior conformity analysis under aging cycles. Figure 5 (c) in the image is a screenshot summarizing the characteristic logic test under the aging cycle. Figure 5 (d) in the image is a screenshot of the report summary section under the aging cycle. Figure 4 and Figure 5 As shown in the partial screenshot of the interpretable report, the summary section of the interpretable report includes the final estimated health status value and model confidence (reliability). In the 20th cycle, the true health status is 97.7%, and the inferred health status is 94.5%, a 0.7% decrease in MAE compared to the average of 93.8%. This demonstrates the effectiveness of injecting prior knowledge in improving the accuracy of the final health status estimation. In the 153rd cycle, the true health status is 79.59%, and the inferred health status is 82.5%, a 3.28% decrease in MAE compared to the average of 85.78%. Furthermore, through prior knowledge of aging, the report infers that the current model may have a mismatch, indicating abnormal battery aging, and provides further suggested measures, further demonstrating the effectiveness and interpretability of the method described in this invention.

[0101] The large language models used in this invention include, but are not limited to, GLM4, Qwen3, Deepseek-R1, and Deepseek-V3. It should be understood that the performance of large language models directly affects the performance of downstream equipment maintenance tasks; therefore, the best-performing large language model should be used whenever the equipment's computing power allows.

[0102] Corresponding to the aforementioned embodiments of the large model-enabled method for generating interpretable reports on energy storage battery state estimation, the present invention also provides embodiments of a large model-enabled system for generating interpretable reports on energy storage battery state estimation.

[0103] See Figure 6 The present invention provides a large model-enabled energy storage battery state estimation interpretable report generation system, including one or more processors and a memory, wherein the memory is coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement the large model-enabled energy storage battery state estimation interpretable report generation method in the above embodiment.

[0104] The embodiments of the large-model-enabled interpretable report generation system for energy storage battery state estimation of this invention can be applied to any device with data processing capabilities, such as a computer or other similar device or system. The system embodiments can be implemented through software, hardware, or a combination of both. Taking software implementation as an example, as a logical system, it is formed by the processor of any data-processing device loading the corresponding computer program instructions from non-volatile memory into memory for execution. From a hardware perspective, such as... Figure 6 The diagram shown illustrates a hardware structure of any data-processing-capable device within the large-model-enabled interpretable report generation system for energy storage battery state estimation, as described in this invention. (Except for...) Figure 6 In addition to the processor, memory, network interface, and non-volatile memory shown, any data processing device in which the system is located in the embodiment may also include other hardware depending on the actual function of the data processing device, which will not be described in detail here.

[0105] The implementation process of the functions and roles of each unit in the above system is detailed in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0106] For the system embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the present invention according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0107] This invention also provides a computer-readable storage medium storing a program that, when executed by a processor, implements the large-model-enabled energy storage battery state estimation interpretable report generation method described in the above embodiments.

[0108] The computer-readable storage medium can be an internal storage unit of any data processing device described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be any data processing device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units of any data processing device and external storage devices. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0109] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for generating interpretable reports of energy storage battery state estimation using a large model, characterized in that, Includes the following steps: (1) Obtain experimental charge-discharge cycle data of the battery's complete historical life cycle as offline training data, including the battery's current, voltage, temperature, state of charge time series data and the health status of each cycle; (2) Preprocess the offline training data to convert it into a standardized, consistent-length model input format; (3) Construct a state estimation model and train the state estimation model using preprocessed offline training data. With minimizing the total loss function as the optimization objective, adjust the parameters of the state estimation model to obtain a well-trained state estimation model. (4) Prior knowledge mining is performed on the preprocessed offline training data through a data-driven prior mining engine to construct prior knowledge prompt words; (5) When applying online, the online data is preprocessed in the same way as the offline training data, and the real-time state of charge and health are estimated by the trained state estimation model. (6) After each battery charge-discharge cycle, start interpretability enhancement reasoning, and based on the preliminary estimated state of charge and health status and prior knowledge prompts, correct the current health status and generate an interpretable report.

2. The method for generating interpretable reports for energy storage battery state estimation based on a large model as described in claim 1, characterized in that, The preprocessing of offline training data specifically includes: First, the offline training data is resampled using window sampling. Then, the length of the input time series data is standardized using zero-value padding. A complete charge-discharge cycle data is divided into multiple sliding windows of length L. All input data are then normalized.

3. The method for generating interpretable reports for energy storage battery state estimation based on a large model as described in claim 1, characterized in that, The state estimation model includes a fusion spatiotemporal encoder, a dual-branch convolutional module, a charged state decoder, and a healthy state decoder. The fusion spatiotemporal encoder includes an embedding layer and a Transformer encoder layer. The dual-branch convolutional module includes a charged feature branch and a healthy feature branch. The charged feature branch includes a convolutional layer and a pooling layer. The healthy feature branch includes a convolutional layer and a pooling layer. The preprocessed current, voltage, and temperature time-series data are fused and encoded by a fusion spatiotemporal encoder to obtain battery fusion features. The battery fusion features are then processed by the charge feature branch and health feature branch of a dual-branch convolution module to obtain charge features and health features, respectively. The charge features are processed by a charge state decoder to obtain an estimate of the charge state, and the health features are processed by a health state decoder to obtain an estimate of the health state.

4. The method for generating interpretable reports for energy storage battery state estimation based on a large model as described in claim 1, characterized in that, The step of training the state estimation model using preprocessed offline training data specifically includes: Based on the preprocessed offline training data, the preprocessed current, voltage, and temperature time-series data are input into the state estimation model to obtain estimates of the state of charge (SOC) and health state. The supervision loss, inter-cycle loss, and consistency loss are calculated based on the estimated SOC and health state and their corresponding preprocessed SOCs. The sum of these losses is used as the total loss function of the state estimation model. With minimizing the total loss function as the optimization objective, gradient descent is used to adjust the parameters of the state estimation model until a preset number of training epochs are reached to obtain a well-trained state estimation model.

5. The method for generating interpretable reports for energy storage battery state estimation based on a large model according to claim 4, characterized in that, The formula for calculating the supervision loss is: In the formula, Indicates monitoring losses, Loss of monitoring indicates health status. This represents the monitoring loss of the state of charge. and Let these represent the true and estimated values ​​of the state of charge of the i-th sample, respectively. and Let represent the true value and the estimated value of the health status of the i-th sample, respectively, and N represent the total number of samples; The formula for calculating the inter-cycle loss is: In the formula, Indicates the inter-cycle loss. Intercycle loss representing health status Represents the intercycle loss in the state of charge. and Let represent the estimated health state and state of charge of all samples in the k-th charge-discharge cycle, respectively, where K represents the total number of battery charge-discharge cycles. Represents the calculation of variance. and The data sets represent charging and discharging respectively. Represents the ReLU activation function. This represents the difference between the current time and the previous time. The formula for calculating the consistency loss is as follows: In the formula, Indicates consistency loss. Represents logarithmic calculation; The formula for calculating the total loss function of the state estimation model is as follows: In the formula, This represents the total loss function of the state estimation model. and These represent the weight parameters for inter-cycle loss and consistency loss, respectively.

6. The method for generating interpretable reports for energy storage battery state estimation based on a large model according to claim 1, characterized in that, Step (4) specifically includes the following sub-steps: (4.1) Voltage-Precision Prior Knowledge Mining: Suppose that the preprocessed offline training data contains N samples, calculate the average voltage and absolute error of health status estimation for each sample; divide the voltage range into M voltage intervals equally, and calculate the average absolute error of each voltage interval based on the absolute error of health status estimation; calculate the global average error based on the absolute error of health status estimation for all samples, and calculate the corresponding precision factor for each voltage interval based on its average absolute error and the global average error; select voltage intervals with precision factors greater than 1 to form the prior knowledge set of estimation precision. (4.2) Knowledge mining of battery aging characteristics: The preprocessed offline training data is divided into three health state intervals: healthy, medium and aged. The constant current charging time, charging temperature difference and charging state of charge change rate of each health state interval are calculated as the aging prior knowledge set. (4.3) Construction of prior knowledge cue words: The prior knowledge set of estimation accuracy and the prior knowledge set of aging are combined into meta-information, and together with the text content of battery data description and model feature description, they form prior knowledge cue words.

7. The method for generating interpretable reports for energy storage battery state estimation based on a large model according to claim 1, characterized in that, Step (6) specifically includes the following sub-steps: (6.1) Through the battery state semantic conversion module, the original time series data of the charge-discharge cycle, the estimated values ​​of the state of charge and health, as well as the charge characteristics and health characteristics, are jointly converted into a structured semantic text description. (6.2) Combine prior knowledge prompts, structured semantic text descriptions, and pre-set professional guidance prompt templates to construct complete prompts for large language models; (6.3) Input the complete prompt words into the pre-trained large language model and drive the large language model to perform multi-step reasoning analysis including internal consistency evaluation, prior conformity analysis and feature logic test; (6.4) Based on the results of multi-step reasoning analysis, the large language model outputs a formatted and interpretable report containing the final health status decision value of the current charge-discharge cycle, decision confidence, reasoning basis analysis, and targeted operation and maintenance suggestions.

8. The method for generating interpretable reports for energy storage battery state estimation based on a large model according to claim 7, characterized in that, The battery state semantic conversion module specifically includes: The original time-series data, estimated values ​​of state of charge and health, and charge and health features are converted into structured semantic text descriptions, including temperature statistics, charging temperature difference, health status estimation statistics for voltage range, charging state of charge change rate, constant current charging time, charge feature variance, and health feature variance.

9. The method for generating interpretable reports for energy storage battery state estimation based on a large model according to claim 7, characterized in that, In the formatted interpretable report, the final health status decision value is a corrected or confirmed value of the health status after a comprehensive evaluation based on consistency assessment, prior conformity analysis, and characteristic logic test of the preliminary estimates of multiple windows within the current charge-discharge cycle.

10. A large-model-enabled interpretable report generation system for energy storage battery state estimation, comprising one or more processors and a memory, characterized in that, The memory is coupled to the processor; wherein the memory is used to store program data, and the processor is used to execute the program data to implement the large model-enabled interpretable report generation method for energy storage battery state estimation as described in any one of claims 1-9.

Images