Large model intelligent operation and maintenance method and system for energy storage cluster time sequence understanding

Abstract

The application discloses a kind of big model intelligent operation and maintenance methods and systems for energy storage cluster time sequence understanding, the method comprises: obtaining energy storage battery time sequence data, extract single body and cluster multi-scale operation representation;Introduce dense time slice mechanism to the semantic mode division of continuous time sequence;Define multi-granularity battery semantic table and standardization mapping rule, convert numerical signal to semantic representation;Adopt multi-agent consensus verification mechanism to guide large language model, generate structured operation archives, support explainable operation and maintenance based on prompt query and retrieval enhancement.The application first defines a new paradigm of battery time sequence understanding, builds the mapping relationship from battery time sequence data to advanced semantic knowledge, bridges the gap between the two, solves the problem of battery knowledge deficiency and numerical time sequence signal processing existing in large language model, significantly improves the credibility and accuracy of time sequence prediction and anomaly detection tasks, provides knowledge-driven automated operation and maintenance support for energy storage systems, and has economic and engineering practicality.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance and time series analysis technology for energy storage battery systems, and in particular to a large-scale intelligent operation and maintenance method and system for time series understanding of energy storage clusters. Background Technology

[0002] With the rapid development of lithium-ion batteries and battery energy storage systems, their applications are becoming increasingly important, requiring them to have high reliability and high efficiency in operation and maintenance. As the scale and complexity of battery systems increase, especially with the promotion of large-scale battery energy storage systems composed of multiple cells connected in series and parallel, their runtime sequence data exhibits characteristics such as multivariability and strong coupling. Traditional analysis methods based on threshold rules or shallow machine learning are unable to fully explore the deep state semantics and anomaly patterns hidden within them, and have significant limitations in terms of real-time performance, interpretability, and cross-operating condition generalization ability.

[0003] In recent years, artificial intelligence technology has been increasingly integrated into battery management systems, with some studies attempting to introduce deep learning models to achieve end-to-end prediction of battery status. However, most of these methods remain in a "black box" operation mode, only outputting numerical results (such as state estimates or binary anomaly markers), lacking semantic descriptions of the decision-making basis and evolution process. This makes it difficult for maintenance personnel to understand the reasoning logic, locate root causes, and formulate maintenance strategies. Meanwhile, Large Language Models (LLMs) have demonstrated powerful sequence modeling and contextual reasoning capabilities in the field of natural language processing and are beginning to be explored for knowledge-based question answering and report generation tasks in the industrial and energy sectors. However, due to the complex characteristics of battery time-series data, such as continuous values, high noise, and cross-variable correlations, there are modal differences between LLMs and the discrete text symbols that LLMs are adapted to. Directly using LLMs to process numerical sequences has limited effectiveness, making it difficult for LLMs to accurately capture key time-series patterns such as charging and discharging dynamics, cell inconsistencies, and pre-fault characteristics.

[0004] Furthermore, most existing battery data analysis methods based on large language models are limited to single cells or simple scenarios, failing to systematically address the needs for semantic representation of cluster states and analysis of interaction effects under multi-cell coupled topologies in energy storage systems. At the same time, the interpretability, reliability, and conformity of the generated results to domain knowledge logic still lack effective verification and assurance mechanisms, making it difficult to support high-risk decisions in actual operation and maintenance.

[0005] Therefore, in order to address the bottlenecks of the aforementioned large language models, such as difficulty in directly understanding battery time-series data, lack of domain knowledge, inability to perceive time-series dynamics, and insufficient interpretability in complex energy storage system operation and maintenance scenarios, there is an urgent need for a battery-specific narrative framework that can integrate time-series feature engineering and semantic generation capabilities to achieve reliable and automated conversion from numerical sequences to operation and maintenance knowledge, thereby improving the transparency and intelligence level of battery management. Summary of the Invention

[0006] The purpose of this invention is to address the bottlenecks in existing technologies, such as the lack of interpretability of operation and maintenance results, the difficulty of large language models directly understanding battery time-series data, the lack of domain knowledge, and the inability to perceive time-series dynamics. This invention provides a large-scale intelligent operation and maintenance method and system for understanding the time-series data of energy storage clusters. By constructing a time-series-semantic mapping, this invention solves the problem that large language models cannot directly understand battery time-series data, automatically transforming the runtime time-series data of individual energy storage batteries and the cluster into structured, interpretable semantic descriptions. This can be applied to energy storage battery health management, including but not limited to: battery time-series prediction, anomaly detection, and fault diagnosis, among other operation and maintenance tasks.

[0007] The objective of this invention is achieved through the following technical solution: The first aspect of this invention provides a large-scale intelligent operation and maintenance method for time-series understanding of energy storage clusters, comprising the following steps:

[0008] (1) Obtain battery operation sequence data, including the current of each battery module and the voltage, temperature and state of charge of each battery cell in the battery module;

[0009] (2) Define a multi-granularity battery semantic table that covers knowledge in the field of battery operation and maintenance;

[0010] (3) Construct a time-series-semantic modality partitioning module driven by dense time slices to divide the battery runtime time sequence data into multiple time slices, and based on the battery semantic table, map the data patterns in each time slice to specific semantic representations, and merge and text-convert the same semantic representations to generate a set of semantic sentences;

[0011] (4) Construct a self-evolving narrative generation module to input the set of semantic sentences into a multi-agent system composed of multiple large language models, establish a battery operation description file generation process through a multi-agent consensus verification mechanism, and fine-tune the large language model through a self-evolving training paradigm.

[0012] (5) When the energy storage battery operation and maintenance is applied online, the structured direct prompt query or the retrieval enhancement generation method in the question and answer scenario is used to generate and output the operation and maintenance solution in natural language form by using the fine-tuned large language model.

[0013] Furthermore, the battery semantic table defines a series of descriptive semantic entries for voltage, temperature, charge state, and current from three perspectives: the overall cluster level, the cluster inconsistency level, and the individual battery level. Each semantic entry includes shape-related semantics and statistically related semantics. Shape-related semantics include trends, inflection points, volatility, and outliers, while statistically related semantics include average, standard deviation, initial value, and final value.

[0014] Furthermore, step (3) specifically includes the following sub-steps:

[0015] (3.1) Time slice division: Divide the battery runtime sequence data of length T to be analyzed into S short time units, i.e. time slices;

[0016] (3.2) Local semantic detection: For each time slice, semantic detection rules are determined according to predefined mathematical rules, and the data patterns in the time slice are mapped to specific semantic categories defined in the battery semantic table according to the semantic detection rules;

[0017] (3.3) Local semantic pattern enhancement: Scan the local semantic detection results of all time slices in chronological order, merge multiple consecutive time slices with the same semantics into a unified semantic segment, and update its duration range to generate an intermediate semantic representation;

[0018] (3.4) Semantic transformation: Transform the intermediate semantic representation into a basic natural language sentence to obtain a set of semantic sentences.

[0019] Furthermore, the length of the time slice is set according to the data sampling frequency and business requirements.

[0020] Furthermore, the mathematical rules include: judging the trend of change based on first-order difference and threshold, judging volatility based on coefficient of variation, judging outliers based on Z-score method, and judging inflection points based on sliding window and fitting error.

[0021] Furthermore, step (4) specifically includes the following sub-steps:

[0022] (4.1) Define the battery operation description file template;

[0023] (4.2) Input the set of semantic sentences into a multi-agent system consisting of multiple large language models. Each large language model generates a draft of the battery runtime sequence description based on the battery runtime description profile template.

[0024] (4.3) Determine whether the semantic similarity between the battery runtime description drafts generated by each large language model is greater than or equal to the preset threshold. If yes, output the text with the highest consensus as the final battery runtime description file and save it in the database. If no, the domain expert intervenes to write or revise it and saves it as the final battery runtime description file in the database.

[0025] (4.4) Supervised training and fine-tuning of all large language models are performed using battery runtime description files, with the goal of minimizing the supervised loss function, and the low-rank fitting parameters of the large language models are iteratively optimized; consensus-guided training and fine-tuning of all large language models are performed using battery runtime sequence description drafts, with the goal of minimizing the consensus loss function, and the low-rank fitting parameters of the large language models are iteratively optimized.

[0026] Furthermore, the large language model includes Qwen, DeepSeek, and GPT.

[0027] Furthermore, the supervised loss function adopts the standard cross-entropy loss, and its calculation formula is as follows:

[0028]

[0029] In the formula, This indicates a supervised loss function. This represents the i-th input sample of the large language model, that is, the i-th semantic sentence in the set of semantic sentences. The parameters representing the large language model. express The corresponding battery operation description file, This means that given the input is and parameters Under the condition of large language model output The probability of M is given by M, which represents the total number of training samples.

[0030] The formula for calculating the consensus loss function is as follows:

[0031]

[0032] In the formula, Let N represent the consensus loss function, and let N represent the number of large language models in the agent system. and These represent the battery runtime sequence description drafts generated by the j-th and k-th large language models, respectively. Distance metric representing the semantic similarity between two battery runtime sequence description drafts.

[0033] Furthermore, the direct prompt query specifically includes: combining the battery operation description file with a natural language query to form prompt words, directly inputting them into the fine-tuned and trained large language model agent, and the agent providing the text in the natural language form with the highest consensus as the generated answer, which is the operation and maintenance solution;

[0034] The enhanced retrieval generation method specifically includes: decomposing historically generated battery operation description files into fragments and constructing a dedicated vector database; when a new query question is input, first retrieve the battery operation description file fragment most relevant to the current query question from the dedicated vector database, and then send this fragment and the query question together into the finely tuned and trained large language model agent to generate the answer, which is the operation and maintenance solution.

[0035] A second aspect of the present invention provides a large-scale intelligent operation and maintenance system for time-series understanding of energy storage clusters, 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 above-mentioned large-scale intelligent operation and maintenance method for time-series understanding of energy storage clusters.

[0036] The beneficial effects of this invention are as follows: This invention transforms complex numerical time-series data into natural language descriptions familiar to human experts through an automated "signal-semantic-text" conversion pipeline, making the decision-making process of deep learning models transparent, traceable, and easy to understand; This invention ensures that the analysis can simultaneously cover the macroscopic state of the battery cluster and the microscopic behavior of individual batteries through battery semantic tables and battery operation description profiles, revealing key issues such as inconsistencies, providing comprehensive analytical dimensions; Through a time-slice-driven temporal-semantic modality partitioning module, it transforms the raw time-series data processing that large models are not good at into text understanding and generation tasks that they excel at, achieving high-quality time-series data analysis without the need for complex, raw data-oriented pre-training of large language models; Simultaneously, through a multi-agent consensus verification mechanism, it effectively avoids the illusions or errors that may arise from a single large language model, ensuring the professionalism and accuracy of the output text while improving automation; Compared with existing technologies, the battery operation description profiles generated by this invention can be directly used to generate reports, assist in decision-making, and even drive automated operation and maintenance processes, greatly improving operation and maintenance efficiency. Attached Figure Description

[0037] Figure 1 The flowchart below shows the intelligent operation and maintenance method for large-scale models based on time-series understanding of energy storage clusters according to the present invention.

[0038] Figure 2 This is a battery operation sequence diagram of the present invention; wherein, Figure 2 (a) in the figure is an example of battery operation sequence diagram corresponding to the average voltage of the lithium battery module; Figure 2(b) in the figure is an example of battery operation sequence diagram corresponding to the voltage standard deviation of the lithium battery module; Figure 2 (c) in the figure is an example diagram of the battery operation sequence corresponding to the current of the lithium battery module; Figure 2 (d) in the figure is an example diagram of the battery operation sequence corresponding to the voltage of single cell #1; Figure 2 (e) in the figure is an example diagram of the battery operation sequence corresponding to the temperature of single cell #1; Figure 2 (f) in the figure is an example of battery operation sequence diagram corresponding to the SOC of single cell #1;

[0039] Figure 3 This is a runtime sequence example diagram of the method of the present invention in an anomaly detection task;

[0040] Figure 4 This is another runtime sequence example diagram of the method of the present invention in an anomaly detection task;

[0041] Figure 5 This is a runtime sequence example diagram of the method of the present invention in a time series prediction task;

[0042] Figure 6 This is a schematic diagram of a large-scale intelligent operation and maintenance system for time-series understanding of energy storage clusters according to the present invention. Detailed Implementation

[0043] 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.

[0044] 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.

[0045] 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."

[0046] 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.

[0047] See Figure 1 The intelligent operation and maintenance method for large-scale models of energy storage clusters based on time-series understanding of the present invention specifically includes the following steps:

[0048] (1) Obtain battery operation sequence data, including the current of each battery module and the voltage, temperature and state of charge (SOC) of each battery cell in the battery module.

[0049] Specifically, the battery operation sequence data collected by the battery management system in the energy storage battery system is obtained. The battery module is the smallest unit of analysis. The battery module is composed of several battery cells connected in series. The battery cells are all lithium-ion batteries and are the smallest unit of energy storage. The operation sequence data collected for each battery module includes the voltage, temperature, state of charge of each battery cell in the battery module and the current of each battery module.

[0050] In this embodiment, the battery module consists of 16 lithium iron phosphate battery cells connected in series. All battery operation timing data were sampled at a frequency of 1 minute, resulting in a total of 100 sets of operation timing data collected from 28 battery modules. This timing data covers charging, discharging, and idle states under various power settings.

[0051] (2) Define a multi-granularity battery semantic table that covers knowledge of battery operation and maintenance.

[0052] Furthermore, the battery semantic table defines a series of descriptive semantic entries for signals such as voltage, temperature, state of charge, and current from three perspectives: the overall cluster level, the cluster inconsistency level, and the individual battery level. Each semantic entry includes shape-related semantics and statistically related semantics. Shape-related semantics include trends, inflection points, volatility, and outliers, while statistically related semantics include averages, standard deviations, initial values, and final values. In this embodiment, the battery semantic table is specifically shown in Table 1, where checked tags represent features selected as semantic expressions.

[0053] Table 1: Battery Semantic Table

[0054]

[0055] (3) Construct a time-series-semantic modality partitioning module driven by dense time slices to divide the battery runtime time-series data into multiple time slices. Based on the battery semantic table, map the data patterns in each time slice to specific semantic representations, and merge and text-convert identical semantic representations to generate a set of semantic sentences. The time-series-semantic modality partitioning module includes four sub-steps: time slice partitioning, local semantic detection, local semantic pattern enhancement, and semantic conversion.

[0056] (3.1) Time Slicing: The multivariate time series of length T to be analyzed (i.e., battery runtime time series data) is divided into S short time units, i.e., time slices, each slice having a length of... ,in Slice length The settings can be configured according to the data sampling frequency and business requirements to ensure that the data pattern within each time slice is relatively simple and stable. In this embodiment, T is set to 100, and the slice length is... If the value is 10, then S is 10.

[0057] (3.2) Local semantic detection: For each time slice, semantic detection rules are determined according to predefined and generalizable mathematical rules, and the data patterns in the time slice are mapped to specific semantic categories defined in the battery semantic table according to the semantic detection rules.

[0058] Furthermore, the mathematical rules include, but are not limited to: judging trends based on first-order differences and thresholds, judging volatility based on the coefficient of variation, judging outliers based on the Z-score method, judging inflection points based on sliding windows and fitting errors, and so on. Based on these mathematical rules, semantic detection rules can be determined, and these rules can then be used to judge shape-related semantics such as trends, inflection points, volatility, and outliers, as well as statistically related semantics such as mean, standard deviation, initial value, and final value. The defined semantic detection rules are detailed in Table 2.

[0059] Table 2: Semantic Detection Rule Table

[0060]

[0061] In this embodiment, The value is 0.001. The value is 0.005. The value is 1.25; The value is 0.01; This represents the base sample size, which has a value of 4. This represents the number of consecutive error samples, and its value is 3. The value is 0.2.

[0062] (3.3) Local semantic pattern enhancement: Scan the local semantic detection results of all time slices in chronological order, merge multiple consecutive time slices with the same semantics into a unified semantic segment, and update its duration range, thereby generating an intermediate semantic representation with higher information density and more conciseness.

[0063] Specifically, starting from the first time slice, the semantic information of adjacent time slices is compared. If a consistent pattern is found, their descriptions are merged. For example, if the voltage change trend of a certain battery cell shows a "continuous increase" in the first to fourth time slices, it can be integrated into a unified local semantic "continuous increase".

[0064] (3.4) Semantic conversion: The intermediate semantic representation obtained in step (3.3) above (e.g., time range: points 1-50; target signal: voltage of battery No. 1; trend of change: rising) is converted into basic natural language sentences (e.g., "from time point 1 to time point 50, the voltage of battery No. 1 continues to rise") to obtain a set of semantic sentences.

[0065] (4) Construct a self-evolving narrative generation module to input the set of semantic sentences into a multi-agent system composed of multiple large language models. Establish a reliable and highly automated battery operation description file generation process through a multi-agent consensus verification mechanism. Fine-tune the large language model through a self-evolving training paradigm to improve the temporal narrative generation performance of the large language model.

[0066] (4.1) Battery Operation Description Profile Template Definition: The final organizational structure of the predefined battery operation description profile template usually follows a global to local order. Specifically, it first describes the overall operation of the cluster, then describes inconsistencies within the cluster, and finally describes the behavior of individual batteries that are significant or problematic. In addition, due to the limitations of the input and output of the large language model, cluster operation and individual cell operation are input separately, and each individual cell template can only contain 4 individual cells per input, that is, 16 individual cells need to be input in 4 separate inputs.

[0067] In this embodiment, the battery operation description file template defined in the cluster operation description is as follows:

[0068] "Lithium-ion battery time series description text: "

[0069] Description period: {description_begin_period} ~ {description_end_period}

[0070] Description object: {description_object}

[0071] I. Overall Operational Conclusion:

[0072] 1. **Overall Operation Process**: {overall_operation}

[0073] 2. **Overall inconsistency**: {overall_inconsistency}

[0074] II. **Cluster Perspective**:

[0075] 1.1. Voltage characteristics:

[0076] 1.1.1. Overall Operation Status:

[0077] a) Average voltage of the lithium battery module:

[0078] **Trend**: {average_voltage_trend}

[0079] **Turning Point**: {average_voltage_transition_events}

[0080] **Fluctuation**: {average_voltage_fluctuation}

[0081] **Abnormal Phenomenon: {average_voltage_outlier_phenomenon}

[0082] **Mean and Standard Deviation (Steady Phase)**: {average_voltage_mean_variance}

[0083] **Initial / Final Value (Increase or Decrease Phase)**: {average_voltage_initial_final}

[0084] b) Maximum voltage of the lithium battery module:

[0085] **Trend**: {max_voltage_trend}

[0086] **Turning Point**: {max_voltage_transition_events}

[0087] **Abnormal Phenomenon: {max_voltage_outlier_phenomenon}

[0088] **Mean and Standard Deviation (Steady Phase)**: {max_voltage_mean_std}

[0089] **Initial / Final Value (Increase or Decrease Phase)**: {max_voltage_initial_final}

[0090] c) Minimum voltage of lithium battery module:

[0091] **Trend**: {min_voltage_trend}

[0092] **Turning Point**: {min_voltage_transition_events}

[0093] **Abnormal Phenomenon: {min_voltage_outlier_phenomenon}

[0094] **Mean and Standard Deviation (Steady Phase)**: {min_voltage_mean_std}

[0095] **Initial / Final Value (Increase or Decrease Phase)**: {min_voltage_initial_final}

[0096] 1.1.2. Voltage inconsistency between battery cells:

[0097] a) Standard deviation:

[0098] **Amplitude Level**: {voltage_std_amplitude}

[0099] **Fluctuation Level**: {voltage_std_fluctuation}

[0100] **Mean and Standard Deviation**: {voltage_std_mean_std}

[0101] b) Shannon entropy:

[0102] **Amplitude Level**: {voltage_entropy_amplitude}

[0103] **Fluctuation Level**: {voltage_entropy_fluctuation}

[0104] **Mean and Standard Deviation**: {voltage_entropy_mean_std}

[0105] 1.2. Temperature characteristics:

[0106] 1.2.1. Average voltage of the lithium battery module:

[0107] ...

[0108] 1.3. SOC Characteristics

[0109] 1.3.1. Average SOC of lithium battery module:

[0110] ...

[0111] (The template for temperature and SOC is the same as that for voltage)

[0112] 1.4. Current characteristics:

[0113] **Trend**: {current_trend}

[0114] **Turning Point**: {current_transition_events}

[0115] **Fluctuation**: {current_fluctuation}

[0116] **Abnormal Phenomenon: {current_outlier_phenomenon}**

[0117] **Mean and Standard Deviation (Stability Phase)**: {current_mean_std}

[0118] **Initial / Final Value (Increase or Decrease Phase)**: {current_initial_final}

[0119] Please extract the data from the lithium battery module based on the provided information and return it as a JSON array.

[0120] All elements constitute a single JSON object containing the following keys:

[0121] {-"Overall_operation": {"overall_operation": "...", "overall_inconsistency": "..."};

[0122] -"Voltage":{"average":{"trend": "...", "transition_events": "...", "fluctuation": "...", "outliers": "...", "mean_and_std": "...", "initial_final": "..."},

[0123] "maximum":{"trend": "...", "transition_events": "...", "outliers":"...", "mean_and_std": "...", "initial_final": "..."},

[0124] "minimum":{"trend": "...", "transition_events": "...", "outliers":"...", "mean_and_std": "...", "initial_final": "..."},

[0125] "standard_deviation":{"amplitude": "...", "fluctuation": "...", "mean_and_std": "..."},

[0126] "shannon_entropy":{"amplitude": "...", "fluctuation": "...", "mean_and_std": "..."}};

[0127] -"Temperature":{"average":{"trend": "...", "transition_events":"...", "fluctuation": "...", "outliers": "...", "mean_and_std": "...", "initial_final": "..."},

[0128] "maximum":{"trend": "...", "transition_events": "...", "outliers":"...", "mean_and_std": "...", "initial_final": "..."},

[0129] "minimum":{"trend": "...", "transition_events": "...", "outliers":"...", "mean_and_std": "...", "initial_final": "..."},

[0130] "standard_deviation":{"amplitude": "...", "fluctuation": "...", "mean_and_std": "..."},

[0131] "shannon_entropy":{"amplitude": "...", "fluctuation": "...", "mean_and_std": "..."}};

[0132] -"SOC":{"average":{"trend": "...", "transition_events": "...", "fluctuation": "...", "outliers": "...", "mean_and_std": "...", "initial_final": "..."},

[0133] "maximum":{"trend": "...", "transition_events": "...", "outliers":"...", "mean_and_std": "...", "initial_final": "..."},

[0134] "minimum":{"trend": "...", "transition_events": "...", "outliers":"...", "mean_and_std": "...", "initial_final": "..."},

[0135] "standard_deviation":{"amplitude": "...", "fluctuation": "...", "mean_and_std": "..."},

[0136] "shannon_entropy":{"amplitude": "...", "fluctuation": "...", "mean_and_std": "..."}};

[0137] -"Current":{"trend": "...", "transition_events": "...", "fluctuation": "...", "outliers": "...", "mean_and_std": "...", "initial_final": "..."}}”。

[0138] In the battery operation description file template corresponding to the above cluster operation, `description_begin_period` represents the start period of the description, `description_end_period` represents the end period of the description, `description_object` represents the description object, `overall_operation` represents the overall operation status, `overall_inconsistency` represents overall inconsistency, `average_voltage_trend` represents the average voltage trend, `average_voltage_transition_events` represents the average voltage transition points, `average_voltage_fluctuation` represents the average voltage fluctuation, `average_voltage_outlier_phenomenon` represents the average voltage outlier phenomenon, `average_voltage_mean_variance` represents the mean and standard deviation of the average voltage, `average_voltage_initial_final` represents the initial / final value of the average voltage, `max_voltage_trend` represents the maximum voltage trend, `max_voltage_transition_events` represents the maximum voltage transition point, and `max_voltage_outlier_phenomenon` represents the maximum voltage outlier phenomenon. `enon` represents the maximum voltage anomaly; `max_voltage_mean_std` represents the mean and standard deviation of the maximum voltage; `max_voltage_initial_final` represents the initial / final value of the maximum voltage; `min_voltage_trend` represents the minimum voltage trend; `min_voltage_transition_events` represents the minimum voltage transition points; `min_voltage_outlier_phenomenon` represents the minimum voltage anomaly; `min_voltage_mean_std` represents the mean and standard deviation of the minimum voltage; `min_voltage_initial_final` represents the initial / final value of the minimum voltage; `voltage_std_amplitude` represents the amplitude of the voltage standard deviation; `voltage_std_fluctuation` represents the fluctuation of the voltage standard deviation; `voltage_std_mean_std` represents the mean and standard deviation of the voltage standard deviation; `voltage_entropy_amplitude` represents the amplitude of the voltage Shannon entropy; `voltage_entropy_fluctuation` represents the fluctuation of the voltage Shannon entropy; `voltage_entropy_mean_std` represents the mean and standard deviation of the voltage Shannon entropy; `current_trend` represents the current trend.`current_transition_events` represents current transition points, `current_fluctuation` represents current fluctuations, `current_outlier_phenomenon` represents current anomalies, `current_mean_std` represents the current mean and standard deviation, and `current_initial_final` represents the initial / final current value. The JSON object contains the following keys:

[0139] {-"Overall Operation Status": {"Overall Operation Status": "...", "Overall Inconsistency": "..."};

[0140] -"Voltage":{"Average":{"Trend": "...", "Turning Point": "...", "Fluctuation": "...", "Outliers": "...", "Mean and Standard Deviation": "...", "Initial / Final Value": "..."},

[0141] "Maximum":{"Trend": "...", "Turning Point": "...", "Outliers": "...", "Mean and Standard Deviation": "...", "Initial / Final Value": "..."},

[0142] "Minimum":{"Trend": "...", "Turning Point": "...", "Outliers": "...", "Mean and Standard Deviation": "...", "Initial / Final Value": "..."},

[0143] "Standard deviation":{"Amplitude": "...", "Volatility": "...", "Mean and standard deviation": "..."},

[0144] Shannon entropy:{"Amplitude": "...", "Volatility": "...", "Mean and Standard Deviation": "..."}};

[0145] -"Temperature":{"Average":{"Trend": "...", "Turning Point": "...", "Volatility": "...", "Outliers": "...", "Mean and Standard Deviation": "...", "Initial / Final Values": "..."},

[0146] "Maximum":{"Trend": "...", "Turning Point": "...", "Outliers": "...", "Mean and Standard Deviation": "...", "Initial / Final Value": "..."},

[0147] "Minimum":{"Trend": "...", "Turning Point": "...", "Outliers": "...", "Mean and Standard Deviation": "...", "Initial / Final Value": "..."},

[0148] "Standard deviation":{"Amplitude": "...", "Volatility": "...", "Mean and standard deviation": "..."},

[0149] Shannon entropy:{"Amplitude": "...", "Volatility": "...", "Mean and Standard Deviation": "..."}};

[0150] -"State of Charge":{"Average":{"Trend": "...", "Turning Point": "...", "Volatility": "...", Outliers": "...", "Mean and Standard Deviation": "...", "Initial / Final Values": "..."},

[0151] "Maximum":{"Trend": "...", "Turning Point": "...", "Outliers": "...", "Mean and Standard Deviation": "...", "Initial / Final Value": "..."},

[0152] "Minimum":{"Trend": "...", "Turning Point": "...", "Outliers": "...", "Mean and Standard Deviation": "...", "Initial / Final Value": "..."},

[0153] "Standard deviation":{"Amplitude": "...", "Volatility": "...", "Mean and standard deviation": "..."},

[0154] Shannon entropy:{"Amplitude": "...", "Volatility": "...", "Mean and Standard Deviation": "..."}};

[0155] -"Current":{"Trend": "...", "Turning Point": "...", "Volatility": "...", "Outliers":"...", "Mean and Standard Deviation": "...", "Initial / Final Values": "..."}}.

[0156] In this embodiment, the battery operation description file template defined in the individual operation description is as follows:

[0157] Output template example:

[0158] Single cell #1:\n

[0159] **Voltage**:\n

[0160] Trend: {cell_1_voltage_trend}\n

[0161] Turning point: {cell_1_voltage_transition_events}\n

[0162] Fluctuation: {cell_1_voltage_fluctuation}\n

[0163] Outliers: {cell_1_voltage_outliers}\n

[0164] Mean and standard deviation (steady-state phase): {cell_1_voltage_mean_std}\n

[0165] Initial / Final Value (Increase or Decrease Phase): {cell_1_voltage_initial_final}\n

[0166] **Temperature**:\n

[0167] ...

[0168] **SOC**:\n

[0169] ...

[0170] (The template for temperature and SOC is the same as that for voltage)

[0171] Please extract the data for each individual battery cell (from cell #1 to cell #4) based on the provided information and return it as a JSON array.

[0172] Each array element is a JSON object for each individual battery cell, containing the following keys:

[0173] {"cell id": "...";

[0174] "voltage":{"trend": "...", "transition": "...", "fluctuation": "...","outliers": "...", "mean_and_std": "...", "initial_final": "..."};

[0175] "temperature":{"trend": "...", "transition": "...", "fluctuation":"...", "outliers": "...", "mean_and_std": "...", "initial_final": "..."};

[0176] "soc":{"trend": "...", "transition": "...", "fluctuation": "...", "outliers": "...", "mean_and_std": "...", "initial_final": "..."}};

[0177] After generating a JSON object for each individual battery cell, please arrange them into a JSON array and return it in the following format:

[0178] {"Cell Information": [{JSON for Cell 1}, {JSON for Cell 2}, ..., {JSON for Cell 4}]} (Please check the validity of the format!)

[0179] In the battery operation description file template corresponding to the above individual operation, `cell_1_voltage_trend` represents the voltage trend of cell #1, `cell_1_voltage_transition_events` represents the voltage transition points of cell #1, `cell_1_voltage_fluctuation` represents the voltage fluctuations of cell #1, `cell_1_voltage_outliers` represents the voltage outliers of cell #1, `cell_1_voltage_mean_std` represents the voltage mean and standard deviation of cell #1, and `cell_1_voltage_initial_final` represents the initial / final voltage value of cell #1. The JSON object for each cell contains the following keys:

[0180] {"Cell serial number": "...";

[0181] "Voltage":{"Trend": "...", "Turning Point": "...", "Volatility": "...", "Outliers":"...", "Mean and Standard Deviation": "...", "Initial / Final Value": "..."};

[0182] "Temperature":{"Trend": "...", "Turning Point": "...", "Volatility": "...", "Outliers": "...", "Mean and Standard Deviation": "...", "Initial / Final Value": "..."};

[0183] "State of charge":{"Trend": "...", "Turning point": "...", "Volatility": "...", "Outliers":"...", "Mean and standard deviation": "...", "Initial / final value": "..."}}.

[0184] (4.2) Generate a draft battery runtime sequence description: Construct an agent system containing at least two different large language models to input all semantic sentence sets obtained in step (3) into each large language model agent. The agent is required to generate a complete draft battery runtime sequence description based on the battery runtime description archive template defined in step (4.1). Figure 2 Taking the battery runtime sequence diagram shown as an example, this section explains the draft battery runtime sequence description generated by the large language model. For instance, Figure 2 As shown in (a), the draft battery runtime description generated for the average voltage of the lithium battery module is: "Average voltage of the lithium battery module:

[0185] **trend**:

[0186] The average voltage of the lithium battery module showed an upward trend, rising from 3.2569V to 3.3317V from the 1st to the 60th time, and then remaining stable from the 61st to the 100th time, rising from 3.3325V to 3.3386V.

[0187] **Turning Point**:

[0188] No significant transition events were observed in the average voltage.

[0189] **fluctuation**:

[0190] Within the stated period, the average voltage fluctuates very little.

[0191] **Abnormal Phenomenon**:

[0192] No abnormalities were observed in the average voltage.

[0193] like Figure 2 As shown in (b), the draft battery operation sequence description generated for the voltage standard deviation of the lithium battery module is: "Voltage inconsistency between batteries:

[0194] **Standard deviation:**

[0195] In the first to the 100th measurements, the standard deviation of the inter-cell voltage was small, with an average value of 0.0007.

[0196] like Figure 2 As shown in (c), the draft battery operation sequence description for current generation of the lithium battery module is: "Current characteristics:

[0197] **trend**:

[0198] The current stabilized between 53.4000A and 56.0000A for the 1st to 10th cycles, between 54.8000A and 57.1000A for the 11th to 30th cycles, between 52.8000A and 54.4000A for the 31st to 40th cycles, and between 51.4000A and 53.1000A for the 41st to 100th cycles.

[0199] **Turning Point**:

[0200] No obvious transition events were observed in the current.

[0201] **fluctuation**:

[0202] The current value fluctuates very little during the time period.

[0203] **Abnormal Phenomenon**:

[0204] No abnormalities were observed in the current situation.

[0205] like Figure 2 As shown in (d), the draft battery operation sequence description generated for the voltage of single cell #1 is: "**Voltage**:

[0206] Trend: In samples 1-60, the value increased from 3.2570 to 3.3320; in samples 61-100, the value stabilized from 3.3330 to 3.3390.

[0207] Turning point: The turning point from increase to stability was reached with the 61st sample.

[0208] Volatility: Low

[0209] Outliers: None

[0210] Mean and variance (stable phase): Mean ~3.336, Variance ~0.000006 (stable phases from samples 61 to 100)

[0211] Initial / Final values ​​(used to increase or decrease phase): Initial value: 3.2570, Final value: 3.3390.

[0212] like Figure 2As shown in (e), the draft battery operation sequence description generated for the temperature of single cell #1 is: "**Temperature**:

[0213] Trend: From the 1st to the 100th sample, the value remained stable between 25.0000 and 26.0000.

[0214] Turning point: None

[0215] Volatility: Low

[0216] Outliers: None

[0217] Mean and variance (stable phase): Mean ~25.5, Variance ~0.25

[0218] Initial / Final Values ​​(used to increase or decrease phase): Initial: 25.0000, Final: 26.0000.

[0219] like Figure 2 As shown in (f), the draft battery runtime sequence description generated for the SOC of single cell #1 is: "**SOC**:

[0220] Trend: From the 1st to the 100th sample, the increase was from 8,000 to 39,000.

[0221] Turning point: None

[0222] Volatility: Low

[0223] Outliers: None

[0224] Mean and variance (for stable phases): Not applicable (no stable phase).

[0225] Initial / Final Values ​​(used to increase or decrease phase): Initial: 8.0000, Final: 39.0000.

[0226] Furthermore, large language models include Qwen, DeepSeek, GPT, etc.

[0227] In this embodiment, the intelligent agent system is composed of two large language models, Qwen3-14b and Qwen2.5-14b, and prompt words are provided corresponding to the battery operation description profile template. The specific prompt words for the cluster description are as follows:

[0228] "You are an expert specializing in lithium-ion battery time series signal analysis and maintenance. Your task is to complete the following 'Lithium-ion Battery Time Series Description Text' template based on the specific information provided by the user."

[0229] Notice:

[0230] 1. Please strictly follow the logic of the text template, fill in the corresponding parts within the curly braces, and convert them to JSON format.

[0231] 2. When you fill in {overall_operation}, you need to describe the charging / discharging behavior of the lithium battery module, for example: "From sample 1 to 50, the lithium battery module is charging, and from sample 50 to 100, it is discharging."

[0232] When you fill in {overall_inconsistency}, you need to highlight any significant inconsistencies in the lithium battery modules, for example: "For most of the time, the individual battery cells showed good consistency. However, a moderate standard deviation was observed in the SOC of samples 1 through 20."

[0233] 3. Descriptions should use concise language. This is especially important for descriptions of "overall operation":

[0234] **In the "Trend" section, please output "Increase (times 1-50), Stabilize (times 50-100)" instead of "From samples 1 to 50, increase from 29.0℃ to 30.0℃; from samples 51 to 100, stabilize from 30.0℃ to 31.0℃."

[0235] **In the "Mean and Standard Deviation" section, only consider the mean and standard deviation during the "steady-state" phase. Please output the mean and standard deviation separately for each steady-state phase, such as "Steady-state at 3.2V, standard deviation of 0.0043 (times 1-50)."

[0236] **In the "Initial / Final Value" section, only consider the initial / final value for the "Increase" (or "Decrease") phase. Please output the values ​​for each increase (or decrease) phase separately, such as "3.2V to 3.3V (times 1-50)."

[0237] Example guidance: When you fill in the {average_voltage_trend} section, you will find "increase from 3.2V to 4.1V" in the user information.

[0238] Then you should fill in "The average voltage of the lithium battery module shows an upward trend, increasing from 3.2V to 4.1V during the described period" in {average_voltage_trend}.

[0239] For descriptions of standalone operation, the specific prompts are as follows:

[0240] "You are an expert specializing in lithium-ion battery time series signal analysis and maintenance. Your task is to complete the following template based on the specific information provided by the user."

[0241] Notice:

[0242] 1. Please extract the data for each lithium battery cell (from cell #1 to cell #4) based on the provided information and return it as a JSON array.

[0243] 2. Descriptions should use concise language. For example:

[0244] **In the "Trend" section, please output "Increase (times 1-50), Stabilize (times 50-100)" instead of "From samples 1 to 50, increase from 29.0℃ to 30.0℃; from samples 51 to 100, stabilize from 30.0℃ to 31.0℃."

[0245] **In the "Mean and Standard Deviation" section, only consider the mean and standard deviation during the "steady-state" phase. Please output something like "Stable at 3.2V, standard deviation of 0.0043 (times 1-50)."

[0246] **In the "Initial / Final Value" section, only consider the initial / final values ​​during the "Increase" (or "Decrease") phase. Please output something like "3.2V to 3.3V (times 1-50)."

[0247] **In the "Volatility" field, if the user information does not provide a specific volatility level, please output "None"; otherwise, please output based on the user information, such as "Volatility was detected at time points 1-20 and 40-50."

[0248] Example guidance: When you fill in the {cell_1_voltage_trend} section, you will find "increase from 3.2V to 4.1V" in the user information.

[0249] Then you should fill in "from sample 1 to 50, increase from 3.2V to 4.1V" in {cell_1_voltage_trend}.

[0250] (4.3) Multi-agent consensus verification mechanism: Calculate the semantic similarity between the battery runtime sequence description drafts generated by each large language model, and determine whether the similarity is greater than or equal to the preset threshold. If the similarity is greater than or equal to the preset threshold, consensus is considered to be achieved, and the text with the highest consensus is output as the final battery runtime description file and stored in the database. Otherwise, if the similarity is less than the preset threshold, the sample is marked as having uncertainty, triggering the manual verification process. Domain experts intervene to write or revise the draft, and the written or revised battery runtime sequence description draft is stored as the final battery runtime description file in the database.

[0251] Furthermore, semantic similarity can be obtained using metrics such as BERT score (BERTScore) or ROUGE (Recall-Oriented Understudy for Gisting Evaluation). In this embodiment, the BERTScore metric is used, with a preset threshold of 0.9.

[0252] (4.4) Self-evolutionary training of large language models: Supervised training and fine-tuning of all large language models of the agent system are performed using the high-reliability battery runtime description files stored in the database in step (4.3). During the fine-tuning process, the low-rank adaptation (LoRA) method is used to iteratively optimize the low-rank adaptation parameters of the large language models with the goal of minimizing the supervised loss function. For the remaining samples that lack high-reliability verification, i.e., the battery runtime sequence description drafts generated by the large language models that have not been saved in the database, these samples lacking high-reliability verification are used to perform consensus-guided training and fine-tuning of all large language models of the agent system. The aim is to minimize the output differences between different large language models, thereby prompting them to form a unified and reliable consensus understanding of the same time series. During the fine-tuning process, the low-rank adaptation parameters of the large language models are iteratively optimized using the LoRA method with the goal of minimizing the consensus loss function.

[0253] Furthermore, the supervised loss function adopts the standard cross-entropy loss, and its calculation formula is as follows:

[0254]

[0255] In the formula, This represents a supervised loss function; This represents the i-th input sample of the large language model, that is, the i-th semantic sentence in the set of semantic sentences; Represents the parameters of a large language model; express The corresponding battery operation description file; This means that given the input is and parameters Under the condition of large language model output The probability of ; M represents the total number of training samples.

[0256] Furthermore, the formula for calculating the consensus loss function is as follows:

[0257]

[0258] In the formula, Represents the consensus loss function; N represents the number of large language models in the agent system; and These represent the battery runtime sequence description drafts generated by the j-th and k-th large language models, respectively; Distance metric representing the semantic similarity between two battery runtime sequence description drafts, when The smaller the value, the more semantically similar the two battery runtime sequence description drafts are. Indicators such as BERTScore or ROUGE can be used.

[0259] In this embodiment, the BERTScore metric is used, and the fine-tuning parameters are set with a learning rate of 0.0001, 5 training epochs, a batch size of 4, a LoRA Rank of the low-rank matrix of 8, a LoRA Alpha of the low-rank matrix of 16, and a LoRA Dropout of 0.05.

[0260] (5) When the energy storage battery operation and maintenance is applied online, the structured direct prompt query or the retrieval enhancement generation method in the question and answer scenario is used to generate and output the operation and maintenance solution in natural language form by using the fine-tuned large language model.

[0261] Furthermore, the direct prompt query specifically includes: combining the battery operation description file with natural language queries (such as: "What is the trend of the average SOC change of the module?", "Are there any abnormal batteries?", etc.) to form prompt words, and directly inputting them into the large language model agent after fine-tuning training obtained in step (4). The agent gives the text in the form of natural language with the highest consensus as the generated answer, which is the operation and maintenance solution.

[0262] Furthermore, the enhanced retrieval generation method specifically includes: decomposing a large number of historically generated battery operation description files into fragments to construct a dedicated vector database; when a new query question is input, the battery operation description file fragment most relevant to the current query question is first retrieved from the dedicated vector database, and then this fragment and the query question are fed into the finely tuned and trained large language model agent to generate a more accurate and context-rich answer, which is the operation and maintenance solution.

[0263] Furthermore, maintenance tasks include, but are not limited to: battery timing prediction, anomaly detection, and fault diagnosis.

[0264] In this embodiment, the prompt word for the anomaly detection task is:

[0265] "You are a professional battery diagnostics expert, specializing in time series signal analysis of lithium-ion batteries."

[0266] Please analyze the time series of the following battery module (composed of 16 cells connected in series) and determine if there are any anomalies during this time period.

[0267] The following is some basic information about the relevant lithium battery cells: Nominal voltage: 3.2V, Operating voltage: 2.5~3.65V, Charging / discharging protocol: Constant power, Ambient temperature: 25℃.

[0268] This basic information shows the expected normal operating range of the battery, which is very important for fault diagnosis tasks.

[0269] You should perform monitoring tasks at both the cluster level (cluster behavior) and the individual level (each specific cell).

[0270] If a significant anomaly is detected, you should select the most likely anomaly type from the following options:

[0271] A. Abnormal temperature within the cells of a specific individual;

[0272] B. The temperature of the entire lithium battery module is abnormal;

[0273] C. Voltage anomalies in specific individual cells;

[0274] D. The voltage of the entire lithium battery module is abnormal.

[0275] In addition, you should also identify the start and end samples of the detected anomalies, for example: "samples 1 through 50".

[0276] Output format: You should use "operation as" as <ans> normal< / ans> The "or" operation is <ans> abnormal< / ans> The answer should be returned in the format of ", followed by relevant analysis.

[0277] If an anomaly is detected, you should output its fault type, start time, and end time.

[0278] The above prompts are input into a large language model to perform an anomaly detection task. Figure 3 Taking the runtime sequence diagram shown as an example, the anomaly detection task is executed, and the question is: "Please diagnose the potential anomalies in this time series, considering the following factors:"

[0279] Cluster level: Identifies systematic deviations in collective behavior patterns, indicating anomalies across the entire module.

[0280] Individual level: By cross-comparing with similar cells, abnormalities in individual cells are detected. For example, if the voltage of cell No. 1 is different from that of other cells, an abnormality is detected in cell No. 1.

[0281] You should select the most likely exception type from these options:

[0282] A. Abnormal temperature of a specific individual battery; B. Abnormal temperature of the entire lithium battery module.

[0283] C. Charge / discharge command communication failure; D. Individual cell voltage abnormality. The correct answer is: "A. Battery No. 1 has an abnormal temperature." Using the method described in this invention to perform the anomaly detection task, the final result generated is: "Anomaly type is..." <ans> A< / ans> .

[0284] **analyze:**

[0285] **Cluster-level analysis:**

[0286] Cluster-level data shows a systematic problem with inconsistent temperatures. The temperature standard deviation for the first 50 samples is reported as "slight," but it becomes "moderate" starting from samples 51-80 and 91-100... This significant increase in dispersion indicates an anomaly across the module, with cells no longer maintaining uniform temperature...

[0287] **Individual-level analysis:**

[0288] A detailed review of the temperature profiles of individual cells identified cell #1 as a clear outlier. Its behavior was distinctly different from all 15 other cells:

[0289] *During the initial discharge phase (samples 41-50), when the current became negative, the temperature of all other batteries remained stable (according to their descriptions: "Stable (cycles 1-60"), "Stable (cycles 1-80"), etc.). In contrast, **the temperature of battery #1 rose significantly from 29°C to 34°C**.

[0290] *Subsequently, its temperature stabilized at a higher value (35.5℃) than that of similar samples 51-70.

[0291] *Subsequently, during the dormancy period (samples 81-90), when the current was zero, the temperature of all other batteries remained stable. However, **the temperature of battery #1 plummeted from 32°C to 26°C**, indicating that its cooling rate was abnormally high among all batteries.

[0292] by Figure 4 Taking the runtime sequence diagram shown as an example, the anomaly detection task is executed, and the question is: "Please determine whether there are any anomalies during this time period. You should perform the anomaly detection task using a tiered evaluation approach."

[0293] Cluster level: Identifies systematic deviations in collective behavior patterns, indicating anomalies across the entire module.

[0294] Individual level: By cross-comparing with similar cells, abnormalities in individual cells are detected. For example, if the voltage of cell 1 is different from others, an abnormality is detected in cell 1. The reference answer is: "Operational abnormality in this cycle; the temperature of cell 16 is significantly higher than other cells." Using the method described in this invention to perform the abnormality detection task, the final result generated is: "Operation is..." <ans> abnormal< / ans> .

[0295] **analyze:**

[0296] Based on a hierarchical evaluation of the provided time-series data, anomalies were detected at the individual battery cell level. Cluster-level analysis showed that the charging module exhibited normal and expected behavior, but cross-comparisons of individual batteries revealed significant biases.

[0297] **Cluster-level assessment (normal):** ...

[0298] **Individual-level assessment (abnormal):**

[0299] ***Cross-comparison Anomaly:** A detailed review of the temperature data for individual batteries revealed a significant anomaly. Throughout the entire testing period, the temperature of 15 out of 16 battery cells (numbered #1 to #15) remained consistently within the range of **26.5℃-27.0℃**.

[0300] ***Identified Faulty Units:** In stark contrast, **cell #16** exhibits a significantly different and abnormal temperature profile. The temperature is stable, but the range is wider.

[0301] In this embodiment, the prompt words for the time series prediction task are:

[0302] "You are a professional battery diagnostics expert, specializing in time series signal analysis of lithium-ion batteries."

[0303] Please analyze the time series data of the following battery module and answer the questions:

[0304] Please predict the possible average SOC of all battery cells in the next 10 minutes.

[0305] Please answer in the following format: "Average SOC of the battery cell:" <ans> xx< / ans> %", where xx is your predicted value, followed by relevant analysis. (symbol) <ans> and< / ans> It's necessary for locating the answer!

[0306] The above prompts are input into a large language model to perform a time-series prediction task. Figure 5Taking the runtime sequence diagram shown as an example, the time-series prediction task is performed with the question: "Please predict the possible value of the average SOC of all lithium battery modules in the next 10 minutes." The reference answer is: "The average SOC value of the battery is approximately 79.0625%." Using the method described in this invention to perform the time-series prediction task, the final result is: "According to the provided **lithium-ion battery time series description text**: At sample 100 (end of data), the average SOC is 83.0625%. During the stable discharge phase (samples 71 to 100), the SOC decreases linearly, from approximately 96.0625% to 83.0625%, a total decrease of 13% over 30 minutes. Discharge rate: 13% / 30 = 0.4333% per minute… Predicted SOC decrease = 4.333%. Predicted SOC: 83.0625% - 4.333% = 78.7295%."

[0307] ###predict:

[0308] The average SOC of the lithium battery module is approximately <ans>78.7295< / ans> %.

[0309] For example, experimental verification was conducted using real energy storage power station operation and maintenance data to illustrate and verify the effectiveness of the steps of the present invention. Furthermore, to more clearly demonstrate the superiority of the method described in this invention for energy storage battery health management, a large language model without fine-tuning, Qwen3-14B, and a large model for time-series data, ChatTS, were used for comparison with the method described in this invention. The experimental results are shown in Table 3. For the charge state estimation task, the mean square error (MSE) and mean absolute error (MAE) were selected as evaluation criteria; for the anomaly detection task, accuracy and F1 score were selected as evaluation criteria. The prompts for the two comparison methods are the same as those in the method described in this invention.

[0310] Table 3: Comparison of quantification results from different methods

[0311]

[0312] As shown in Table 3, the method described in this invention, compared with the other two methods, demonstrates its effectiveness. Compared to models that do not perform semantic transformation and self-learning, the method described in this invention achieves optimal results on all metrics of both tasks.

[0313] The large language models used in this invention include, but are not limited to, GLM4, Qwen2.5, 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.

[0314] See Figure 6 The present invention provides a large-scale intelligent operation and maintenance system for time-series understanding of energy storage clusters, 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 large-scale intelligent operation and maintenance method for time-series understanding of energy storage clusters in the above embodiment.

[0315] The embodiments of the large-scale intelligent operation and maintenance system for time-series understanding of energy storage clusters proposed in this invention can be applied to any device with data processing capabilities, such as a computer or other similar equipment. The device embodiments can be implemented through software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, 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 is a hardware structure diagram of any device with data processing capabilities within the large-scale intelligent operation and maintenance system for time-series understanding of energy storage clusters according to the present invention. (Except for...) Figure 6 In addition to the processor, memory, network interface, and non-volatile memory shown, any data processing device 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.

[0316] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0317] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device 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.

[0318] This invention also provides a computer-readable storage medium storing a program thereon, which, when executed by a processor, implements the large-model intelligent operation and maintenance method for time-series understanding of energy storage clusters described in the above embodiments.

[0319] 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.

[0320] 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.< / ans>

Claims

1. A large-scale intelligent operation and maintenance method for time-series understanding of energy storage clusters, characterized in that, Includes the following steps: (1) Obtain battery operation sequence data, including the current of each battery module and the voltage, temperature and state of charge of each battery cell in the battery module; (2) Define a multi-granularity battery semantic table that covers knowledge in the field of battery operation and maintenance; the battery semantic table specifies a series of descriptive semantic entries for voltage, temperature, charge state and current from three perspectives: the overall cluster level, the cluster inconsistency level and the individual battery level. (3) Construct a time-series-semantic modality partitioning module driven by dense time slices to divide the battery runtime time-series data into multiple time slices, and based on the battery semantic table, map the data patterns in each time slice to specific semantic representations, and merge and text-convert the same semantic representations to generate a set of semantic sentences; step (3) specifically includes the following sub-steps: (3.1) Time slice division: Divide the battery runtime sequence data of length T to be analyzed into S short time units, i.e. time slices; (3.2) Local semantic detection: For each time slice, semantic detection rules are determined according to predefined mathematical rules, and the data patterns in the time slice are mapped to specific semantic categories defined in the battery semantic table according to the semantic detection rules; (3.3) Local semantic pattern enhancement: Scan the local semantic detection results of all time slices in chronological order, merge multiple consecutive time slices with the same semantics into a unified semantic segment, and update its duration range to generate an intermediate semantic representation; (3.4) Semantic transformation: converting intermediate semantic representations into basic natural language sentences to obtain a set of semantic sentences; (4) Construct a self-evolving narrative generation module to input the semantic sentence set into a multi-agent system composed of multiple large language models, establish a battery operation description file generation process through a multi-agent consensus verification mechanism, and fine-tune the large language model through a self-evolving training paradigm; step (4) specifically includes the following sub-steps: (4.1) Define the battery operation description file template; (4.2) Input the set of semantic sentences into a multi-agent system consisting of multiple large language models. Each large language model generates a draft of the battery runtime sequence description based on the battery runtime description profile template. (4.3) Determine whether the semantic similarity between the battery runtime description drafts generated by each large language model is greater than or equal to the preset threshold. If yes, output the text with the highest consensus as the final battery runtime description file and save it in the database. If no, the domain expert intervenes to write or revise it and saves it as the final battery runtime description file in the database. (4.4) Supervised training and fine-tuning of all large language models are performed using battery runtime description files, with the goal of minimizing the supervised loss function, and the low-rank fitting parameters of the large language models are iteratively optimized; consensus-guided training and fine-tuning of all large language models are performed using battery runtime sequence description drafts, with the goal of minimizing the consensus loss function, and the low-rank fitting parameters of the large language models are iteratively optimized. (5) When the energy storage battery operation and maintenance is applied online, the structured direct prompt query or the retrieval enhancement generation method in the question and answer scenario is used to generate and output the operation and maintenance solution in natural language form by using the fine-tuned large language model.

2. The large-scale intelligent operation and maintenance method for time-series understanding of energy storage clusters according to claim 1, characterized in that, In the battery semantic table, each semantic entry contains shape-related semantics and statistical semantics. Shape-related semantics include trend of change, turning point, volatility, and outlier. Statistical semantics include mean, standard deviation, initial value, and final value.

3. The large-scale intelligent operation and maintenance method for time-series understanding of energy storage clusters according to claim 1, characterized in that, The length of the time slice is set according to the data sampling frequency and business requirements.

4. The large-scale intelligent operation and maintenance method for time-series understanding of energy storage clusters according to claim 1, characterized in that, The mathematical rules include: judging the trend of change based on first-order difference and threshold, judging volatility based on coefficient of variation, judging outliers based on Z-score method, and judging inflection points based on sliding window and fitting error.

5. The large-scale intelligent operation and maintenance method for time-series understanding of energy storage clusters according to claim 1, characterized in that, The large language models include Qwen, DeepSeek, and GPT.

6. The large-scale intelligent operation and maintenance method for time-series understanding of energy storage clusters according to claim 1, characterized in that, The supervised loss function uses the standard cross-entropy loss, and its calculation formula is as follows: In the formula, This indicates a supervised loss function. This represents the i-th input sample of the large language model, that is, the i-th semantic sentence in the set of semantic sentences. The parameters representing the large language model. express The corresponding battery operation description file, This means that given the input is and parameters Under the condition of large language model output The probability of M is given by M, which represents the total number of training samples. The formula for calculating the consensus loss function is as follows: In the formula, Let N represent the consensus loss function, and let N represent the number of large language models in the agent system. and These represent the battery runtime sequence description drafts generated by the j-th and k-th large language models, respectively. Distance metric representing the semantic similarity between two battery runtime sequence description drafts.

7. The large-scale intelligent operation and maintenance method for time-series understanding of energy storage clusters according to claim 1, characterized in that, The direct prompt query specifically includes: combining the battery operation description file with a natural language query to form prompt words, which are then directly input into the finely tuned and trained large language model agent. The agent provides the text in the form of natural language with the highest consensus as the generated answer, which is the operation and maintenance solution. The enhanced retrieval generation method specifically includes: decomposing historically generated battery operation description files into fragments and constructing a dedicated vector database; when a new query question is input, first retrieve the battery operation description file fragment most relevant to the current query question from the dedicated vector database, and then send this fragment and the query question together into the finely tuned and trained large language model agent to generate the answer, which is the operation and maintenance solution.

8. A large-scale intelligent operation and maintenance system for time-series understanding of energy storage clusters, characterized in that, It includes 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 intelligent operation and maintenance method for time-series understanding of energy storage clusters as described in any one of claims 1-7.

Images