Maintenance method and system of battery energy storage system based on large language model

Abstract

The invention discloses a battery energy storage system maintenance method and system based on a large language model, and belongs to the technical field of predictive maintenance of battery energy storage systems. Comprising the steps that multi-mode real-time operation data and historical charge-discharge cycle time sequence data of the battery energy storage system are acquired and stored, and the multi-mode real-time operation data comprise voltage, current, temperature and charge state data; inputting the multi-modal real-time operation data and the historical charge-discharge cycle time sequence data into a dynamic labeling engine formed by a large language model, and generating a battery health state semantic label with probability weight through multi-modal embedding, space-time attention extraction and expert routing distribution; and inputting the battery health state semantic label to a fault mode correlator, and outputting a battery health dynamic risk score and a corresponding maintenance action suggestion by the fault mode correlator based on a relational graph attention network embedded into a battery electrochemical degradation physical equation.

Description

Technical Field

[0001] This invention belongs to the field of predictive maintenance technology for battery energy storage systems, and relates to a maintenance method and system for battery energy storage systems based on a large language model. Background Technology

[0002] With the large-scale grid connection of renewable energy sources such as wind power and photovoltaics, the intermittent nature of their output poses a challenge to grid stability. As the core of power regulation, the reliability and full life cycle management of battery energy storage systems have become key bottlenecks restricting the large-scale application of renewable energy. Battery health status (SOH) monitoring and life prediction are the core links of energy storage operation and maintenance.

[0003] Traditional battery health monitoring mainly falls into two categories: one is rule-based fixed threshold monitoring systems, which trigger alarms by setting thresholds such as voltage deviation and capacity decay. Although simple to deploy, battery degradation is a nonlinear process involving multiple electrochemical reactions such as SEI film thickening, lithium deposition, and cathode structure collapse. Fixed thresholds cannot adapt to dynamic degradation patterns, and the false alarm rate exceeds 30% in the mid-life stage, which can easily lead to confusion in operation and maintenance decisions. The other is traditional statistical models such as support vector machines and random forests. Although they have basic nonlinear fitting capabilities, they cannot establish a correlation between degradation phenomena and deep electrochemical mechanisms due to limited feature extraction dimensions. The prediction error at the end of the life can reach more than 15%, and the sensitivity is difficult to meet the requirements.

[0004] Deep learning technology offers direction for monitoring upgrades. Temporal neural networks such as LSTM and GRU can uncover deep correlations in runtime data, improving SOH prediction accuracy. However, their "black box" nature leads to a lack of interpretability in the prediction results. Maintenance personnel cannot distinguish the root cause of anomalies and must resort to a full shutdown for investigation, significantly increasing maintenance costs and downtime. In recent years, LLM has shown potential in the field of industrial predictive maintenance, but it has obvious application limitations: First, it is mostly an offline analysis mode, lacking real-time data access and dynamic annotation capabilities, which does not match the real-time maintenance needs of energy storage systems; second, it still relies on static degradation labels, which cannot accurately characterize the nonlinear degradation process; third, multimodal data such as electrochemical parameters and operating logs are fragmented and lack physical correlation, making it difficult to identify multi-factor coupled degradation risks; fourth, general pre-trained models are not adapted to the battery failure mechanism, making it unable to accurately interpret specialized degradation scenarios and prone to prediction distortion.

[0005] In summary, existing technologies have significant shortcomings in dynamic degradation characterization, multimodal data fusion, model interpretability, and domain knowledge adaptation, making it difficult to meet the high reliability, high precision, and interpretable health management requirements of large-scale energy storage systems. Innovative technologies are urgently needed to overcome these bottlenecks. Summary of the Invention

[0006] The purpose of this invention is to address the significant shortcomings of existing technologies in dynamic degradation characterization, multimodal data fusion, model interpretability, and domain knowledge adaptation, and to provide a maintenance method and system for battery energy storage systems based on a large language model.

[0007] To achieve the above objectives, the present invention employs the following technical solution: In a first aspect, this invention discloses a maintenance method for a battery energy storage system based on a large language model, comprising: Acquire and store multimodal real-time operating data and historical charge-discharge cycle timing data of the battery energy storage system. The multimodal real-time operating data includes voltage, current, temperature and state of charge data. The multimodal real-time operating data and historical charge-discharge cycle time series data are input into a dynamic annotation engine composed of a large language model. Through multimodal embedding, spatiotemporal attention extraction and expert routing allocation, a semantic label of battery health status with probability weights is generated. The battery health status semantic label is input into the fault mode correlator. The fault mode correlator is based on the relational graph attention network embedded with the battery electrochemical degradation physical equation. It constructs a heterogeneous graph of label nodes, failure mode nodes and maintenance action nodes and completes multi-hop inference. It outputs a battery health dynamic risk score and corresponding maintenance action suggestions. The dynamic risk score and maintenance action suggestions are integrated into the battery management system control loop to trigger an adaptive maintenance strategy. At the same time, the maintenance effect data is fed back to the dynamic annotation engine and the fault mode correlator to achieve online model optimization.

[0008] Further improvements are made in the following aspects: The large language model is a sparse expert hybrid architecture large language model, which contains at least 12 domain expert modules for different battery failure modes. The multimodal embedding is achieved through the following formula:

[0009] in, For data from the i-th type of sensor; For trainable projection matrices; unified embedding dimension ; The spatiotemporal attention extraction is achieved through the following formula:

[0010] in, For query vector, As the key vector, It is a numerical vector. for Dimension size, It is a mask matrix; The expert route allocation is achieved through the following formula:

[0011] Each expert's output label distribution The final result is a weighted fusion:

[0012] Where j is the index of the j-th expert, e is a natural constant, and T represents the matrix transpose.

[0013] The semantic tags contain textual information describing the battery health status and corresponding confidence levels. The textual information at least covers health status descriptions of categories such as local overheating, accelerated SEI film growth, and increased risk of lithium plating on the negative electrode, and the tag confidence levels dynamically decay over time.

[0014] The battery electrochemical degradation physical equations include at least a lithium deposition kinetics model and the Arrhenius equation; the edge relationships in the heterogeneity graph include five predefined association types: electrochemical causality, temporal evolution, control dependence, data statistics, and domain knowledge; the dynamic risk scoring is achieved through the following formula:

[0015] The Readout function performs mean pooling of the node states. For the weight vector, For activation function, This is the node state matrix.

[0016] The adaptive maintenance strategy includes at least one of reducing charging current, initiating impedance spectrum diagnosis, and activating the cooling system; and the semantic tag can be directly mapped to the energy management system alarm code, while automatically generating maintenance work orders that conform to the IEC62443 standard.

[0017] The node types of the heterogeneous graph include label nodes, failure nodes, and action nodes; the edge relationships of the heterogeneous graph include electrochemical causality, temporal evolution, control dependence, data statistics, and domain knowledge.

[0018] The multimodal real-time operating data of the battery energy storage system is obtained through a sensor network, and the acquisition accuracy meets the following requirements: voltage acquisition accuracy ±1mV, current acquisition accuracy ±10mA, temperature acquisition accuracy ±0.5℃, and the time resolution of the time series database is not less than 1 second.

[0019] Secondly, this invention discloses a maintenance system for a battery energy storage system based on a large language model, comprising: The data acquisition unit is used to acquire and store multimodal real-time operating data and historical charge-discharge cycle timing data of the battery energy storage system. The multimodal real-time operating data includes voltage, current, temperature and state of charge data. The dynamic probabilistic annotation unit is used to input the multimodal real-time running data and historical charge-discharge cycle time series data into the dynamic annotation engine composed of a large language model, and generate a battery health status semantic label with probability weight through multimodal embedding, spatiotemporal attention extraction and expert routing allocation. The fault mode association and risk assessment unit is used to input the semantic label of the battery health status into the fault mode association unit. The fault mode association unit is based on the relational graph attention network embedded with the physical equation of battery electrochemical degradation, constructs a heterogeneous graph of label nodes, failure mode nodes and maintenance action nodes and completes multi-hop inference, and outputs the battery health dynamic risk score and corresponding maintenance action suggestions. The closed-loop maintenance control unit is used to integrate the dynamic risk score and maintenance action suggestions into the battery management system control loop, trigger the adaptive maintenance strategy, and at the same time feed back the maintenance effect data to the dynamic annotation engine and fault mode correlator to realize online model optimization.

[0020] Thirdly, the present invention discloses an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the aforementioned maintenance system for a battery energy storage system based on a large language model.

[0021] Fourthly, the present invention discloses a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned maintenance system for a battery energy storage system based on a large language model.

[0022] Compared with the prior art, the present invention has the following beneficial effects: This invention discloses a maintenance method for battery energy storage systems based on a large language model. On the one hand, this method integrates multimodal real-time operating data and historical charge-discharge time-series data through a dynamic annotation engine. By leveraging multimodal embedding, spatiotemporal attention extraction, and expert routing allocation, it generates semantic tags with probability weights, overcoming the limitations of traditional fixed threshold monitoring. This method can accurately capture the nonlinear electrochemical degradation process of the battery and avoid the alarm lag or false alarm problems of traditional methods. On the other hand, the fault mode correlator embeds the physical equation of battery electrochemical degradation, which can establish the correlation between semantic tags and underlying failure mechanisms, enabling root cause identification of battery faults. Combined with the dynamic risk score output by multi-hop inference, it can predict the trend of battery health deterioration in advance, significantly extending the fault warning window period compared to traditional solutions and reserving sufficient time for operation and maintenance. This method directly integrates dynamic risk scoring and maintenance action recommendations into the battery management system control loop, which can automatically trigger adaptive maintenance strategies, replacing the traditional manual judgment and operation mode. This not only significantly improves the timeliness of maintenance response, but also enables differentiated maintenance based on the actual health status of the battery, avoiding resource waste caused by over-maintenance or safety hazards caused by under-maintenance. At the same time, maintenance effect data can be fed back to the dynamic labeling engine and fault mode correlator to complete online model optimization, forming a closed-loop management of "monitoring-decision-maintenance-feedback-optimization". This allows the maintenance strategy to dynamically iterate with the battery's operating conditions and degradation status, continuously improving the adaptability and effectiveness of the maintenance plan. Traditional deep learning monitoring models suffer from the drawback of being "black boxes," while the semantic labels for battery health status generated by this method are descriptive texts that can clearly present the specific types and probabilities of battery health problems. The heterogeneous graph reasoning process of the fault mode correlator can also clearly show the correlation logic between health labels, failure modes, and maintenance actions, allowing maintenance personnel to trace the basis for risk assessment and improve the credibility of maintenance decisions. In addition, standardized semantic labels can achieve a unified expression of health status across different battery modules and energy storage sites, breaking down the barriers to the expression of health data at different sites and on different devices. This facilitates cross-site maintenance coordination and experience reuse, improving the standardization level of overall maintenance management. The expert routing allocation mechanism of the dynamic annotation engine can configure dedicated processing modules for different battery failure modes, adapting to the maintenance needs of various chemical systems such as lithium iron phosphate and ternary lithium batteries. It can also quickly respond to the monitoring needs of new battery failure modes by adding expert modules. At the same time, the multimodal data processing architecture is compatible with the types of sensor data added in the future, and the model design that embeds physical equations reduces the dependence on pure data fitting, enabling the model to have stronger generalization ability in different energy storage application scenarios, and providing flexible architectural support for subsequent technology iteration and functional expansion. Attached Figure Description

[0023] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart of a maintenance method for a battery energy storage system based on a large language model, as described in an embodiment of the present invention. Figure 2 This is an overall architecture diagram of the maintenance system of the battery energy storage system based on a large language model in an embodiment of the present invention; Figure 3 This is a detailed architecture of the predictive maintenance system for the battery energy storage system based on a large language model, as described in this embodiment of the invention. Figure 4 This describes the dynamic annotation engine workflow of the maintenance system for a battery energy storage system based on a large language model, as described in this embodiment of the invention. Figure 5 This is an example of a fault mode association diagram for the maintenance system of a battery energy storage system based on a large language model, as described in an embodiment of the present invention. Figure 6 This is a block diagram of a battery energy storage system based on a large language model, as described in this embodiment of the invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0026] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0027] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0028] The present invention will now be described in further detail with reference to the accompanying drawings: See Figure 1This invention discloses a maintenance method for a battery energy storage system based on a large language model, comprising: S1, acquire and store multimodal real-time operating data and historical charge-discharge cycle timing data of the battery energy storage system, wherein the multimodal real-time operating data includes voltage, current, temperature and state of charge data; S2, the multimodal real-time running data and historical charge-discharge cycle time series data are input into the dynamic annotation engine composed of a large language model, and a battery health status semantic label with probability weight is generated through multimodal embedding, spatiotemporal attention extraction and expert routing allocation. S3, input the battery health status semantic label to the fault mode correlator. The fault mode correlator is based on the relational graph attention network embedded with the battery electrochemical degradation physical equation, constructs a heterogeneous graph of label nodes, failure mode nodes and maintenance action nodes and completes multi-hop inference, and outputs the battery health dynamic risk score and corresponding maintenance action suggestions. S4. The dynamic risk score and maintenance action suggestions are integrated into the battery management system control loop to trigger the adaptive maintenance strategy. At the same time, the maintenance effect data is fed back to the dynamic annotation engine and the fault mode correlator to realize online model optimization.

[0029] This invention, in addition to advancing the mean time to failure (MTBF) by 3-5 charge-discharge cycles and reducing unplanned downtime losses by over 30%, can also accurately pinpoint the core degradation causes of batteries at different cycle stages (such as abnormal SEI film thickening and lithium plating on the negative electrode) based on refined semantic tags generated by a dynamic labeling engine. This provides data support for optimizing battery positive and negative electrode material formulations and improving cell manufacturing processes, thereby enhancing the cycle life and safety performance of new batteries. Simultaneously, by using the risk evolution path output by the fault mode correlator, it can optimize the reliability testing scheme before battery delivery, reducing testing costs and cycles, and improving product qualification rates. Besides achieving standardized descriptions of cross-site health status and improving operational efficiency through semantic tags, it can also rely on a unified health status semantic system to achieve centralized monitoring and hierarchical management of the health status of battery energy storage systems at multiple sites. This facilitates the grid side's overall coordination of charging and discharging strategies for various energy storage sites, ensuring the stability and accuracy of grid power regulation. Furthermore, standardized semantic tags enable interoperability and mutual recognition of health data from different manufacturers and models of batteries, breaking down data barriers and providing a data foundation for the integration and optimized allocation of grid-side energy storage resources. In addition to providing a scalable MoE architecture and supporting rapid model iteration for novel failure modes, this method embeds a relational graph attention network of the battery electrochemical degradation physical equations. This provides a dual-driven modeling paradigm of "data + physical mechanism" for algorithm development, reducing the algorithm's dependence on pure data fitting and improving the model's generalization ability. Simultaneously, the online model optimization mechanism can continuously iterate the algorithm based on actual operation and maintenance feedback, reducing the development cycle for adapting the algorithm to different application scenarios. Furthermore, the multimodal processing framework of the dynamic annotation engine is compatible with future additions of sensor data types, reserving ample space for algorithm functional expansion. On one hand, the closed-loop maintenance control system can automate the execution of maintenance actions from risk warning to maintenance actions, significantly reducing manual intervention and lowering the workload and professional threshold for operation and maintenance personnel. On the other hand, dynamic risk scoring can quantitatively assess the battery health status, facilitating the development of differentiated maintenance plans by the operation and maintenance team, avoiding over-maintenance or under-maintenance, and reducing overall operation and maintenance costs. Moreover, the feedback mechanism of maintenance effect data can continuously verify the effectiveness of maintenance strategies, gradually forming an optimal operation and maintenance solution adapted to the specific operating conditions of the power station. For batteries intended for secondary use, this method can generate unique secondary life assessment labels to accurately determine the remaining value and applicable scenarios of retired batteries, thereby improving the economic efficiency and safety of secondary use. For battery data throughout its entire life cycle, standardized semantic tags can be used to achieve full-process data traceability from production and use to retirement, providing data support for battery recycling and liability determination, and helping to build a complete green closed-loop battery industry chain.

[0030] The present invention will be further described below with reference to specific embodiments: Step 1: Acquire and store multimodal real-time operating data and historical charge-discharge cycle timing data of the battery energy storage system. The multimodal real-time operating data includes voltage, current, temperature and state of charge data. The multimodal real-time operating data of the battery energy storage system is obtained through a sensor network, and the acquisition accuracy meets the following requirements: voltage acquisition accuracy ±1mV, current acquisition accuracy ±10mA, temperature acquisition accuracy ±0.5℃, and the time resolution of the time series database is not less than 1 second.

[0031] Step 2: Input the multimodal real-time running data and historical charge-discharge cycle time series data into the dynamic annotation engine composed of a large language model. Through multimodal embedding, spatiotemporal attention extraction and expert routing allocation, generate semantic labels for battery health status with probability weights. The large language model is a sparse expert hybrid architecture large language model, which contains at least 12 domain expert modules for different battery failure modes. The multimodal embedding is achieved through the following formula:

[0032] in, For data from the i-th type of sensor; For trainable projection matrices; unified embedding dimension ; The spatiotemporal attention extraction is achieved through the following formula:

[0033] in, For query vector, As the key vector, It is a numerical vector. for Dimension size, It is a mask matrix; The expert route allocation is achieved through the following formula:

[0034] Each expert's output label distribution The final result is a weighted fusion:

[0035] Where j is the index of the j-th expert, e is a natural constant, and T represents the matrix transpose.

[0036] Step 3: Input the semantic label of battery health status into the fault mode correlator. The fault mode correlator is based on the relational graph attention network embedded with the physical equation of battery electrochemical degradation. It constructs a heterogeneous graph of label nodes, failure mode nodes and maintenance action nodes and completes multi-hop inference. It outputs the dynamic risk score of battery health and corresponding maintenance action suggestions. The node types of the heterogeneous graph include label nodes, failure nodes, and action nodes; the edge relationships of the heterogeneous graph include electrochemical causality, temporal evolution, control dependence, data statistics, and domain knowledge.

[0037] The semantic tags contain textual information describing the battery health status and corresponding confidence levels. The textual information at least covers health status descriptions of categories such as local overheating, accelerated SEI film growth, and increased risk of lithium plating on the negative electrode, and the tag confidence levels dynamically decay over time.

[0038] The battery electrochemical degradation physical equations include at least a lithium deposition kinetics model and the Arrhenius equation; the edge relationships in the heterogeneity graph include five predefined association types: electrochemical causality, temporal evolution, control dependence, data statistics, and domain knowledge; the dynamic risk scoring is achieved through the following formula:

[0039] The Readout function performs mean pooling of the node states. For the weight vector, For activation function, This is the node state matrix.

[0040] Step four: Integrate the dynamic risk score and maintenance action suggestions into the battery management system control loop to trigger the adaptive maintenance strategy. At the same time, feed back the maintenance effect data to the dynamic annotation engine and fault mode correlator to achieve online model optimization.

[0041] The adaptive maintenance strategy includes at least one of reducing charging current, initiating impedance spectrum diagnosis, and activating the cooling system; and the semantic tag can be directly mapped to the energy management system alarm code, while automatically generating maintenance work orders that conform to the IEC62443 standard.

[0042] The working principle of this invention is as follows: This invention proposes a predictive maintenance framework for battery energy storage systems based on Large Language Model (LLM). Its core innovation lies in transforming traditional rule-based battery health monitoring into a dynamic semantic reasoning process. This framework overcomes the limitations of existing systems through three key technological breakthroughs: Dynamic probabilistic labeling: Employs a finely tuned LLM architecture to parse multi-source sensor data in real time and generate semantic labels with probability weights (such as "high temperature causes SEI to grow faster, confidence level 82%)", replacing the traditional binary alarm mechanism. Electrochemical knowledge integration: Embedding battery degradation physics equations (such as lithium deposition kinetic models) into graph neural networks to establish an interpretable connection between semantic tags and underlying electrochemical processes; Closed-loop control integration: Risk scores are directly injected into the BMS control loop to achieve end-to-end automation from semantic understanding to maintenance actions.

[0043] The technical value of this invention is reflected in: (1) For battery manufacturers: the average fault warning time can be advanced by 3-5 charge-discharge cycles, reducing unplanned downtime losses by more than 30%; (2) For grid operators: cross-site health status standardization description is achieved through semantic tags, improving operation and maintenance efficiency; (3) For algorithm developers: Provides a scalable MoE architecture to support rapid model iteration for new failure modes.

[0044] like Figure 2 As shown, the method in this invention is implemented through the following system architecture: Battery energy storage system: an infrastructure including battery packs, sensor networks, and power conversion devices; Battery pack: a lithium-ion battery stack providing energy storage capacity; Sensors and monitoring devices: collecting real-time parameters such as voltage, current, and temperature; Power conversion system: realizing DC-AC bidirectional conversion; Predictive maintenance system: the core innovative module of this invention; Dynamic labeling engine: converting sensor data into probabilistic semantic tags; Fault mode correlator: analyzing the causal relationships between tags and calculating risk scores; Battery management system: a traditional control unit; BMS control logic: adjusting operating parameters according to risk scores; Data flow: Sensor data → Dynamic labeling engine → Fault mode correlator → BMS control → Power conversion system Innovation: For the first time, LLM semantic understanding is directly coupled with the traditional BMS control loop; continuous state assessment is achieved through probabilistic tags, replacing the traditional threshold alarm mechanism.

[0045] This system comprises the following core modules and their interaction relationships: Data Acquisition Layer: High-precision sensor network: voltage acquisition (Accuracy ±1mV), Current (±10mA), temperature (±0.5°C) and impedance spectral data; Time-series database: Stores historical charge-discharge cycle data with a time resolution of up to 1 second.

[0046] Dynamic annotation engine (DLE): Input: Real-time sensor data stream (k is the length of the sliding window); Output: Probability label set ,in For semantic tags (such as "local overheating"), This represents the corresponding probability.

[0047] Pattern Correlator (FMC): Built on RGAT, nodes represent labels / failure modes, and edge weights reflect the strength of causal relationships; Output dynamic risk score And recommended actions (such as "immediately reduce charging rate").

[0048] BMS Interface Layer: Protocol conversion module: converts semantic instructions into CAN bus control commands; Feedback channel: Record maintenance results for online model learning.

[0049] Key data flow: Raw data → DLE performs feature extraction and semantic mapping → FMC performs risk reasoning → BMS executes control strategies → Performance data is fed back to model optimization.

[0050] The core technology implementation method is as follows: 1. Dynamic Annotation Engine (DLE) The processing flow of the GPT-4 variant using a sparse MoE architecture is as follows: (1) Multimodal embedding: Sensor data is converted into a unified representation through a modality-specific encoder:

[0051] in: : Data from the i-th type of sensor (e.g., voltage sequence) Trainable projection matrix Unified Embedding Dimension (2) Spatiotemporal attention: An improved Transformer layer is used to capture cross-modal and temporal dependencies:

[0052] The mask matrix accomplish: Spatial attention: Limiting the depth of interaction between different sensor types Time attention: Forced focus on the last 30 time steps (3) Expert routing: Features are assigned to 12 domain experts (such as Expert_thermal, which specializes in handling thermal runaway) through a gating mechanism:

[0053] Each expert's output label distribution The final result is a weighted fusion:

[0054] 2. Fault Mode Correlator (FMC) Construct a heterogeneous graph containing three types of nodes: (1) Node type: Tag node: from DLE

[0055] Failure points: such as "lithium dendrite growth" Action nodes: such as "reduce charging current" (2) Edge relationships: Five types of edges are defined to reflect different relationships: Electrochemical causality (e.g., "high temperature" → "SEI thickening") Time-series evolution (e.g., "early stage of lithium plating" → "severe lithium plating") Controlling dependencies (e.g., "capacity decay" → "equalization charging") Data statistics (co-occurrence frequency) Domain knowledge (related to literature verification) (3) Message passing: Graph reasoning is performed using R-GAT:

[0056] in: Relationship-specific attention weights : Transformation matrix corresponding to relation type r (4) Risk scoring: Calculated at the final level:

[0057] The Readout function performs mean pooling of node states.

[0058] Actual implementation process of energy storage power stations: System extensions: Deploy a cloud-edge collaborative architecture, with each battery compartment equipped with an edge computing node to run local DLE, and the central cloud platform aggregating multi-node data to train the global FMC; Add an acoustic fingerprint sensor to monitor the rate of gas generation from electrolyte decomposition.

[0059] Multi-dimensional tags: Generate comparison labels across battery stacks (e.g., "inter-cluster capacity deviation exceeds 15%"); Long-term degradation trend labels (such as an average annual capacity decay rate of 2.3%) are generated by combining calendar aging data.

[0060] Operations and Maintenance Integration: Directly map semantic tags to EMS system alarm codes; Automatically generate maintenance work orders that conform to the IEC62443 standard.

[0061] See Figure 3 Dynamic labeling engine: the core of multimodal data processing; Data preprocessing: standardization and windowing of raw sensor data; Large language model: Transformer model using MoE architecture; Label generation: outputs semantic labels with confidence (such as SEI growth anomaly); Fault mode correlator: risk reasoning module; Graph neural network: constructs a heterogeneous graph of label-fault-action; Risk score calculation: generates 0-1 risk values ​​based on node states; Key interaction: preprocessed time series data is input into LLM for feature extraction; The generated labels are used as input to GNN nodes, and the final risk score triggers the BMS control strategy.

[0062] Innovation: The system employs a relational graph attention network to model electrochemical causal relationships, enabling end-to-end interpretable reasoning from raw data to control commands.

[0063] See Figure 4 Input layer: raw data processing; sensor data: multi-source signals such as voltage, current, and temperature; sliding window segmentation: generating time-series data segments (default 30-second window); processing layer: feature extraction and fusion; multimodal coding: independent embedding of different types of sensor data; spatiotemporal attention: capturing cross-modal and time-series dependencies; expert routing: assigning to specific domain experts based on features; output layer: semantic results; probabilistic labels: weighted fusion results of the outputs of each expert; Innovation: The spatiotemporal attention mechanism simultaneously handles the relationships between sensors and the time dimension; expert division of labor enables specialized identification of different failure modes.

[0064] See Figure 5 Tag nodes: Semantic tags generated by DLE; High temperature alarm: Temperature exceeds normal fluctuation range; SEI thickening: Accelerated growth of solid electrolyte interface layer; Voltage fluctuation: Abnormal oscillation of charge and discharge platform; Lithium dendrite risk: Increased possibility of lithium plating on negative electrode; Fault nodes: Potential failure modes; Thermal runaway: Chain exothermic reaction; Internal short circuit: Direct contact between positive and negative electrodes; Capacity decay: Continuous decrease in available energy; Action nodes: Maintenance strategies; Reduce charging current: Relieve electrode stress; Activate cooling system: Suppress temperature rise.

[0065] Innovation: Explicit modeling of electrochemical degradation chain reaction pathways; supports multi-hop inference (e.g., high temperature → SEI thickening → capacity decay → current reduction).

[0066] This invention transforms sensor data into semantic tags in real time (e.g., increased risk of lithium plating at the negative electrode), breaking through the static threshold limitation; it uses graph attention networks to establish a mapping relationship between electrochemical parameters and text tags, solving the problem of multimodal fragmentation; and it adopts a fine-tuning strategy enhanced by electrochemical knowledge (e.g., injecting Arrhenius equation priors) to improve domain specificity.

[0067] See Figure 6 This invention also discloses a maintenance system for a battery energy storage system based on a large language model, comprising: The data acquisition unit is used to acquire and store multimodal real-time operating data and historical charge-discharge cycle timing data of the battery energy storage system. The multimodal real-time operating data includes voltage, current, temperature and state of charge data. The dynamic probabilistic annotation unit is used to input the multimodal real-time running data and historical charge-discharge cycle time series data into the dynamic annotation engine composed of a large language model, and generate a battery health status semantic label with probability weight through multimodal embedding, spatiotemporal attention extraction and expert routing allocation. The fault mode association and risk assessment unit is used to input the semantic label of the battery health status into the fault mode association unit. The fault mode association unit is based on the relational graph attention network embedded with the physical equation of battery electrochemical degradation, constructs a heterogeneous graph of label nodes, failure mode nodes and maintenance action nodes and completes multi-hop inference, and outputs the battery health dynamic risk score and corresponding maintenance action suggestions. The closed-loop maintenance control unit is used to integrate the dynamic risk score and maintenance action suggestions into the battery management system control loop, trigger the adaptive maintenance strategy, and at the same time feed back the maintenance effect data to the dynamic annotation engine and fault mode correlator to realize online model optimization.

[0068] A third objective of this invention is to provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the maintenance method of the battery energy storage system based on a large language model.

[0069] The maintenance method for the battery energy storage system based on a large language model includes the following steps: Acquire and store multimodal real-time operating data and historical charge-discharge cycle timing data of the battery energy storage system. The multimodal real-time operating data includes voltage, current, temperature and state of charge data. The multimodal real-time operating data and historical charge-discharge cycle time series data are input into a dynamic annotation engine composed of a large language model. Through multimodal embedding, spatiotemporal attention extraction and expert routing allocation, a semantic label of battery health status with probability weights is generated. The battery health status semantic label is input into the fault mode correlator. The fault mode correlator is based on the relational graph attention network embedded with the battery electrochemical degradation physical equation. It constructs a heterogeneous graph of label nodes, failure mode nodes and maintenance action nodes and completes multi-hop inference. It outputs a battery health dynamic risk score and corresponding maintenance action suggestions. The dynamic risk score and maintenance action suggestions are integrated into the battery management system control loop to trigger an adaptive maintenance strategy. At the same time, the maintenance effect data is fed back to the dynamic annotation engine and the fault mode correlator to achieve online model optimization.

[0070] A fourth objective of this invention is to provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the maintenance method for the battery energy storage system based on a large language model.

[0071] The maintenance method for the battery energy storage system based on a large language model includes the following steps: Acquire and store multimodal real-time operating data and historical charge-discharge cycle timing data of the battery energy storage system. The multimodal real-time operating data includes voltage, current, temperature and state of charge data. The multimodal real-time operating data and historical charge-discharge cycle time series data are input into a dynamic annotation engine composed of a large language model. Through multimodal embedding, spatiotemporal attention extraction and expert routing allocation, a semantic label of battery health status with probability weights is generated. The battery health status semantic label is input into the fault mode correlator. The fault mode correlator is based on the relational graph attention network embedded with the battery electrochemical degradation physical equation. It constructs a heterogeneous graph of label nodes, failure mode nodes and maintenance action nodes and completes multi-hop inference. It outputs a battery health dynamic risk score and corresponding maintenance action suggestions. The dynamic risk score and maintenance action suggestions are integrated into the battery management system control loop to trigger an adaptive maintenance strategy. At the same time, the maintenance effect data is fed back to the dynamic annotation engine and the fault mode correlator to achieve online model optimization.

[0072] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0073] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0074] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0075] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0076] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A maintenance method for a battery energy storage system based on a large language model, characterized in that, include: Acquire and store multimodal real-time operating data and historical charge-discharge cycle timing data of the battery energy storage system. The multimodal real-time operating data includes voltage, current, temperature and state of charge data. The multimodal real-time operating data and historical charge-discharge cycle time series data are input into a dynamic annotation engine composed of a large language model. Through multimodal embedding, spatiotemporal attention extraction and expert routing allocation, a semantic label of battery health status with probability weights is generated. The battery health status semantic label is input into the fault mode correlator. The fault mode correlator is based on the relational graph attention network embedded with the battery electrochemical degradation physical equation. It constructs a heterogeneous graph of label nodes, failure mode nodes and maintenance action nodes and completes multi-hop inference. It outputs a battery health dynamic risk score and corresponding maintenance action suggestions. The dynamic risk score and maintenance action suggestions are integrated into the battery management system control loop to trigger an adaptive maintenance strategy. At the same time, the maintenance effect data is fed back to the dynamic annotation engine and the fault mode correlator to achieve online model optimization.

2. The maintenance method for a battery energy storage system based on a large language model according to claim 1, characterized in that, The large language model is a sparse expert hybrid architecture large language model, which contains at least 12 domain expert modules for different battery failure modes. The multimodal embedding is achieved through the following formula: in, For data from the i-th type of sensor; For trainable projection matrices; unified embedding dimension ; The spatiotemporal attention extraction is achieved through the following formula: in, For query vector, As the key vector, It is a numerical vector. for Dimension size, It is a mask matrix; The expert route allocation is achieved through the following formula: Each expert's output label distribution The final result is a weighted fusion: Where j is the index of the j-th expert, e is a natural constant, and T represents the matrix transpose.

3. The maintenance method for a battery energy storage system based on a large language model according to claim 1, characterized in that, The semantic tags contain textual information describing the battery health status and corresponding confidence levels. The textual information at least covers health status descriptions of categories such as local overheating, accelerated SEI film growth, and increased risk of lithium plating on the negative electrode, and the tag confidence levels dynamically decay over time.

4. The maintenance method for a battery energy storage system based on a large language model according to claim 1, characterized in that, The battery electrochemical degradation physical equations include at least a lithium deposition kinetics model and the Arrhenius equation; the edge relationships in the heterogeneity graph include five predefined association types: electrochemical causality, temporal evolution, control dependence, data statistics, and domain knowledge; the dynamic risk scoring is achieved through the following formula: The Readout function performs mean pooling of the node states. For the weight vector, For activation function, This is the node state matrix.

5. The maintenance method for a battery energy storage system based on a large language model according to claim 1, characterized in that, The adaptive maintenance strategy includes at least one of reducing charging current, initiating impedance spectrum diagnosis, and activating the cooling system; and the semantic tag can be directly mapped to the energy management system alarm code, while automatically generating maintenance work orders that conform to the IEC62443 standard.

6. The maintenance method for a battery energy storage system based on a large language model according to claim 1, characterized in that, The node types of the heterogeneous graph include label nodes, failure nodes, and action nodes; the edge relationships of the heterogeneous graph include electrochemical causality, temporal evolution, control dependence, data statistics, and domain knowledge.

7. The maintenance method for a battery energy storage system based on a large language model according to claim 1, characterized in that, The multimodal real-time operating data of the battery energy storage system is obtained through a sensor network, and the acquisition accuracy meets the following requirements: voltage acquisition accuracy ±1mV, current acquisition accuracy ±10mA, temperature acquisition accuracy ±0.5℃, and the time resolution of the time series database is not less than 1 second.

8. A maintenance system for a battery energy storage system based on a large language model, characterized in that, include: The data acquisition unit is used to acquire and store multimodal real-time operating data and historical charge-discharge cycle timing data of the battery energy storage system. The multimodal real-time operating data includes voltage, current, temperature and state of charge data. The dynamic probabilistic annotation unit is used to input the multimodal real-time running data and historical charge-discharge cycle time series data into the dynamic annotation engine composed of a large language model, and generate a battery health status semantic label with probability weight through multimodal embedding, spatiotemporal attention extraction and expert routing allocation. The fault mode association and risk assessment unit is used to input the semantic label of the battery health status into the fault mode association unit. The fault mode association unit is based on the relational graph attention network embedded with the physical equation of battery electrochemical degradation, constructs a heterogeneous graph of label nodes, failure mode nodes and maintenance action nodes and completes multi-hop inference, and outputs the battery health dynamic risk score and corresponding maintenance action suggestions. The closed-loop maintenance control unit is used to integrate the dynamic risk score and maintenance action suggestions into the battery management system control loop, trigger the adaptive maintenance strategy, and at the same time feed back the maintenance effect data to the dynamic annotation engine and fault mode correlator to realize online model optimization.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements a maintenance system for a battery energy storage system based on a large language model as described in any one of claims 1-7.

10. A computer-readable storage medium storing a computer program that, when executed by a processor, implements the maintenance system for the battery energy storage system based on a large language model as described in any one of claims 1-7.

Images