Response method for energy storage system and electronic device

Abstract

The embodiment of the application provides a response method for an energy storage system and an electronic device, and relates to the technical field of large language models. The response method comprises: obtaining multi-dimensional text data in the energy storage field, and constructing a training data set comprising a plurality of training samples based on the multi-dimensional text data, each training sample comprising: input text and output text. The training data set is used to train a base model to obtain an energy storage special model; for a received user question, a plurality of target knowledge units related to the user question are determined in a plurality of knowledge units divided by the multi-dimensional text data, and each knowledge unit has labeled hierarchical path information; based on the user question and the plurality of target knowledge units, response content corresponding to the user question output by the energy storage special model is obtained.

Description

Technical Field

[0001] This invention relates to the field of large language model technology, and more specifically to a response method and electronic device for energy storage systems. Background Technology

[0002] With the large-scale deployment of electrochemical energy storage systems, operation and maintenance data exhibits characteristics such as multi-source heterogeneity, a high proportion of unstructured data, and frequent knowledge updates. In the field of intelligent operation and maintenance of existing electrochemical energy storage systems, related research and practice mainly focus on rule-based or data-driven automated analysis methods. Some solutions rely on the data interface between the Supervisory Control and Data Acquisition (SCADA) system and the operation and maintenance platform, using thresholds or expert rules for alarm screening and trend prediction. While these methods can achieve basic anomaly detection, they lack the ability to understand the logical connections and semantics of complex multi-causal events, still requiring significant manual intervention for judgment and confirmation. Another type of research attempts to use machine learning or deep learning models to perform health prediction and fault diagnosis on numerical data such as cell parameters, ambient temperature, and charging and discharging behavior. This can improve the automation level of analysis to some extent, but the models are mostly limited to specific equipment or operating conditions, making it difficult to generalize to actual operating scenarios involving multiple manufacturers and systems. They also cannot handle unstructured text knowledge, such as maintenance records, operating procedures, and standard clauses. The above methods are often used for local operation and maintenance analysis and early warning of energy storage systems, but they usually cannot form a complete systemic intelligent operation and maintenance framework.

[0003] It is evident that traditional operation and maintenance methods largely rely on structured threshold alarms or single data-driven models, making it difficult to effectively integrate unstructured textual knowledge such as standards, maintenance records, and incident reports. This results in insufficient causal correlation analysis capabilities for complex faults. Furthermore, when general-purpose language models are directly applied to energy storage scenarios, the lack of domain knowledge alignment can easily lead to misunderstandings of technical terms and factual illusions. In addition, full parameter fine-tuning is computationally expensive and prone to catastrophic amnesia, failing to meet the needs of enterprise-level private deployments and continuous iteration. Therefore, constructing an intelligent operation and maintenance architecture that integrates multi-dimensional textual data retrieval and efficient adaptation of domain parameters has become a key technology for overcoming existing knowledge utilization bottlenecks and improving system security governance capabilities. Summary of the Invention

[0004] The purpose of this invention is to provide a response method and electronic device for energy storage systems, which deeply integrates multi-dimensional text data fusion, retrieval enhancement generation optimization and parameter fine-tuning, and constructs a complete technical system from knowledge extraction and semantic retrieval to domain reasoning and decision generation.

[0005] To achieve the above objectives, the present invention provides a response method for energy storage systems, comprising: Acquire multi-dimensional text data in the field of energy storage, and based on the multi-dimensional text data, construct a training dataset including multiple training samples, each training sample including: input text and output text; The basic model is trained based on the training dataset to obtain a dedicated energy storage model; For a received user question, multiple target knowledge units related to the user question are determined from multiple knowledge units obtained by dividing the multi-dimensional text data. Each knowledge unit has been labeled with its hierarchical path information. Based on the user's question and the multiple target knowledge units, the response content corresponding to the user's question is obtained from the output of the energy storage-specific model.

[0006] The present invention also provides an electronic device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the response method for energy storage systems as described above.

[0007] The present invention also provides a computer-readable storage medium, which is a non-volatile or non-transient storage medium, on which a computer program is stored, and which, when executed by a processor, performs the above-described response method for energy storage systems.

[0008] In one embodiment, based on the multi-dimensional text data, a training dataset comprising multiple training samples is constructed, including: The multi-dimensional text data is semantically normalized and text normalized to obtain a normalized text sequence, and the normalized text sequence is segmented to obtain multiple initial corpus fragments; Based on semantic analysis and knowledge extraction of each initial corpus fragment, each initial corpus fragment is converted into an input-output pair, each input-output pair including: input text and output text; For each input-output pair, if the input-output pair involves risk indicators of the energy storage system, then based on the risk coupling relationship network of the energy storage system, the risk indicators in the output text of the input-output pair are associated according to the coupling relationship; The obtained input-output pairs are used as training samples to form a training dataset.

[0009] In one embodiment, a base model is trained based on the training dataset to obtain an energy storage-specific model, including: A LoRA module is added between each key linear transformation layer of the base model and its connected feedforward layer to obtain an intermediate model; The intermediate model is trained using training samples from the training dataset, and the weight parameters of each key linear transformation layer in the base model are updated to form a dedicated energy storage model. During the training of the intermediate model, the low-rank decomposition matrix corresponding to each LoRA module is updated, and the weight parameters of the key linear transformation layer connected to the LoRA module are updated based on the low-rank decomposition matrix corresponding to each LoRA module.

[0010] In one embodiment, the method further includes: The basic model is trained based on the aforementioned multi-dimensional text data; The basic model is trained based on the aforementioned training dataset to obtain a dedicated energy storage model, including: The trained base model is trained using the training dataset to obtain a dedicated energy storage model.

[0011] In one embodiment, the specific method for dividing the multi-dimensional text data into multiple knowledge units is as follows: Based on preset segmentation parameters and semantic similarity thresholds, the multi-dimensional text data is divided into multiple knowledge units, and the hierarchical path information of each knowledge unit is labeled through hierarchical identification; wherein the segmentation parameters include: the length of a single segment, the maximum text span of a segment, and the overlap ratio of adjacent segments.

[0012] In one embodiment, for a received user question, determining multiple target knowledge units related to the user question from multiple knowledge units obtained by segmenting the multi-dimensional text data includes: Multiple candidate knowledge units similar to the user's question are retrieved from the multiple knowledge units obtained by dividing the multi-dimensional text data. The relevance scores between the user's question and each of the candidate knowledge units are calculated respectively, and the target knowledge units are selected from the candidate knowledge units based on the relevance scores corresponding to the candidate knowledge units.

[0013] In one embodiment, based on the user question and the plurality of target knowledge units, the response content corresponding to the user question output by the energy storage-specific model is obtained, including: The user's question and the multiple target knowledge units are respectively embedded into preset prompt word templates to form enhanced input information; The enhanced input information is input into the energy storage-specific model, and the energy storage-specific model outputs the response content corresponding to the user's question.

[0014] In one embodiment, the multi-dimensional text data includes at least one of the following: standard data, operation and safety management data, scientific research data, and operation monitoring data.

[0015] In one embodiment, the response content includes: hierarchical path information for each of the target knowledge units.

[0016] In one embodiment, if the user question is an operation and maintenance question, then based on the user question, the multiple target knowledge units, and the risk assessment results of the energy storage system, the response content corresponding to the user question output by the energy storage-specific model is obtained. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the response method for energy storage systems in the first embodiment of the present invention; Figure 2 This is a schematic diagram of the knowledge base constructed in the first embodiment of the present invention; Figure 3 This is a schematic diagram of the training loss curves of experimental groups under different experimental conditions in the first embodiment of the present invention; Figure 4 This is a schematic diagram of the hierarchical path in the first embodiment of the present invention; Figure 5 This is a schematic diagram of the three target knowledge units retrieved in response to a user's question in the first embodiment of the present invention; Figure 6 A schematic diagram of the technical route for responding to user maintenance inquiries in the first embodiment of the present invention; Figure 7 A schematic diagram illustrating the response to a user's knowledge-related question in the first embodiment of the present invention; Figure 8 This is a schematic diagram illustrating the response to a user's operation and maintenance inquiry in the first embodiment of the present invention. Detailed Implementation

[0018] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings to provide a clearer understanding of the purpose, features, and advantages of the present invention. It should be understood that the embodiments shown in the drawings are not intended to limit the scope of the present invention, but are merely illustrative of the essential spirit of the technical solution of the present invention.

[0019] In the following description, certain specific details are set forth for the purpose of illustrating various disclosed embodiments in order to provide a thorough understanding of the various disclosed embodiments. However, those skilled in the art will recognize that embodiments may be practiced without one or more of these specific details. In other instances, well-known apparatuses, structures, and techniques associated with this application may not have been shown or described in detail to avoid unnecessarily obscuring the description of the embodiments.

[0020] Unless the context requires otherwise, throughout the specification and claims, the word “comprising” and its variations, such as “including” and “having”, shall be understood to have an open, inclusive meaning, that is, to be interpreted as “including, but not limited to”.

[0021] Throughout this specification, references to "an embodiment" or "an embodiment" indicate that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment. Therefore, the appearance of "in an embodiment" or "an embodiment" in various places throughout the specification does not necessarily refer to the same embodiment. Furthermore, a particular feature, structure, or characteristic may be combined in any manner in one or more embodiments.

[0022] The singular forms “a” and “the” used in this specification and the appended claims include plural references unless otherwise expressly stated herein. It should be noted that the term “or” is generally used to include the meaning of “or / and” unless otherwise expressly stated herein.

[0023] In the following description, in order to clearly demonstrate the structure and working method of the present invention, a number of directional terms will be used. However, terms such as "front", "back", "left", "right", "outside", "inside", "outward", "inward", "up", and "down" should be understood as convenient terms and not as limiting terms.

[0024] Existing technologies have significant shortcomings in unstructured knowledge fusion and model domain adaptation: First, traditional retrieval methods struggle to identify chapter levels and semantic boundaries in energy storage documents, leading to low cross-document evidence recall and contextual breaks, thus affecting reasoning quality. Second, general-purpose models, without targeted optimization, exhibit knowledge domain bias when directly processing energy storage-specific operation and maintenance corpora, resulting in conclusions prone to terminological bias and factual illusions, failing to meet the decision-making requirements of high-security-standard scenarios. Third, full parameter updates are computationally expensive and lack lightweight continuous learning mechanisms, making it difficult to support enterprise-level private deployments and dynamic iteration of knowledge bases. Furthermore, when general-purpose large language models are directly applied to energy storage scenarios, the lack of targeted optimization incorporating domain knowledge leads to knowledge domain bias and unstable logical reasoning, easily generating erroneous or ambiguous conclusions, failing to meet the requirements of energy storage systems for accuracy, interpretability, and security in operation and maintenance decisions.

[0025] This invention provides a response method for energy storage systems based on retrieval enhancement generation and vertical fine-tuning of a large model for the energy storage industry. By deeply integrating a large language model with a domain knowledge base, it realizes automatic retrieval, reasoning, and decision generation of operation and maintenance knowledge.

[0026] First, a knowledge base for energy storage operation and maintenance (O&M) is constructed using a Retrieval Enhanced Generation (RAG) mechanism. This involves semantic segmentation and vectorization of multi-source documents, including power plant operation data, equipment manuals, industry standards, and historical O&M reports. Then, during the retrieval phase, title awareness and reordering algorithms are introduced to improve the accuracy and traceability of cross-document and cross-version knowledge matching, enabling efficient retrieval and evidence retrieval of O&M knowledge. Finally, considering the specialized knowledge characteristics of energy storage scenarios, a domain-based fine-tuning strategy using the LoRA method is employed. Energy storage standards and specifications, alarm corpora, and expert Q&A data are used as training sets to perform lightweight parameter updates on the large language model. This improves the model's accuracy and stability in understanding technical terms, system structure, and operational logic. A feedback correction mechanism is then used to achieve continuous self-learning and dynamic knowledge updates.

[0027] This method utilizes a Retrieval Enhanced Generation (RAG) mechanism to construct a vector knowledge base for the energy storage field and introduces title awareness and reordering algorithms to improve the accuracy and relevance of the model in extracting key information from complex documents. By using a large language model fine-tuned for electrochemical energy storage systems as the core reasoning foundation, the model can better understand the characteristics of energy storage devices, system topology, and operational patterns. Through the above technical solutions, this invention can significantly reduce the reliance on human experience in the operation and maintenance of energy storage systems, realizing a shift from "manual analysis—passive response" to "intelligent diagnosis—proactive early warning." This not only improves the automation and timeliness of operation and maintenance work but also enhances the accuracy and consistency of anomaly identification and decision-making results, providing technical support for the safe and reliable operation of energy storage systems.

[0028] The first embodiment of this invention relates to a response method for energy storage systems, applied to electronic devices, such as electronic devices, servers, or common computer devices, such as desktop computers and laptops. The electronic devices can respond to user queries from energy storage systems based on this response method. A systematic risk indicator system covering eight categories and dozens of detailed factors is constructed based on multiple risk indicators involved in the actual operation of electrochemical energy storage systems. Based on this systematic risk indicator system, a multi-level risk indicator system covering all key aspects can be constructed, forming a risk factor tree. This risk factor tree is a structured risk index system for electrochemical energy storage systems (EESS), namely the ERI (EESS Risk Indexing) system. The ERI system, based on risk identification and factor decomposition, integrates national standards, industry specifications, expert knowledge, and multi-source text data to classify and dynamically characterize potential risks of energy storage systems in terms of structural safety, battery safety, and functional reliability. The ERI system not only serves as a core supporting tool for risk assessment but also provides a quantitative basis for subsequent risk warning and operation and maintenance strategies.

[0029] For example, the risk indicators of an energy storage system are divided into four levels, which are as follows from top to bottom: subsystem (risk surface), risk factor, secondary risk factor, and risk item. Each level includes at least one risk indicator, and the risk indicators in the higher level are associated with one or more risk indicators in the lower level (referred to as lower-level risk indicators).

[0030] At the subsystem level, there are eight core subsystems: cell subsystem, battery pack subsystem, battery cluster subsystem, energy storage cabinet / station subsystem, thermal management subsystem, fire protection subsystem, power electronic devices and BMS subsystem (BMS, Battery Management System); the above eight core subsystems constitute the first-level risk classification covering the main subsystems of the energy storage system.

[0031] For each subsystem, based on the physical attributes and impact paths that generate risks, four risk factors are identified: electrical characteristics, thermal characteristics, mechanical characteristics, and environmental characteristics, covering the most important risk dimensions in energy storage systems. These four factors comprehensively reflect the potential sources of problems in energy storage systems and, together with the aforementioned subsystem classification, constitute a two-dimensional risk factor matrix of "subsystem × risk factor".

[0032] Each risk factor within a subsystem can include multiple sub-risk factors. For example, the electrical characteristics of a cell subsystem can include sub-risk characteristics such as performance failure, internal short circuit, and abnormal use. Each sub-risk factor can also include multiple risk items. For instance, the sub-risk factor "performance failure" can further include two risk factors: low coulombic efficiency and excessive self-discharge rate. Other risk factors, sub-factors, and risk items can be set as needed and will not be detailed here.

[0033] After constructing the systemic risk indicator system, a coupling relationship network is introduced to characterize the mutual influence between different risk indicators. The constructed coupling relationship network adopts a directed graph structure, which can characterize the coupling relationship between risk indicators, including the coupling relationship between risk indicators at the same level and the coupling relationship between risk indicators at different levels. The coupling edges characterize the influence relationship between different risk indicators.

[0034] In one example, a risk-coupled network can be represented by an adjacency matrix, specifically; in, This represents the coupling relationship between the i-th risk indicator and the j-th risk indicator. =1, indicating a coupling relationship between the i-th risk indicator and the j-th risk indicator; if =0 indicates that there is no coupling relationship between the i-th risk indicator and the j-th risk indicator.

[0035] like Figure 1 The diagram shown is a detailed flowchart of the response method for energy storage systems in this embodiment.

[0036] Step 101: Obtain multi-dimensional text data in the field of energy storage, and construct a training dataset including multiple training samples based on the multi-dimensional text data. Each training sample includes: input text and output text.

[0037] Specifically, this involves collecting multi-dimensional textual data in the energy storage field. This multi-dimensional textual data includes at least one of the following: standard data, operation and safety management data, research data, and operation monitoring data; among which: Regarding standard data, national and industry-issued standards related to energy storage safety were collected, such as the "Safety Technical Specifications for Energy Storage Systems" and the "Design Specifications for Energy Storage Power Stations." These documents clearly define the technical requirements and safety boundaries for energy storage systems in design, construction, installation, operation, and maintenance, providing authoritative standard basis for risk assessment and intelligent operation and maintenance models. Additionally, technical manuals and user guides from major energy storage equipment manufacturers were collected, including equipment structural parameters, operating procedures, maintenance specifications, and safe operation guidelines. This type of data, as an important extension of standard documents, provides accurate structured references for model training and equipment-level knowledge extraction.

[0038] Regarding operation and safety management data, safety inspection reports and accident investigation reports from energy storage power stations were collected. Safety inspection reports reveal typical hidden dangers and abnormal patterns exposed during long-term operation of the power stations, possessing high empirical value. Accident investigation reports detail the causes, evolution, and handling measures of fault events, providing key case evidence for causal reasoning and risk classification in the model. This type of data can be used in the intelligent reasoning module for abnormal pattern recognition and causal chain construction, helping to improve the accuracy and interpretability of the model in fault prediction and safety early warning tasks during training.

[0039] In terms of research data, it integrates research findings related to risk assessment, diagnosis, and operation and maintenance of energy storage systems from academic papers, research reports, and industry technical literature. By extracting structured information such as risk factor trees, coupling relationship diagrams, and factor weights, it systematically depicts the hierarchical relationships and interaction mechanisms of risk factors in energy storage systems. Simultaneously, it performs semantic extraction and knowledge mapping on different algorithm models, fault mechanisms, and control strategies to supplement the deficiencies of standard and report data at the mechanism and algorithm levels, constructing a knowledge network with depth and logical connections.

[0040] Regarding operational monitoring data, historical operating records of the energy storage power station were collected, including battery charge / discharge curves, voltage, current, temperature, and key operating parameters. This type of data is used for risk model verification and parameter calibration, enabling the fusion of quantitative information and semantic knowledge, and improving the computational reliability and generalization ability of the intelligent operation and maintenance model.

[0041] Based on the collected multi-dimensional text data in the energy storage field, a knowledge base was formed through multi-level and multi-dimensional data fusion. Please refer to [link / reference]. Figure 2The knowledge base includes technical documents, standard files, historical cases, and expert knowledge. This establishes a comprehensive data system encompassing standard constraints, operational feedback, mechanistic knowledge, and research findings, providing a unified data foundation and scalable support environment for subsequent fine-tuning of dataset construction and retrieval enhancement generation mechanisms. In other words, by systematically collecting, organizing, and integrating multi-dimensional data resources, it provides unified and standardized basic data support for subsequent knowledge retrieval, model training, and intelligent reasoning.

[0042] After forming a knowledge base containing multi-dimensional text data, structured, semantically clear, and logically computable training samples are generated based on multi-source heterogeneous multi-dimensional text data. These training samples provide knowledge carriers with high semantic density and task-oriented features for subsequent model training, and thus constitute the training dataset.

[0043] First, the multi-dimensional text data is semantically normalized and textually standardized to obtain a standardized text sequence. Then, the standardized text sequence is segmented to obtain multiple initial corpus fragments. Specifically: First, the multi-dimensional text data is screened and cleaned. For the multi-dimensional text data, unified semantic normalization and text standardization are performed to eliminate symbol redundancy and language inconsistency caused by differences in sources. Specific processing methods include: symbol standardization, redundant sentence cleaning, segmented syntax optimization, format repair and abnormal character detection. The text data can be distilled through an LLM large language model.

[0044] For example, for multi-dimensional text data, a unified strategy is adopted, including full-width to half-width conversion, stop word filtering, industry terminology standardization, and punctuation replacement. Specifically: full-width English letters, numbers, and common symbols (such as commas, parentheses, and spaces) are uniformly converted to half-width forms; high-frequency words without actual semantic contribution are removed based on general or domain-specific stop word lists; for specific domains (such as finance, healthcare, and law), a thesaurus or rules are established to standardize non-standard expressions, abbreviations, and aliases into standardized terms; and Chinese punctuation is appropriately processed according to application scenarios, including deleting redundant punctuation and replacing special symbols according to rules. This ensures the linguistic consistency and logical coherence of multi-dimensional text data. The formulaic representation is as follows: ; in, Represents a normalized text sequence. Representing multi-dimensional text data, This represents the normalization processing function. This represents the corpus cleaning function.

[0045] The standardized text sequence obtained above can significantly reduce lexical noise and spelling variations in the corpus, unify the expression of key terms, and provide a high-quality input foundation for subsequent structured representation and semantic annotation.

[0046] The obtained normalized text sequence is logically segmented and semantically clustered using an automated text segmentation algorithm to obtain multiple initial corpus fragments with clear semantic boundaries and normalized grammatical structure.

[0047] Based on semantic analysis and knowledge extraction of each initial corpus fragment, each initial corpus fragment is converted into an input-output pair, each input-output pair including: input text and output text; in detail: For each initial corpus fragment, semantic parsing and knowledge extraction are performed. Semantic parsing identifies and classifies the initial corpus fragments to determine whether they are knowledge-based question-and-answer tasks or operation and maintenance instruction tasks. Knowledge extraction extracts tags from the initial corpus fragments, such as risk factors, safety measures, and operational indicators. Based on the semantic classification results of the initial corpus fragments, input text is generated. Based on the extracted knowledge tags and the initial corpus fragments, output text corresponding to the input text is generated. The input text and the corresponding output text constitute an input-output pair. This input-output pair can be in an "Instruction-Response" or "Question-and-Answer" (Q&A) data format that conforms to the language model learning mechanism.

[0048] Based on the above process, input-output pairs corresponding to all initial corpus fragments can be generated. The core of this process is: the model constructs an explicit task semantic objective, enabling it to possess domain reasoning capabilities when understanding instruction intent and generating response content; this process can be formalized as follows: in, Indicates an input / output pair. This indicates the input text. Represents the initial corpus fragment, Indicates based on initial corpus fragments The input text generated by reasoning, Indicates input text The expected output text after being input into the model. Indicates the initial corpus fragment The output text generated for the target.

[0049] For example, input text The output text is: "Please analyze the operational risks of the energy storage system based on the following clauses". The risk type corresponding to this clause is thermal characteristic risk, which mainly affects the battery cell.

[0050] In the above processing, the self-supervised generation capability of a general-purpose large language model can be used to perform preliminary semantic classification, label extraction, and input-output pair generation on the initial corpus fragments. For the labels extracted from the initial corpus fragments, domain experts can be further prompted for review to ensure consistency and accuracy. In particular, key labels such as "risk factors," "safety measures," and "operational indicators" can be manually corrected to ensure the professionalism and logical consistency of the labeled semantics. This consistency mechanism significantly improves the semantic accuracy and task coverage of the initial corpus fragments, enabling the fine-tuned model to more effectively learn energy storage safety knowledge and reasoning logic during subsequent training.

[0051] Furthermore, the input-output pairs are structured and modeled to enable the model to learn the complex causal networks and risk propagation logic in the energy storage system from these input-output pairs. This process is achieved by introducing a risk coupling relationship network, in which the adjacency matrix corresponding to the risk coupling relationship network can be serialized to form a hierarchical knowledge representation structure in JSON format, so as to support the model in capturing causal relationships and multidimensional logical structures during training.

[0052] The specific method is as follows: For each input-output pair, if the input-output pair involves risk indicators of the energy storage system, then based on the risk coupling relationship network of the energy storage system, the risk indicators in the output text of the input-output pair are associated according to the coupling relationship. For example, for each input-output pair, semantic similarity comparison is used to determine whether there is a risk indicator in the risk coupling relationship network that matches the input text of the input-output pair. If there is, it is determined that the output text of the input-output pair can be modified to embed the causal relationship represented by the indicator coupling relationship in the risk coupling relationship network into the output text. If there is no such indicator, it means that the input-output pair is unrelated to the risk coupling relationship network and no modification is required. For example, the risk coupling relationship network is represented by an adjacency matrix. The method of modifying the output text is as follows: extract the target risk indicators contained in the output text, and then associate two target risk indicators with a coupling relationship with the corresponding element "1" in the adjacency matrix; for two target risk indicators without a coupling relationship, associate them with the corresponding element "0" in the adjacency matrix.

[0053] Based on the above process, the input-output pairs involving risk indicators in energy storage systems can be modified. By injecting structured semantics and fine-tuning with command-based samples, causal relationships are explicitly embedded into the input-output pairs, enabling the model to learn the potential risk propagation network structure through a self-attention mechanism. When using these input-output samples for subsequent training, the model not only understands the semantic meaning of individual risk indicators but also grasps the causal dependencies and propagation paths between risk indicators. The coupling relationship construction in the large language model is achieved not through explicit graph computation but through implicit graph structure learning based on semantic encoding. A large number of structured input-output pairs enable the self-organized expression of risk propagation paths in the model's semantic space.

[0054] For example, consider an input-output pair, a command-response pair. The command is "What secondary risks in the cell subsystem of the ERI system have significant propagation relationships?"; the modified response is "Cell performance failure → Cell thermal abuse" and "Cell internal short circuit → Cell performance failure". This formatted expression is encoded as a sequence dependency by the model during training. Through a self-attention mechanism, directional weights are formed between word vectors, enabling the model to gradually learn to represent the triggering logic between "cell internal short circuit" and "cell performance failure" in the semantic space. When a large number of similar input-output pairs are used to train the model, the co-occurrence patterns between different factors and risk paths are aggregated into the attention weight matrix, forming an implicit "semantic graph," that is, a risk propagation network is constructed in the model's internal parameter space.

[0055] For example, taking another input-output pair as an example, the input text is "Please explain which secondary risks and their corresponding risk factors are specifically included in the electrical failure dimension of the cell subsystem in the ERI system"; the modified output text is "Performance failure - low coulombic efficiency, excessive self-discharge rate; internal short circuit - internal short circuit of the cell; abnormal use - overcharge and over-discharge, charging and discharging outside the window"; such input-output pairs participating in model training can help the model establish a hierarchical structure cognition, enabling it to understand the nested logic of "risk dimension - secondary risk - risk factor" at the semantic level.

[0056] By repeatedly learning from this type of structured instruction-response data, the model's multi-head attention layer can capture "hierarchical relationships," "lateral propagation paths," and "causal strength mappings" in different dimensions, achieving an organic fusion from linguistic representation to causal relationship modeling. The result of this training method is that even when faced with new instructions that have not appeared before, the model can infer possible propagation chains and risk dependency structures based on internalized causal weights and semantic similarity. For example, if the input is "Please analyze the secondary risks that may be caused by an increase in the internal resistance of the battery cell," the model will infer based on the learned pattern of "increased internal resistance ≈ performance failure → thermal abuse," generating a logically consistent response.

[0057] The input-output pairs obtained from the aforementioned process are used as training samples to form the training dataset.

[0058] Furthermore, the system can perform format checks on input and output pairs to ensure field completeness, type matching, and semantic consistency. For samples with detected structural errors or semantic ambiguities, the system will automatically go back to the previous annotation stage to regenerate and correct them.

[0059] In one embodiment, to enhance the linguistic diversity and robustness of the model, strategies such as semantic paraphrasing, template replacement, back-translation, and multi-task hybrid instruction generation are used to expand the input-output pairs. This can yield more input-output pairs. These expanded input-output pairs are then added to the training dataset as training samples. During model training, this allows the model to maintain high consistency in understanding and stability in generation when faced with different representations, semantic variations, and task context switching.

[0060] After the above steps, the final training dataset is output in a high-quality JSON structure, which has both rigorous semantic hierarchy and logical connection, and meets the dual requirements of format standardization and knowledge density for training large-scale language models in the domain. This provides solid data support for efficient LoRA parameter adaptation and task instruction fine-tuning in the future.

[0061] Step 102: Train the base model based on the multi-dimensional text data.

[0062] Step 103: Use the training dataset to train the trained base model to obtain a dedicated energy storage model.

[0063] Specifically, after completing the construction of a high-quality training dataset, a hierarchical and progressive fine-tuning process was designed for the operation and maintenance of electrochemical energy storage, which includes two steps: semantic alignment pre-tuning and instruction fine-tuning.

[0064] The semantic alignment pre-training process involves using a pre-trained generalized large language model (LLM) as the base model and training it with collected multi-dimensional text data. This training process employs a low-learning-rate fine-tuning strategy to enable the model to adapt to specialized corpora in the energy storage field. This low-learning-rate training method can specifically enhance the model's understanding of specific terms, professional concepts, and logical structures in the energy storage field without compromising the base model's original general semantic understanding capabilities.

[0065] The objective function used in this training process is: In this process, the multi-dimensional text data is divided into N segments and used to train the base model. This represents the i-th segment of text. This represents the output obtained after the i-th text segment is input into the base model. The parameters of the basic model.

[0066] By minimizing this training objective function, the trained base model can better maintain semantic consistency when generating or understanding domain text, improve its ability to recognize and apply specialized terms, and thus lay a solid foundation for the subsequent task fine-tuning stage.

[0067] After completing semantic alignment pre-training, the model enters the task instruction fine-tuning stage, which aims to enable the model to generate high-quality, business-logical responses based on specific operation and maintenance tasks or user instructions.

[0068] In the constructed training dataset, the training samples are input-output pairs transformed from operational scenarios, fault analysis, and state assessment related to energy storage operation and maintenance. These training samples are used to fine-tune the semantically aligned pre-trained base model. During this training process, the base model not only needs to understand the semantics of the instruction text but also needs to learn how to execute corresponding operations or generate specific responses based on the instructions, thereby improving the usability and response accuracy of the base model in real-world operation and maintenance scenarios.

[0069] The objective function used for training is the Conditional Maximum Likelihood Estimation (CMLE) function: in, This represents the input text in the i-th training sample (input-output pair). This represents the output text in the i-th training sample (input-output pair). This represents the task context or constraint information attached to the i-th training sample (input-output pair), which limits the relevant information of the energy storage field currently applied by the basic model. By minimizing this objective function, the basic model can generate answers that meet the expected logic and task requirements when receiving input instructions, and gradually form the ability to operate on complex tasks.

[0070] Furthermore, when training the base model that has been pre-trained with semantic alignment using the training dataset, a LoRA module is added between each key linear transformation layer of the base model and its connected feedforward layer to obtain an intermediate model. The intermediate model is trained using training samples from the training dataset, and the weight parameters of each key linear transformation layer in the base model are updated to form a dedicated energy storage model. During the training of the intermediate model, the low-rank decomposition matrix corresponding to each LoRA module is updated, and the weight parameters of the key linear transformation layer connected to the LoRA module are updated based on the low-rank decomposition matrix corresponding to each LoRA module.

[0071] As can be seen from the above, this application, while freezing the weight parameters of the original base model, introduces a trainable LoRA module only on each key linear transformation layer. During training, only the low-rank factorization matrix of the LoRA module is trained. After each training round, the weight parameters of the connected key linear transformation layers are updated based on the low-rank factorization matrix of the LoRA module, thereby achieving targeted adaptation to specific tasks with very few additional parameters. Thus, the Low-Rank Adaptation (LoRA) mechanism is used to efficiently fine-tune the parameters of the base model, thereby significantly reducing memory usage while maintaining stable model performance. Key linear transformation layers include, for example, attention layers.

[0072] For example, the original weight parameters of the key linear transformation layer are: Then the low-rank dimension of the low-rank decomposition matrix of the LoRA module it connects to. The low-rank decomposition matrix of the LoRA module can be represented as: ;in, , Through this decomposition, only updates are needed during training. and By freezing the weights of the key linear transformation layers, the amount of parameter updates is greatly reduced.

[0073] During training, gradient update is used to update the low-rank decomposition matrix of the LoRA module, i.e., update... and The expression for gradient update is: ; Based on this gradient update expression, we can obtain and The weight parameters of the key linear transformation layer can then be updated based on the following calculation formula. for: .

[0074] The above steps update the weight parameters of all key linear transformation layers in the base model, resulting in the required base model for energy storage. The LoRA module is injected between the Self-Attention Layer and Feed-Forward Layer of each Transformer module in the base model, enabling the base model to quickly learn new knowledge mapping relationships for energy storage system-specific tasks while maintaining global semantic stability. Experimental results show that the base model fine-tuned using the LoRA mechanism has only 0.3%–0.7% of the parameter size of the original model, but its performance in energy storage tasks is close to or even surpasses that of full fine-tuning.

[0075] The introduction of the LoRA mechanism significantly improves the model's adaptability and transferability in the field of energy storage security, enabling it to converge quickly under limited computing resources while maintaining high-precision modeling capabilities for complex structures such as risk factor coupling logic and causal chain propagation mechanisms. This approach technically solves three major problems: First, it lowers the resource threshold—the model can complete domain migration in a single-card environment, making it suitable for enterprise-level private deployments; second, it prevents language degradation—because the backbone parameters are frozen, the model retains its original semantic fluency and generative diversity; and finally, it supports continuous incremental learning—LoRA's modular weight design allows for flexible loading or replacement, facilitating rapid adaptation to new domains during model updates or knowledge base expansion.

[0076] For example, three open-source large-scale models, LLaMA-3.2-3B-Instruct, Deepseek-R1-Distill-Qwen7B, and Qwen2.5-7B-Instruct, were selected as comparison objects. These large-scale models were trained in two steps: semantic alignment pre-training and instruction fine-tuning, as described above, to compare the adaptation effects of different model architectures in semantic learning and instruction following tasks. A uniform data partitioning and parameter settings were used: the learning rate was 2×10⁻⁶. -4 The gradient accumulation step size was 8, the maximum truncation length was 2048, the number of training epochs was 20, the optimizer was AdamW, and the Cosine Annealing scheduling strategy was enabled. To balance computational overhead and training convergence speed, gradient clipping and mixed precision (bfloat16) calculations were performed in full-parameter mode.

[0077] like Figure 3As shown, the loss curves during training vary with the number of training updates (steps) for experimental groups under different experimental conditions. The experimental groups include: a standard group, a low learning rate group, a group comparing the number of training epochs, a group comparing gradient accumulation (GA) and a group comparing stage lengths. All experimental groups use LLaMA-3.2-3B-Instruct as the base model, and then adjust the corresponding training hyperparameters to analyze the impact of different parameter settings on domain adaptation. In these experimental groups, "original" represents the original training loss curve of the base model trained in two steps: semantic alignment pre-training and instruction fine-tuning; "smoothed" represents the training loss curve obtained after smoothing the original training loss curve. Additionally, the experimental groups also include a model comparison group, which is used to verify the adaptability of different base model architectures in the energy storage domain corpus fine-tuning task.

[0078] In the above experimental groups, the overall trend shows that the basic model in each group showed a significant downward trend in the early stage of training, indicating that the basic model can effectively learn the professional language patterns and knowledge distribution in the corpus of energy storage.

[0079] In the standard group, the training loss rapidly decreased from about 3.2 to about 1.4 within the first 100 steps, and then gradually stabilized, indicating that Llama-3.2-3B-Instruct can complete domain semantic adaptation relatively quickly under the current training configuration.

[0080] Low learning rate group (Learning Rate = 1 × 10) -4 In the study, the descent process was slower in terms of the influence of hyperparameters, and the final training loss was higher than that of the standard group. This indicates that the low learning rate limited the range of parameter updates of the basic model, resulting in insufficient ability of the basic model to transfer domain knowledge.

[0081] In the group comparing the number of training epochs, the results show that increasing the number of training epochs can further reduce the training loss, but the downward trend gradually slows down in the later stages, and the model's returns tend to saturate.

[0082] In the gradient accumulation comparison group (gradient accumulation GA=8), a lower final training loss was obtained while maintaining a relatively stable downward trend, indicating that appropriately increasing the effective batch size helps to improve the stability of the training process.

[0083] In the truncation length comparison group (truncation length set to 2048), longer text truncation can retain more contextual information, enabling the base model to maintain better convergence performance in the mid-to-late training stages.

[0084] In the model comparison group, the Qwen2.5-7B-Instruct model and the Llama-3.2-3B-Instruct model were compared, with other parameters remaining consistent. Both are lightweight, instruction-based fine-tuning large language models with similar parameter sizes (Qwen2.5-7B-Instruct has approximately 7 billion parameters, while Llama-3.2-3B-Instruct has approximately 3 billion parameters). Both possess strong general language understanding capabilities, thus effectively reflecting the differences in domain adaptability between different models. Experimental results show that the Llama-3.2-3B-Instruct model exhibits a smoother training loss curve and a lower final convergence value, indicating better knowledge transfer ability and domain adaptability during fine-tuning of specialized corpora. In contrast, although Qwen2.5-7B-Instruct has a larger parameter size, it lags behind in terms of depth of specialized domain knowledge learning and convergence stability.

[0085] It can be seen that the model exhibits a stable downward trend within the first 100 steps. In the standard group, the loss dropped rapidly from 3.2 to 1.4 and then gradually stabilized, indicating that the model can effectively capture domain language patterns in the stored corpus. Low learning rate group (1×10⁻⁶) -4 The first model exhibits a slow convergence but a high final loss, indicating that a low learning rate affects the semantic adaptation depth. Conversely, the combination of gradient growth (GA=8) and a truncation length of 2048 shows the best convergence performance in the mid-to-late stages, with significantly better fluency and consistency in the generated text compared to the short-truncation model. Comparisons between different models show that Qwen2.5-7B-Instruct outperforms the other two basic models in both convergence speed and stability, exhibiting the smoothest loss curve and the lowest final convergence point.

[0086] Further question-and-answer test results show that the Qwen2.5-7B-Instruct model performs well in the tasks of parsing energy storage standard clauses and generating operation and maintenance suggestions, with standardized output format, complete logic, and accurate terminology. Deepseek-R1-Distill performs similarly in general language tasks, but is slightly inferior in specialized semantic understanding. LLaMA-3.2-3B, due to its smaller parameter size, suffers from information omissions and repetitions in long text generation. Analysis of the loss curve and output results indicates that a higher learning rate (2×10⁻⁶) is more effective. -4The optimal results can be achieved without causing overfitting by setting the gradient accumulation GA to 8 and the cutoff length to 2048.

[0087] In summary, the Llama-3.2-3B-Instruc model, as a foundational model, demonstrates superior training stability and domain adaptability in current fine-tuning tasks within the energy storage field. Furthermore, it exhibits higher linguistic accuracy and professional consistency in subsequent tasks such as knowledge question answering, risk assessment, and operation and maintenance strategy generation.

[0088] Step 104: For the received user question, determine multiple target knowledge units related to the user question from multiple knowledge units obtained by dividing the multi-dimensional text data, and each knowledge unit has been labeled with its hierarchical path information.

[0089] Step 105: Based on the user question and the multiple target knowledge units, obtain the response content corresponding to the user question output by the energy storage dedicated model.

[0090] Specifically, multi-dimensional text data, as a knowledge source, needs to be segmented first to divide it into multiple knowledge units. The specific process involves: based on preset segmentation parameters and semantic similarity thresholds, dividing the multi-dimensional text data into multiple knowledge units, and then labeling the hierarchical path information of each knowledge unit using hierarchical identification. The segmentation parameters include: the length of a single segment, the maximum text span of a segment, and the overlap ratio of adjacent segments. More details: The first step is to preprocess and slice the multi-dimensional text data, as the granularity of knowledge retrieval directly affects the recall accuracy and semantic matching effect. First, the multi-dimensional text data is formatted, converting the heterogeneous multi-source text data into high-quality text data based on the lightweight markup language Markdown hierarchical structure. Then, chunking is performed, dividing the long document into several semantically independent text segments.

[0091] Let D be the collection of documents containing multi-dimensional text data: ; Each document Classified as Individual blocks: ; in, Document The first sub-block obtained from the division, and so on, Document The division obtained Each sub-block can serve as a knowledge unit.

[0092] In this embodiment, the RecursiveCharacterTextSplitte module from the LangChain library can be used, primarily based on paragraph hierarchy, supplemented by an adaptive segmentation strategy based on Markdown hierarchical structure, thereby ensuring the semantic continuity of chapters. The length L of each text block is... c Keep the number of characters between 500 and 1000 and set the overlap ratio of adjacent blocks to r=0.1 to ensure a smooth transition of contextual information at the boundaries.

[0093] The knowledge units obtained from the aforementioned process can be embedded into a model. After processing, it is mapped to a vector space. ,Right now: .

[0094] in, Representing knowledge units semantic vector representation, dimension .

[0095] Furthermore, this embodiment employs a semantically adaptive chunking algorithm to dynamically divide each document into independent knowledge units based on title level, logical paragraphs, and contextual semantic changes. Each chunk corresponds to a semantically complete minimum knowledge fragment, and the original chapter path information is preserved during segmentation to ensure traceability and contextual integrity when referencing. In addition, to enhance document structure recognition capabilities, the system introduces a title-aware module mechanism during chunking, reconstructing the document structure using Markdown hierarchical markers (#, ##, ###, etc.) to enable the model to automatically identify the hierarchical relationship between titles and content during retrieval.

[0096] The n knowledge units obtained after dividing a document set D of multi-dimensional text data into blocks can be represented as: ; in, This represents a heading detection function used to identify chapter numbers and heading levels in a document (such as "Chapter ×", "Article ××", or Markdown level symbols #, ##, ###). This is the window length parameter, used to limit the maximum text span of the block; The function is based on a semantic similarity threshold. Contextual continuity is assessed to segment the text while preserving semantic integrity. This allows for the annotation of hierarchical path information for each knowledge unit, such as: #Level 1 heading, ##Level 2 heading, ###Level 3 heading, ..., #####Level 5 heading, etc. Figure 4 As shown.

[0097] This method preserves the hierarchical logic of the original document in terms of structure, so that each block contains a complete context and has independent semantic boundaries, providing a knowledge unit that can be precisely located for subsequent retrieval.

[0098] The knowledge units obtained from the aforementioned process can be embedded into a model. After processing, it is mapped to a vector space. ,Right now: .

[0099] in, Representing knowledge units semantic vector representation, dimension .

[0100] In the above process, the BAAI / bge-m3 model is used as the embedding model in the system implementation of the vector embedding and storage stages. This model is based on contrastive learning and achieves cross-language, multimodal representation alignment by maximizing the cosine similarity of vectors of similar semantic texts. Vector storage uses the Chroma vector database, whose core indexing mechanism is based on Approximate Nearest Neighbor (ANN) search.

[0101] Based on the aforementioned process, multiple knowledge units can be obtained by dividing multidimensional text data, which can then be used for subsequent knowledge retrieval.

[0102] After receiving a user's question, the knowledge retrieval stage begins. The LangChain framework can dynamically connect the user's question input to knowledge retrieval and then to context injection. A retrieval-question-answering linkage mechanism can be built based on the LangChain framework's RetrievalQA Chain. Specifically: The user's question q is passed through the same embedding model. Mapped to user question vectors From the multiple knowledge units obtained by dividing the multi-dimensional text data, multiple candidate knowledge units similar to the user's question are retrieved; that is, from the knowledge unit vector database, vectors similar to the user's question are retrieved. The set of candidate vectors corresponding to the top-k most similar candidate knowledge units : Cosine similarity can be used to characterize the similarity between vectors, then: Subsequently, the relevance scores between the user's question and each of the candidate knowledge units are calculated. Based on the relevance scores of the candidate knowledge units, the target knowledge units are selected from the candidate knowledge units. Thus, the initial k candidate knowledge units can be further filtered and sorted, specifically using a cross-encoder. We jointly model the user question q and candidate knowledge unit h, and recalculate the relevance score between the user question q and candidate knowledge unit h from three dimensions: semantic, logical, and contextual consistency. , can be represented as: .

[0103] Subsequently, based on a preset scoring threshold, knowledge units with relevance scores greater than the threshold are selected from the k candidate knowledge units. l The candidate knowledge units are selected as the target knowledge units. This allows the attention mechanism to determine whether the user's question is truly relevant to the document of the candidate knowledge unit, thereby more accurately filtering out noisy documents and improving the accuracy and context matching of knowledge retrieval.

[0104] In the above process, text segmentation ensures the rationality of semantic unit granularity and the integrity of references, while rearrangement ensures the relevance of context and the degree of logical convergence during model generation. Ultimately, this achieves an evolution from "generalized retrieval" to "semantic matching and reasoning coupling," enabling the generated final answer to reach industry-level standards in terms of factual citation, logical coherence, and professional accuracy. For example, if a user asks, "What are the testing standards for the insulation performance of battery modules in high-altitude environments?", the three target knowledge units retrieved would include: Document 1, Document 2, and Document 3. See [link to documentation] for details. Figure 5 As shown, the hierarchical path information for each target knowledge unit is also marked.

[0105] Then, user questions q With this l Target knowledge unit The input information is dynamically embedded into a preset prompt template via context injection to form enhanced input information. In this enhancement generation stage, a general large language model (LLM) can be used to implement the retrieved context information. Explicitly injected into the model input, guiding the generated results to align with external knowledge, the generation process can be represented as follows: ; in, This indicates enhanced input information. This represents the model parameters of the general large language model LLM; for the general large language model LLM, the objective is to maximize the generation likelihood under the conditional probability distribution, i.e.: .

[0106] To ensure the controllability and interpretability of the generated results, a structured preset prompt template is used to construct enhanced input, which incorporates user questions. q With this l Target knowledge unit Unified embedding, with a preset prompt template for example: "Given the following reference: {context}. Please answer the following question based on the above content: {question}." Here, context is... l Target knowledge unit 'question' refers to a question asked by a user. q Conditional generation and semantic fusion are performed by the general large language model LLM, thereby realizing the generation of knowledge-driven augmented input.

[0107] Subsequently, the enhanced input information is input into the energy storage-specific model, which then outputs the response content corresponding to the user's question; thus, the energy storage-specific model can perceive during the generation stage. l The document content of each target knowledge unit is then used to generate the final answer as the response content for the user's question, and provided to the user. This process can be abstractly represented as follows: in, This represents the final answer from the energy storage-specific model. This represents the generator function executed by the energy storage-specific model after receiving enhanced input information. This mechanism achieves an end-to-end semantic flow from semantic retrieval to knowledge injection, balancing linguistic fluency and factual accuracy in the generated results. Furthermore, the response content generated by the energy storage-specific model includes hierarchical path information for each of the target knowledge units.

[0108] In this way, energy storage-specific models can reason based on external knowledge during generation, effectively reducing the "hallucination" phenomenon and improving the consistency and reliability of the generated results. Compared with answers that simply rely on the "memory" of the energy storage-specific model, this retrieval-based dynamic generation allows the model to reference external document knowledge in real time, significantly improving the factual accuracy and interpretive consistency of the answers. Practice shows that incorporating retrieval results can improve the accuracy of answers by more than 10% in tasks such as energy storage standard interpretation, operational safety judgment, and fault tracing analysis, while reducing the incidence of hallucination problems. This solves the problems of knowledge lag and factual bias in traditional models, and provides a highly scalable knowledge support system for intelligent question answering, standard comparison, accident tracing, and risk decision-making in energy storage operation and maintenance, laying a solid foundation for subsequent deployment on actual operation and maintenance platforms.

[0109] For example, if a user asks, "What measures should be taken when an energy storage system experiences [specific fault symptoms]?", the key information of "specific fault symptoms" is first extracted. Then, knowledge related to this fault symptom is retrieved from a vector database, such as possible causes and historical handling cases, to obtain relevant target knowledge units. Based on these retrieved target knowledge units, the accuracy of the energy storage-specific model in answering the question is enhanced, further improving the efficiency of risk assessment analysis and operation and maintenance plan generation. For instance, the energy storage-specific model can generate more accurate and detailed operation and maintenance plans based on retrieved historical cases and the specific circumstances of the current fault symptom.

[0110] In some embodiments, if the user question is an operation and maintenance question, the response content corresponding to the user question is obtained from the energy storage-specific model based on the user question, the multiple target knowledge units, and the risk assessment results of the energy storage system. Specifically, after the user inputs an operation and maintenance question, the system first accurately retrieves relevant clauses and cases (target knowledge units) from the vector library. Then, based on the retrieved evidence (target knowledge units) and the operation and maintenance question, the enhanced input information is input into the large-scale energy storage-specific model fine-tuned by LoRA. The model combines internalized domain parameters and external real-time evidence execution conditions to generate a structured report containing risk rating, causal analysis, and disposal suggestions. The risk score can be interpreted, and targeted operation and maintenance strategies can be proposed for the maintenance and optimization of the energy storage system. The system supports manual verification and feedback feedback, driving the continuous iteration of the knowledge base and adaptation weights. This architecture effectively connects the technical links between unstructured knowledge retrieval and domain model generation, providing underlying support for the standardized operation and maintenance of large-scale energy storage power stations. The overall technical roadmap is described in [link to technical details]. Figure 6 Among them, the results of energy storage risk assessment can be obtained based on existing relevant technologies.

[0111] As can be seen, during the question-and-answer phase, this invention already possesses the capability to provide highly accurate and interpretable responses for the field of energy storage safety. The model can combine risk assessment results with operation and maintenance scenarios to generate logically clear and well-organized response content. The responses include explicit factual references, standard clauses, and causal analysis, significantly improving the credibility of the information and its readability for experts.

[0112] For example, please refer to Figure 7 The user's questions were knowledge-based, specifically: "What are the hierarchical division and functional requirements of the intelligent electrochemical energy storage power station system architecture?" and "What are the insulation performance test standards for battery modules in high-altitude environments?" The retrieved documents and the response content given by the energy storage-specific model were provided. Please refer to Figure 8 The user's question, specifically an operations and maintenance (O&M) question, reads: "The original scores of a power station's subsystems in July are as follows: Battery Cell (C1): 89.589; Battery Pack (C2): 84.840; Battery Cluster (C3): 82.145; Energy Storage Cabinet / Power Cabinet (C4): 80.781; Thermal Management System (C5): 87.296; Fire Protection System (C6): 64.71; Power Electronic Devices (C7): 78.145; Battery Management System (C8): 65.228; Total Score: 79.573. Please provide O&M suggestions based on the reasoning process." The user also provided their thought process and the response from the energy storage-specific model.

[0113] As can be seen from the above, this embodiment realizes the closed-loop implementation from knowledge integration to intelligent question answering, and constructs a basic framework for intelligent operation and maintenance collaboration for energy storage safety, providing a good foundation for subsequent risk assessment and intelligent operation and maintenance of energy storage systems.

[0114] In this embodiment, multi-dimensional text data fusion, retrieval enhancement and generation optimization, and efficient parameter fine-tuning are deeply integrated to construct a complete technical system from knowledge extraction and semantic retrieval to domain reasoning and decision generation. This method addresses engineering bottlenecks in traditional operations and maintenance, such as low utilization of unstructured knowledge, insufficient cross-document retrieval accuracy, and high model domain adaptation costs, achieving structured management of operations and maintenance knowledge and evidentiary support for the reasoning process.

[0115] By systematically integrating various types of information, including energy storage standards, operation reports, accident cases, and expert experience, a semantically unified and scalable energy storage knowledge base was constructed. Based on this, a title-aware segmentation and rearrangement optimization algorithm was designed, significantly improving the accuracy and timeliness of knowledge retrieval. A knowledge retrieval-enhanced mechanism enables the large model to reference external knowledge sources in real time, avoiding the knowledge lag and factual illusion problems inherent in closed-parameter models. As the core of the system, the dedicated energy storage model, supported by LoRA parameter injection and domain corpus fine-tuning, possesses semantic understanding and multi-layered logical reasoning capabilities for energy storage operation status, risk factors, and control strategies. It can complete risk identification, causal judgment, and strategy generation under complex operating conditions. Through the above modular collaboration, a comprehensive intelligent operation and maintenance system covering data acquisition, knowledge management, semantic retrieval, logical reasoning, and strategy output was formed, constructing an intelligent path from static analysis to dynamic cognition in the field of energy storage operation and maintenance.

[0116] The retrieval enhancement and generation algorithm based on title awareness and semantic reordering is used to achieve high-precision knowledge retrieval and dynamic evidence citation; the LoRA-based energy storage model parameter injection and knowledge fine-tuning mechanism is used to enhance the model's domain semantic understanding and task adaptability; and the intelligent operation and maintenance collaborative framework formed on this basis realizes a technical closed loop from multi-source data vectorization, semantic retrieval enhancement, low-rank parameter adaptation to evidence generation. These key components form a complete systematic methodology in structure and achieve deep integration of knowledge, logic, and model capabilities in function, providing a highly reliable and scalable intelligent technical solution for risk assessment, operation and maintenance diagnosis, and safety strategy formulation of energy storage power stations.

[0117] By constructing a collaborative architecture of "multi-source data foundation + retrieval enhancement optimization + LoRA parameter efficient fine-tuning," this system systematically solves core challenges such as knowledge retrieval fragmentation, model domain shift, and high adaptation costs, addressing the entire process from knowledge acquisition and model adaptation to system output. Its fundamental principle lies in: using a title-aware segmentation and cross-coding rearrangement algorithm based on the RAG mechanism to enable the system to possess "high-precision semantic recall and logical verification" capabilities; and using the LoRA low-rank adaptation mechanism to enable the model to possess "lightweight domain transfer and knowledge internalization" capabilities. The synergistic effect of these two mechanisms constructs a traceable, anti-illusion intelligent operation and maintenance technology system from underlying data parsing to upper-level decision generation.

[0118] In terms of methodology, firstly, a knowledge base for the energy storage field is constructed using multi-dimensional text data. Combined with title-aware segmentation and rearrangement algorithms, high-precision semantic retrieval of standards, reports, and documents is achieved, solving the problem of traditional retrieval's inability to identify hierarchical relationships. Secondly, through parameter fine-tuning and adaptive optimization of a large-scale energy storage-specific model, the model possesses domain-specificity and knowledge understanding capabilities at the semantic level, enabling it to output customized strategies and explanations for specific scenarios (such as thermal runaway early warning, cell fault diagnosis, or system safety assessment). Finally, at the system level, a result verification and feedback learning mechanism is introduced. By aligning the generated content with facts and checking consistency, the model achieves continuous optimization and long-term evolution, enabling the intelligent operation and maintenance system to possess self-correction and knowledge update capabilities.

[0119] In summary, this embodiment achieves end-to-end innovation from "precise knowledge retrieval" to "lightweight model adaptation" and then to "evidence-based decision generation." Compared with existing technologies, its advantages are mainly reflected in the following aspects: First, at the retrieval algorithm level, it achieves high-precision matching and dynamic injection of cross-level knowledge, effectively solving the problem of semantic fragmentation in unstructured documents; second, at the fine-tuning mechanism level, it achieves domain semantic alignment and continuous iteration capabilities with extremely low computational cost, breaking through the resource bottleneck of full-parameter optimization; third, at the engineering deployment level, it constructs a verifiable, anti-illusion, and formatted end-to-end collaborative framework, realizing seamless integration between algorithm output and industrial business flow. This method can stably output operation and maintenance decisions with factual basis, logical consistency, and compliance with industry standards under complex energy storage conditions, significantly improving the intelligent management level, risk governance efficiency, and engineering feasibility of electrochemical energy storage systems.

[0120] The second embodiment of the present invention relates to an electronic device, such as an electronic device, a server, or a common computer device, such as a desktop computer or a laptop computer.

[0121] The electronic device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the response method for an energy storage system in the first embodiment.

[0122] It should be noted that the memory may include random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device. Similarly, the processor can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0123] Since the first embodiment corresponds to this embodiment, this embodiment can be implemented in conjunction with the first embodiment. The relevant technical details mentioned in the first embodiment remain valid in this embodiment, and the technical effects achievable in the first embodiment can also be achieved in this embodiment. To reduce repetition, they will not be repeated here. Correspondingly, the relevant technical details mentioned in this embodiment can also be applied to the first embodiment.

[0124] The third embodiment of the present invention relates to a computer-readable storage medium, which is a non-volatile storage medium or a non-transient storage medium, on which a computer program is stored. The computer program is executed by a processor to perform the response method for energy storage systems in the first embodiment.

[0125] The preferred embodiments of the present invention have been described in detail above, but it should be understood that, if necessary, aspects of the embodiments can be modified to utilize aspects, features, and concepts from various patents, applications, and publications to provide other embodiments.

[0126] In light of the detailed description above, these and other changes can be made to the embodiments. Generally, the terminology used in the claims should not be considered limited to the specific embodiments disclosed in the specification and claims, but should be understood to include all possible embodiments together with the full scope of equivalents enjoyed by these claims.

Claims

1. A response method for energy storage systems, characterized in that, include: Acquire multi-dimensional text data in the field of energy storage, and construct a training dataset including multiple training samples based on the multi-dimensional text data. Each training sample includes: input text and output text. The basic model is trained based on the training dataset to obtain a dedicated energy storage model; For a received user question, multiple target knowledge units related to the user question are determined from multiple knowledge units obtained by dividing the multi-dimensional text data. Each knowledge unit has been labeled with its hierarchical path information. Based on the user's question and the multiple target knowledge units, the response content corresponding to the user's question is obtained from the output of the energy storage-specific model.

2. The response method for energy storage systems according to claim 1, characterized in that, Based on the aforementioned multi-dimensional text data, a training dataset comprising multiple training samples is constructed, including: The multi-dimensional text data is semantically normalized and text normalized to obtain a normalized text sequence, and the normalized text sequence is segmented to obtain multiple initial corpus fragments; Based on semantic analysis and knowledge extraction of each initial corpus fragment, each initial corpus fragment is converted into an input-output pair, each input-output pair including: input text and output text; For each input-output pair, if the input-output pair involves risk indicators of the energy storage system, then based on the risk coupling relationship network of the energy storage system, the risk indicators in the output text of the input-output pair are associated according to the coupling relationship; The obtained input-output pairs are used as training samples to form a training dataset.

3. The response method for energy storage systems according to claim 1, characterized in that, The basic model is trained based on the aforementioned training dataset to obtain a dedicated energy storage model, including: A LoRA module is added between each key linear transformation layer of the base model and its connected feedforward layer to obtain an intermediate model; The intermediate model is trained using training samples from the training dataset, and the weight parameters of each key linear transformation layer in the base model are updated to form a dedicated energy storage model. During the training of the intermediate model, the low-rank decomposition matrix corresponding to each LoRA module is updated, and the weight parameters of the key linear transformation layer connected to the LoRA module are updated based on the low-rank decomposition matrix corresponding to each LoRA module.

4. The response method for energy storage systems according to claim 1 or 3, characterized in that, The method further includes: The basic model is trained based on the aforementioned multi-dimensional text data; The basic model is trained based on the aforementioned training dataset to obtain a dedicated energy storage model, including: The trained base model is trained using the training dataset to obtain a dedicated energy storage model.

5. The response method for energy storage systems according to claim 1, characterized in that, The specific method for dividing the multi-dimensional text data into multiple knowledge units is as follows: Based on preset segmentation parameters and semantic similarity thresholds, the multi-dimensional text data is divided into multiple knowledge units, and the hierarchical path information of each knowledge unit is labeled through hierarchical recognition. The segmentation parameters include: the length of a single segment, the maximum text span of a segment, and the overlap ratio of adjacent segments.

6. The response method for energy storage systems according to claim 1, characterized in that, For a received user question, multiple target knowledge units related to the user question are identified from the multiple knowledge units obtained by dividing the multi-dimensional text data, including: Multiple candidate knowledge units similar to the user's question are retrieved from the multiple knowledge units obtained by dividing the multi-dimensional text data. The relevance scores between the user's question and each of the candidate knowledge units are calculated respectively, and the target knowledge units are selected from the candidate knowledge units based on the relevance scores corresponding to the candidate knowledge units.

7. The response method for energy storage systems according to claim 1, characterized in that, Based on the user question and the multiple target knowledge units, the response content corresponding to the user question output by the energy storage-specific model is obtained, including: The user's question and the multiple target knowledge units are respectively embedded into preset prompt word templates to form enhanced input information; The enhanced input information is input into the energy storage-specific model, and the energy storage-specific model outputs the response content corresponding to the user's question.

8. The response method for energy storage systems according to claim 1, characterized in that, The multi-dimensional text data includes at least one of the following: standard data, operation and safety management data, scientific research data, and operation monitoring data.

9. The response method for energy storage systems according to claim 1, characterized in that, The response content includes: hierarchical path information for each of the target knowledge units.

10. The response method for energy storage systems according to claim 1, characterized in that, If the user question is an operation and maintenance question, then based on the user question, the multiple target knowledge units, and the risk assessment results of the energy storage system, the response content corresponding to the user question output by the energy storage-specific model is obtained.

11. An electronic device, characterized in that, include: At least one processor; And a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform a response method for an energy storage system as claimed in any one of claims 1 to 10.

12. A computer-readable storage medium, said computer-readable storage medium being a non-volatile storage medium or a non-transient storage medium, having stored thereon a computer program, characterized in that, The computer program is executed by the processor to perform the response method for energy storage systems as described in any one of claims 1 to 10.

Images