Intelligent generation method and system for power failure first-aid repair scheme

Abstract

The invention belongs to the technical field of power failure repair, and provides an intelligent generation method and system for a power failure repair scheme, and the method comprises the steps: constructing a multi-modal instruction template, obtaining the basic knowledge of the power industry, constructing a data set, training a general large language model through the data set, and employing a hierarchical learning rate strategy to achieve the intelligent generation of the power failure repair scheme. Obtaining a large language model after fine tuning of domain knowledge; obtaining a power failure first-aid repair case, constructing the power failure first-aid repair case as a sample according to a pre-constructed multi-modal instruction template, and optimizing the parameters of the large language model finely adjusted by the domain knowledge by using the sample by taking semantic loss, physical constraint loss and safety loss as a weighted mixed loss function; obtaining a large language model after multi-modal task fine tuning; and processing the target data by using the large language model subjected to multi-modal task fine tuning to obtain a final power failure repair scheme. According to the invention, the accuracy and intelligence of power fault first-aid repair scheme generation are improved.

Description

Technical Field

[0001] This invention belongs to the field of power fault repair technology, specifically relating to an intelligent generation method and system for power fault repair schemes. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Currently, the formulation of power outage emergency repair plans mainly relies on the traditional model of data collection, manual analysis, and experience-based decision-making, supplemented by partial automation systems. However, its technical bottlenecks are becoming increasingly prominent, and it has certain limitations. First, power fault data exhibits multimodal heterogeneous characteristics, but existing technologies for processing multimodal heterogeneous data tend to isolate single modes, and multimodal data fusion remains at the stage of simple splicing integration, resulting in poor fusion effects.

[0004] Furthermore, existing fault diagnosis and solution generation models mainly fall into two categories: rule-based expert systems and data-driven machine learning models. Expert systems have extremely high rule base maintenance costs, while machine learning models lack accuracy in long-tail faults. Finally, the output results are uninterpretable, failing to meet the power industry's requirements for decision traceability.

[0005] In recent years, large language models have begun to be applied to power fault handling due to their strong semantic understanding capabilities. However, there is a gap between their general knowledge and power professional knowledge, and they have shortcomings such as bias in understanding professional terms, lack of numerical feature processing capabilities, and insufficient adaptation to safety regulations.

[0006] In summary, existing technologies have shortcomings in multimodal data fusion, deep adaptation of domain knowledge, and small sample generalization, making it difficult to meet the needs of modern power distribution networks for rapid, accurate, and safe fault repair. Summary of the Invention

[0007] To address the aforementioned problems, this invention proposes an intelligent generation method and system for power fault repair schemes. This invention can integrate multimodal fault data, accurately understand power industry expertise, and efficiently generate compliant schemes. Through targeted fine-tuning, it injects power industry knowledge into a large model, enabling it to integrate the ability to interpret numerical features, understand textual semantics, and generate professional schemes.

[0008] According to some embodiments, the present invention adopts the following technical solution: A method for intelligently generating power fault repair plans includes the following steps: A multimodal instruction template is constructed, which includes a physical feature prompting segment, a multi-source data input segment, and a task instruction segment. The physical feature prompting segment is used to introduce the basic laws of power faults as prior knowledge. The multi-source data input segment is used to integrate structured telemetry data with unstructured text description key-value pairs and natural language mixed format. The task instruction segment is used to construct a standardized instruction format. We acquire basic knowledge of the power industry, construct a dataset, and use the dataset to train a general large language model. We adopt a hierarchical learning rate strategy to obtain a large language model fine-tuned with domain knowledge. Obtain power failure emergency repair cases, construct power failure emergency repair cases as samples according to the pre-built multimodal instruction template, use semantic loss, physical constraint loss and safety loss as weighted mixed loss functions, and use the samples to optimize the parameters of the large language model after fine-tuning the domain knowledge to obtain the large language model after fine-tuning the multimodal task. The system acquires target data for power fault repair, standardizes it according to a pre-built multimodal instruction template, and processes the standardized data using a large language model fine-tuned by the multimodal task to obtain the final power fault repair plan.

[0009] As an alternative implementation, the physical feature prompt segment is used to set the performance characteristics of corresponding physical quantities under various power fault scenarios, thereby making the physical rules explicit and associating the numerical characteristics of physical quantities with the corresponding faults.

[0010] As an alternative implementation, when constructing the multi-source data input segment, the structured data is constructed in the form of key-value pairs where the feature name equals the corresponding value and unit, and the text description retains the original expression while being standardized and cleaned, and supplemented with key scene information. A modal feature mapping function is introduced to realize the quantitative association between numerical values ​​and semantics, and a mapping relationship is constructed between structured data sets and unstructured text feature sets through a linear transformation matrix.

[0011] As an alternative implementation, the instruction format of the task instruction segment is standardized to output fault type, feature correlation analysis, standardized emergency repair steps, and post-power-on verification indicators based on input information.

[0012] As an alternative implementation, the multimodal instruction template is configured with scene adaptation capability, dynamically adjusting the input items and prompts according to the fault type, equipment type, repair stage, or data integrity.

[0013] As a further limitation, the specific content of dynamically adjusting the input items and prompts based on the fault type, equipment type, emergency repair stage, or data integrity includes: for different fault types, strengthening the input weight of corresponding features, and adding problems that the corresponding fault type is likely to cause and suggested repair solutions to the prompt section; For different device types, add relevant semi-structured data input; For different fault diagnosis stages, the input items are mainly real-time telemetry data and preliminary text descriptions, and the task instructions focus on fault type judgment. During the emergency repair implementation stage, incremental input of executed steps and on-site feedback information is supported, and the task instructions are adjusted to optimize subsequent steps. During the power supply verification stage, real-time data after power supply is added, and the task instructions require judgment on whether normal operation has been restored and potential risks. For scenarios with complete data, maintain the multimodal instruction template; for scenarios with missing data, weaken physical prompts and strengthen text semantic understanding. When adding task instructions, prioritize data collection items.

[0014] As an alternative implementation method, the process of acquiring basic knowledge of the power industry and constructing a dataset includes acquiring terminology definitions, physical laws of faults, and safe operating procedures of the power industry. Based on the acquired data, terminology alignment samples, physical law samples, and safety procedure samples are constructed. The terminology alignment samples are used to define the terminology of the power industry, the physical law samples are used to define the numerical characteristics of different physical quantities and the corresponding fault types, and the safety procedure samples are used to define the safety procedures that need to be followed in different emergency repair scenarios.

[0015] As an alternative implementation method, the process of training a general large language model using a dataset and employing a hierarchical learning rate strategy to obtain a domain-knowledge-fine-tuned large language model includes: = ; in, The initial learning rate, l This is the current network layer number. L This represents the total number of floors. γ The attenuation coefficient is used to make the learning rate of the top layer (semantic output layer) higher than that of the bottom layer (feature extraction layer), thereby enhancing the expressive power of domain knowledge. The fine-tuned model knowledge is distilled into a lightweight model, and the core knowledge is preserved through KL divergence.

[0016] As an alternative implementation, the process of optimizing the parameters of the large language model using sample optimization and domain knowledge fine-tuning with semantic loss, physical constraint loss, and safety loss as weighted hybrid loss functions includes: using the LoRA parameter fine-tuning method to update only the low-rank matrix parameters of the model's attention layer; employing a weighted hybrid loss function during fine-tuning to constrain the model's adherence to physical laws and execution of safety procedures; the weighted hybrid loss function includes semantic loss, physical constraint loss, and safety loss; the semantic loss uses cross-entropy loss to calculate the textual difference between the model output and the standard solution; the physical constraint loss limits the constraints that the model output must include based on the voltage and current characteristics of each fault type, otherwise a penalty is applied; the penalty adopts a feature matching penalty mechanism; and the safety loss is used to apply high-weight penalties to steps that violate safety procedures. Each loss is configured with corresponding weights.

[0017] A smart power fault repair solution generation system includes: The instruction template construction module is configured to construct a multimodal instruction template, which includes a physical feature prompting segment, a multi-source data input segment, and a task instruction segment. The physical feature prompting segment is used to introduce the basic laws of power faults as prior knowledge. The multi-source data input segment is used to integrate structured telemetry data with unstructured text description key-value pairs and natural language mixed format. The task instruction segment is used to construct a standardized instruction format. The domain fine-tuning module is configured to acquire basic knowledge of the power industry, build a dataset, use the dataset to train a general large language model, and adopt a hierarchical learning rate strategy to obtain a large language model fine-tuned with domain knowledge. The multimodal task fine-tuning module is configured to acquire power failure repair cases. According to the pre-built multimodal instruction template, the power failure repair cases are constructed as samples. The parameters of the large language model after fine-tuning are optimized by using semantic loss, physical constraint loss and safety loss as weighted mixed loss functions, and the parameters of the large language model after fine-tuning are optimized by optimizing the domain knowledge of the samples, so as to obtain the large language model after fine-tuning of the multimodal task. The intelligent solution generation module is configured to acquire target data for power fault repair, standardize it according to a pre-built multimodal instruction template, and process the standardized data using a large language model fine-tuned by the multimodal task to obtain the final power fault repair solution.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention innovatively provides an intelligent generation method for power fault repair schemes. By constructing multimodal instruction templates to eliminate data barriers and injecting professional capabilities through domain knowledge enhancement fine-tuning, an intelligent generation system for power fault repair schemes is jointly constructed. This not only solves the pain points of traditional models, such as reliance on experience, low efficiency, and uncontrollable risks, but also provides a reusable technical paradigm for intelligent operation and maintenance of power systems through the generalization ability of large models and the deep integration of domain knowledge. It has significant academic value and engineering practicality.

[0019] This invention innovatively provides an intelligent generation method for power fault repair schemes. By constructing a multimodal instruction template to eliminate data barriers, and by forcing the model to learn cross-modal correlation logic through a standardized input format, it achieves semantic-numerical fusion of heterogeneous data. Furthermore, it adopts a three-stage architecture that combines physical feature prompts, multi-source data integration, and task objective constraints to ensure that the input information is complete and logically clear. Moreover, the multimodal instruction template has scene adaptive capabilities, which can dynamically adjust the input items and prompts according to the fault type, equipment type, and repair stage.

[0020] This invention innovatively provides an intelligent generation method for power fault repair plans. It proposes a two-stage domain-enhanced fine-tuning strategy, which injects power domain knowledge into the model through knowledge injection, rule constraints, and small-sample reinforcement, achieving a leap from general capabilities to professional adaptation. It also performs targeted optimization for specific tasks from multimodal instructions to repair plans, enabling the model to output accurate plans by combining structured data and text descriptions.

[0021] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0022] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0023] Figure 1 This is a schematic diagram of a method for intelligently generating power fault repair solutions according to one embodiment. Detailed Implementation

[0024] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0025] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0026] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0027] Example 1 As described in the background section, power fault data exhibits multimodal heterogeneous characteristics, primarily including structured telemetry data, unstructured text data, and semi-structured equipment data. Among these, structured telemetry data is real-time / historical numerical data collected by distribution automation systems (DAS), containing key electrical quantities such as voltage, current, and power, and is characterized by high accuracy and strong time-series characteristics. These numerical features are the core basis for determining the fault type.

[0028] Unstructured text data includes fault repair records, inspection reports, dispatch instructions, etc., which contain semantic information such as fault location, phenomenon, and handling process. However, its expression is flexible, such as colloquial descriptions and mixed use of professional terms, and lacks a unified format.

[0029] Semi-structured equipment data, such as switch models, line topology, and equipment ledgers, are used to locate the scope of fault impact and formulate repair strategies.

[0030] Existing technologies for processing the aforementioned data exhibit a single-modal, isolated characteristic. Structured data processing relies on traditional machine learning models to classify faults by extracting features such as "current surge amplitude" and "voltage drop slope," but it cannot correlate semantic causes such as "insulation breakdown" or "wire contact" in the text.

[0031] For example, the model can identify "phase b current is 0" as a ground fault, but it cannot distinguish whether it is "insulation breakdown of cable joint" or "conductor touching guy wire".

[0032] Text data processing relies on keyword matching or NLP techniques, which can only achieve a rough classification of fault types and cannot perform refined diagnosis by combining numerical features in telemetry data. Multimodal data fusion remains at the level of splicing integration, such as simply concatenating feature vectors from telemetry data with word vectors from text and then inputting them into the model, without establishing a deep causal relationship between numerical features and semantic features, resulting in poor fusion performance.

[0033] On the other hand, rule-based expert systems typically encode the experience of maintenance personnel into IF-THEN rules to determine the fault type. However, their drawback lies in the extremely high maintenance cost of the rule base. Updates to the power grid topology and equipment types require manual modification of each rule, resulting in a high rule conflict rate for complex power grids containing tens of thousands of lines. Furthermore, this approach has limited coverage, only handling known fault types (such as common single-phase grounding), and failing to address rare faults, which, although accounting for a small percentage, take more than three times longer to process than ordinary faults.

[0034] However, fault diagnosis and solution generation based on data-driven machine learning models rely on manual feature engineering, requiring domain experts to design features such as current harmonic components and voltage imbalance. For novel faults, feature design lags behind actual needs. Power fault data exhibits a long-tailed distribution, resulting in insufficient accuracy of models on long-tailed faults, and the output results are uninterpretable. The model can only provide a judgment but cannot explain the cause, making it difficult to meet the power industry's requirements for traceable decision-making.

[0035] Finally, while large language models such as GPT, LLaMA, and Flan-T5 possess strong semantic understanding capabilities, there is a gap between their general knowledge and the specific expertise required in the power industry. Firstly, this manifests in misunderstandings of specialized terminology: general models are prone to common-sense errors in interpreting power-specific terms. For example, they might interpret a ring network box as a ring-shaped enclosure, ignoring its function as a distribution network node, leading to directional errors in the generated solutions. Secondly, they lack the ability to process numerical features: general models lack sensitivity to numerical features in telemetry data; directly inputting text descriptions reduces the probability of correct model judgments and fails to correlate the temporal relationships of some data. Finally, they are insufficiently adapted to safety regulations: power emergency repairs must adhere to the "Power Safety Work Regulations," but solutions generated by general models often violate basic safety standards. For example, in a cable joint replacement solution, the model might omit the voltage testing step and directly suggest stripping the cable insulation, posing a risk of electric shock.

[0036] To address the above problems, this embodiment provides an intelligent method for generating power fault repair plans, such as... Figure 1 As shown, it includes the following steps: A multimodal instruction template is constructed, which includes a physical feature prompting segment, a multi-source data input segment, and a task instruction segment. The physical feature prompting segment is used to introduce the basic laws of power faults as prior knowledge. The multi-source data input segment is used to integrate structured telemetry data with unstructured text description key-value pairs and natural language mixed format. The task instruction segment is used to construct a standardized instruction format. We acquire basic knowledge of the power industry, construct a dataset, and use the dataset to train a general large language model. We adopt a hierarchical learning rate strategy to obtain a large language model fine-tuned with domain knowledge. Obtain power failure emergency repair cases, construct power failure emergency repair cases as samples according to the pre-built multimodal instruction template, use semantic loss, physical constraint loss and safety loss as weighted mixed loss functions, and use the samples to optimize the parameters of the large language model after fine-tuning the domain knowledge to obtain the large language model after fine-tuning the multimodal task. The system acquires target data for power fault repair, standardizes it according to a pre-built multimodal instruction template, and processes the standardized data using a large language model fine-tuned by the multimodal task to obtain the final power fault repair plan.

[0037] The following is a detailed introduction to each step: The first step is the design of multimodal instruction templates, which involves constructing a unified input of structured data and text semantics.

[0038] The contradiction in power fault data lies in the separation between structured numerical features and unstructured semantic information. Time-series data such as voltage and current collected by telemetry systems and textual descriptions reported by maintenance personnel belong to different modalities. Traditional models struggle to establish deep correlations between the two, leading to fault diagnosis relying on subjective human judgment. To address this, this innovative design utilizes a dedicated multimodal instruction template for power faults. By standardizing the input format, it forces the model to learn cross-modal correlation logic, achieving semantic-numerical fusion of heterogeneous data.

[0039] This embodiment focuses on guiding the model to focus on key associations, and adopts a three-stage architecture of physical feature prompts, multi-source data integration, and task objective constraints to ensure that the input information is complete and logically clear.

[0040] Physical feature hints section: Introducing the basic laws of power failure as prior knowledge reduces the model's dependence on domain experience.

[0041] For example, in the case of a ground fault, the prompt is: According to the known laws of power systems: during a ground fault, the current of the fault phase tends to be close to 0, and the total reactive power may be negative; For equipment failure scenarios, the prompt message is: Equipment aging may lead to a decline in insulation performance, manifested as increased phase current fluctuations and reduced line voltage stability, which needs to be analyzed in conjunction with the number of years of operation.

[0042] By making physical rules explicit, the model is guided to associate numerical characteristics with faults.

[0043] Multi-source data input section: Integrates structured telemetry data with unstructured text descriptions, key-value pairs, and natural language mixed formats.

[0044] The structured data in this embodiment is presented in the form of key-value pairs of “feature name = value + unit”, such as “phase a current = 72.49A, phase b current = 0.0A, line voltage ab = 0V, total reactive power = 0kVar”, to ensure numerical accuracy and clarity of physical quantities.

[0045] The text description retains the original wording while undergoing standardized cleaning (e.g., correcting "the cable is broken" to "there are signs of insulation damage at the cable joint"), and adding key scenario information.

[0046] For example, the integrated input is: "Table data: Phase a current = 6.24A, Phase b current = 0A, Phase c current = 18.24A, Line ab voltage = 23108V, Total reactive power = -11.87kVar; Fault description: The high-voltage conductor of poles #43-#13 is in contact with the guy wire, accompanied by spark discharge. There is a record of strong winds in the past."

[0047] To achieve cross-modal data fusion, this embodiment introduces a modal feature mapping function to realize the quantitative correlation between numerical values ​​and semantics. Let the structured dataset be... Ds ={ , ,..., The unstructured text feature set is: Dt ={ , ,..., }, through linear transformation matrix W ∈ Establish mapping relationships: .

[0048] Wherein, Emb() is the feature embedding function (in this embodiment, structured data uses normalized embedding, and text uses BERT word vectors). σ This is a ReLU activation function that outputs a fused feature vector. This function ensures that cross-modal information such as "phase a current = 0A" and "ground fault" are clustered together in high-dimensional space.

[0049] Task instruction segment: Clearly define the core elements of the model output to avoid generating content that deviates from actual needs. The instruction format is standardized as follows: "Please output the following based on the above information: 1. Fault type; 2. Feature correlation analysis; 3. Standardized emergency repair steps; 4. Verification indicators after power restoration."

[0050] Power outage scenarios are diverse, and fixed templates cannot cover all needs. Therefore, the template constructed in this embodiment has scenario-adaptive capabilities, and can dynamically adjust input items and prompts according to the fault type, equipment type, and repair stage.

[0051] The specific process includes: Fault type adaptation: For ground faults, the input weights for features such as "phase current difference" and "sudden drop in line voltage" are strengthened, and the prompt section adds "Ground faults easily lead to an increase in voltage of non-faulty phases, and the faulty section should be isolated first." For short circuit faults, input items such as "peak short circuit current" and "protection operation time" are added, and the prompt section is supplemented with "Short circuit faults may be accompanied by equipment burnout, and insulation resistance should be tested before emergency repairs." For equipment faults, semi-structured data inputs such as "equipment model," "years of operation," and "historical defect records" are added.

[0052] According to the emergency repair stage: the input items in the fault diagnosis stage are mainly real-time telemetry data and preliminary text descriptions, and the task instructions focus on "fault type judgment"; the emergency repair implementation stage supports incremental input of "executed steps" and "on-site feedback", and the task instructions are adjusted to "optimize subsequent steps"; the power supply verification stage adds "real-time data after power supply", and the task instructions require "judging whether normal has been restored and potential risks".

[0053] Adaptation based on data integrity: The template for data-complete scenarios retains the full three-segment structure. For data-missing scenarios, the physical prompts are automatically weakened, while the semantic understanding of the text is strengthened. Task instructions are updated with "priority data collection items" (e.g., "suggest measuring three-phase current and line voltage").

[0054] Next, we will fine-tune the instructions for domain knowledge enhancement.

[0055] While general-purpose large language models possess strong semantic understanding capabilities, they exhibit significant limitations in power fault scenarios: biased understanding of technical terms, insufficient knowledge of the physical laws governing faults, and poor adherence to safety regulations. To address these limitations, this embodiment proposes a two-stage domain-enhanced fine-tuning strategy. Through a combination of "knowledge injection – rule constraints – small-sample reinforcement," power domain knowledge is injected into the model in a targeted manner, achieving a leap from general-purpose capabilities to specialized adaptation.

[0056] Phase 1: Pre-tuning of domain knowledge to build a foundation of power industry knowledge. The goal of this phase is to enable the model to master the basic knowledge system of the power industry, including terminology definitions, physical laws of faults, and safe operating procedures, laying the foundation for subsequent task fine-tuning.

[0057] Fine-tuning the dataset construction: Based on the "Electric Power Safety Work Regulations" and historical fault cases, three types of core samples were constructed.

[0058] The first category is terminology alignment samples: used to solve the problem of misunderstandings in terms, that is, to define unified and correct terms.

[0059] In this embodiment, the sample format is "Input: What is [term]? Output: [Term definition + function + application scenario]".

[0060] For example: "Input: What is a ring main unit? Output: A ring main unit is a power distribution device used for cable line branching and connection in a power distribution network. It contains 1 to 4 switching units and can realize fault section isolation and load transfer. It is widely used in urban cable networks."

[0061] The second category is physical law samples: used to solve the problem of missing feature associations.

[0062] The sample format in this embodiment is "Input: What kind of fault does [voltage / current characteristic] belong to? Output: [Fault type + feature matching logic]".

[0063] For example: "Input: Phase a current 0A, phase b current 70A, ab line voltage 0V, what kind of fault is this? Output: Phase B ground fault. Logic: The faulty phase (phase B) current is normal, the non-faulty phase (phase A) current is interrupted and the voltage drops sharply, which is consistent with the characteristics of a single-phase ground fault."

[0064] The third category is safety procedure samples: used to address operational violations that are prone to occur during power maintenance.

[0065] The sample format in this embodiment is "Input: [Operation Scenario] What steps need to be performed? Output: [Steps + Safety Basis]".

[0066] For example: "Input: What steps need to be performed before emergency repair? Output: Step 1. Power outage (according to 'Electric Power Safety Work Regulations' 3.2.1); Step 2. Voltage testing (using a qualified voltage tester, according to 'Electric Power Safety Work Regulations' 3.3.2); Step 3. Grounding (connecting the grounding end first, then the conductor end, according to 'Electric Power Safety Work Regulations' 3.4.5); Step 4. Setting up a safety fence (according to 'Electric Power Safety Work Regulations' 4.2.3)."

[0067] Preferably, in some embodiments, the dataset size is controlled between 5,000 and 10,000 records to ensure coverage of core knowledge and avoid redundancy.

[0068] This stage of fine-tuning employs a combined strategy of full-parameter fine-tuning and knowledge distillation. Full-parameter fine-tuning uses the aforementioned dataset to train the base model, employing a hierarchical learning rate strategy, as shown in the formula: = .

[0069] in, l This is the current network layer number. L This represents the total number of floors. γ The attenuation coefficient (set to 0.95 in this embodiment) is used to make the learning rate of the top layer (semantic output layer) higher than that of the bottom layer (feature extraction layer), thereby enhancing the ability to express domain knowledge.

[0070] The initial learning rate is set to Iterate 3-5 times to ensure that the model deeply internalizes domain knowledge.

[0071] Knowledge distillation distills the fine-tuned model knowledge into a lightweight model (such as Flan-T5-small), retaining core knowledge through the "teacher-student" loss function (KL divergence), balancing performance and deployment efficiency.

[0072] The knowledge distillation process employs a temperature-coefficient-corrected KL divergence loss, assuming the teacher model output probability distribution is... Pt (Flan-T5-base), the student model output is Ps (Flan-T5-small), then the distillation loss is: = log

[0073] in, T The temperature coefficient is used to enhance the student model's ability to learn edge knowledge (such as rare fault types) by smoothing the probability distribution.

[0074] The second stage involves fine-tuning the multimodal task to enhance the "input-decision" mapping capability.

[0075] This phase builds upon the knowledge gained in the first phase, focusing on targeted optimization of the specific tasks from multimodal instructions to emergency repair plans, enabling the model to combine structured data and textual descriptions to output accurate solutions.

[0076] Fine-tuning dataset construction: Based on real fault cases, samples are constructed according to the multimodal instruction template format. Each sample contains "multimodal input (physical prompts + telemetry data + text description) + standard output (fault type + repair steps)", as shown in Table 1.

[0077] Table 1. Examples of fine-tuned datasets

[0078] Lightweight fine-tuning technique: Employs the LoRA (Low-Rank Adaptation) parameter efficient fine-tuning method, updating only the low-rank matrix parameters of the model's attention layer.

[0079] The low-rank matrix decomposition satisfies: W= +BA in, For pre-trained weights, B∈ A∈ It is a low-rank matrix (rank r=16), with parameter size from Reduced to 2dr.

[0080] Its advantages are as follows: First, it reduces computational costs. Compared to full parameter fine-tuning, the GPU memory required for training is reduced by 90%, and can be completed with a single GPU. Second, it avoids overfitting. In scenarios with small sample sizes, fine-tuning only key layers can preserve the model's general capabilities and prevent over-memorization of limited samples. Finally, it facilitates model fusion. The fine-tuned LoRA weights can be flexibly separated from the base model, supporting quick switching between fine-tuned versions for different scenarios.

[0081] Loss function design: Traditional fine-tuning only uses cross-entropy loss, focusing on the accuracy of text output. This embodiment designs a weighted hybrid loss function, which simultaneously constrains the model to comply with physical laws and execute safety procedures.

[0082] The weighted mixture loss function specifically includes: Semantic loss (L_semantic): Cross-entropy loss is used to calculate the text difference between the model output and the standard solution. In this embodiment, the weight coefficient of this part is 0.4. Physical constraint loss (L_phy): Hard constraints are designed for voltage and current characteristics, such as "If the fault type is B-phase grounding, the model output must include the correlation analysis of "b-phase current ≈ 0A", otherwise a penalty is imposed. In this embodiment, the weight coefficient of this part is 0.3.

[0083] The penalty employs a feature matching penalty mechanism. Let S be the set of feature associations output by the model, and the standard feature set be... ,but:

[0084] Here, λ=10 is the penalty coefficient, which will increase sharply when the model omits key features such as "phase b current ≈ 0A".

[0085] Safety Loss (L_safety): A high-weight penalty is imposed on steps that violate safety procedures. A risk level weighting factor is introduced, where the risk level of the step is 1. ∈[1,5], the violation indicator variable is If ∈{0,1}, then: =

[0086] Ensure that high-risk steps such as "operating without verifying electricity" are avoided. =5) The penalty is more than 10 times that of low-risk steps.

[0087] In this embodiment, the weighting coefficient for this part is 0.3, and the penalty value increases with the step risk level. The total loss function is: L_total =0.4×L_semantic+0.3×L_phy+0.3×L_safety. This fine-tuning strategy overcomes the technical bottleneck of adapting general large-scale models by injecting knowledge layers and customizing loss functions.

[0088] The two-stage design first solidifies the foundation and then focuses on the task, avoiding interference between terminology learning and decision-making ability, and solving the problem of knowledge fragmentation in traditional fine-tuning.

[0089] In this embodiment, the LoRA technique is combined with a weighted hybrid loss function, enabling the model to achieve high accuracy even in small sample scenarios and reducing reliance on labeled data. Physical constraints and safety losses ensure that the model output is not only semantically fluent but also conforms to electrical laws and safety regulations, thus addressing the shortcomings of general models that appear reasonable but are actually illegal.

[0090] Example 2 A smart power fault repair solution generation system includes: The instruction template construction module is configured to construct a multimodal instruction template, which includes a physical feature prompting segment, a multi-source data input segment, and a task instruction segment. The physical feature prompting segment is used to introduce the basic laws of power faults as prior knowledge. The multi-source data input segment is used to integrate structured telemetry data with unstructured text description key-value pairs and natural language mixed format. The task instruction segment is used to construct a standardized instruction format. The domain fine-tuning module is configured to acquire basic knowledge of the power industry, build a dataset, use the dataset to train a general large language model, and adopt a hierarchical learning rate strategy to obtain a large language model fine-tuned with domain knowledge. The multimodal task fine-tuning module is configured to acquire power failure repair cases. According to the pre-built multimodal instruction template, the power failure repair cases are constructed as samples. The parameters of the large language model after fine-tuning are optimized by using semantic loss, physical constraint loss and safety loss as weighted mixed loss functions, and the parameters of the large language model after fine-tuning are optimized by optimizing the domain knowledge of the samples, so as to obtain the large language model after fine-tuning of the multimodal task. The intelligent solution generation module is configured to acquire target data for power fault repair, standardize it according to a pre-built multimodal instruction template, and process the standardized data using a large language model fine-tuned by the multimodal task to obtain the final power fault repair solution.

[0091] The specific processes for each module, as described in Example 1, will not be repeated here.

[0092] 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 one or more computer-usable storage media (including, but not limited to, disk storage, etc.) containing computer-usable program code. CD - ROM It takes the form of a computer program product implemented on (such as optical memory, etc.).

[0093] 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.

[0094] 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.

[0095] These computer program instructions can 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.

[0096] The above description is merely a preferred embodiment of the present invention and is not intended to limit the 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 by those skilled in the art without creative effort within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A power failure repair scheme intelligent generation method, characterized in that, The method comprises the following steps: A multi-modal instruction template is constructed, which comprises a physical feature prompt segment, a multi-source data input segment and a task instruction segment, the physical feature prompt segment is used to introduce the basic law of power failure as prior knowledge, the multi-source data input segment is used to integrate structured telemetry data and unstructured text description key-value pairs and natural language mixed format, and the task instruction segment is used to construct a standardized instruction format; Basic knowledge of the power industry is obtained, a data set is constructed, a general large language model is trained using the data set, a hierarchical learning rate strategy is adopted, and a large language model after domain knowledge fine-tuning is obtained; Power failure repair cases are obtained, and the power failure repair cases are constructed into samples according to the pre-constructed multi-modal instruction template, a weighted mixed loss function is used, which comprises semantic loss, physical constraint loss and safety loss, the parameters of the large language model after domain knowledge fine-tuning are optimized using the samples, and a large language model after multi-modal task fine-tuning is obtained; Power failure repair target data is obtained, and the target data is standardized according to the pre-constructed multi-modal instruction template, the standardized data is processed using the large language model after multi-modal task fine-tuning, and a final power failure repair scheme is obtained. The physical feature prompt segment is used to set the performance characteristics of corresponding physical quantities under corresponding power failure scenarios for various power failure scenarios, to realize explicitness of physical rules, and to associate the numerical characteristics of the physical quantities with the corresponding failures.

2. The intelligent generation method of a power failure repair scheme according to claim 1, characterized in that, When the multi-source data input segment is constructed, the structured data is constructed in the form of a key-value pair with a feature name equal to a corresponding numerical value and a unit, the text description is standardized and cleaned while the original expression is retained, and key scene information is supplemented; 3. The intelligent generation method of a power failure repair scheme according to claim 1, characterized in that, A modal feature mapping function is introduced to realize quantitative association of numerical values and semantics, and a structured data set and an unstructured text feature set are constructed into a mapping relationship through a linear transformation matrix. The instruction format of the task instruction segment is standardized as an instruction format based on input information, output failure type, feature association analysis, standardized repair steps and post-delivery verification indicators.

4. The intelligent generation method of a power failure repair scheme according to claim 1, characterized in that, The multi-modal instruction template is configured with scene self-adaptation capability, and input items and prompt contents are dynamically adjusted according to failure types, equipment types, repair stages or data integrity.

5. The intelligent generation method of power failure repair scheme according to claim 1, characterized in that, The specific contents of dynamically adjusting input items and prompt contents according to failure types, equipment types, repair stages or data integrity include: for different failure types, the input weight of the corresponding feature is respectively strengthened, and the prompt segment increases problems and recommended repair schemes that are easily caused by the corresponding failure type; 6. The intelligent generation method of a power failure repair scheme according to claim 5, characterized in that, For different equipment types, related semi-structured data input is added; For different failure diagnosis stages, real-time telemetry data and preliminary text description are mainly used as input items, the task instruction focuses on failure type judgment, the repair implementation stage supports incremental input of executed steps and field feedback information, and the task instruction is adjusted to optimize subsequent steps; real-time data after power transmission is added in the power transmission verification stage, and the task instruction requires to judge whether it is restored to normal and potential risks; For data complete scenes, the multi-modal instruction template is maintained, for data missing scenes, the physical prompt segment is weakened, the text semantic understanding is strengthened, and the task instruction adds data collection items that need to be recommended in priority. ​ 7. The intelligent generation method of power failure repair scheme according to claim 1, characterized in that, The process of acquiring the basic knowledge of the power industry and constructing the dataset includes acquiring the definition of the terms of the power industry, the physical law of failure and the safety operation specification, and based on the acquired data, constructing a term alignment sample, a physical law sample and a safety procedure sample, wherein the term alignment sample is used to define the terms of the power industry, the physical law sample is used to define the numerical characteristics of different physical quantities and the corresponding failure types, and the safety procedure sample is used to define the safety procedures that need to be followed in different repair scenes.

8. The intelligent generation method of a power failure repair scheme according to claim 1, characterized in that, The process of training a general large language model using a data set, adopting a hierarchical learning rate strategy, and obtaining a large language model after domain knowledge fine-tuning includes: = ; wherein, is the initial learning rate, l is the current network layer number, L is the total number of layers, γ is the decay coefficient, the learning rate of the top layer, i.e., the semantic output layer, is higher than that of the bottom layer, i.e., the feature extraction layer, the expression ability of the domain knowledge is strengthened, the model knowledge after fine-tuning is distilled to the lightweight model, and the core knowledge is preserved through KL divergence.

9. The intelligent generation method of a power failure repair scheme according to claim 1, characterized in that, The process of using the sample to optimize the parameters of the large language model after the field knowledge fine-tuning includes: using the LoRA parameter fine-tuning method to update only the low-rank matrix parameters of the model attention layer, and using the weighted mixed loss function in the fine-tuning process to constrain the model to comply with the physical law and execute the safety procedure, wherein the weighted mixed loss function includes semantic loss, physical constraint loss and safety loss, wherein the semantic loss uses cross-entropy loss to calculate the text difference between the model output and the standard scheme, the physical constraint loss limits the constraints that must be included in the model output for the voltage and current characteristics of each failure type, otherwise a penalty is imposed, and the penalty uses a feature matching penalty mechanism, and the safety loss is used to impose a high weight penalty on steps that violate safety procedures, and each loss is configured with a corresponding weight.

10. A power failure rush-repair scheme intelligent generation system, characterized in that, Comprise: An instruction template construction module configured to construct a multi-modal instruction template, the multi-modal instruction template comprising a physical feature prompt segment, a multi-source data input segment and a task instruction segment, the physical feature prompt segment being used to introduce the basic law of power failure as prior knowledge, the multi-source data input segment being used to integrate structured telemetry data and unstructured text description in a key-value pair and natural language mixed format, and the task instruction segment being used to construct a standardized instruction format; A field fine-tuning module configured to acquire the basic knowledge of the power industry, construct a dataset, train a general large language model using the dataset, and use a hierarchical learning rate strategy to obtain a large language model after fine-tuning of field knowledge; A multi-modal task fine-tuning module configured to acquire power failure repair cases, construct the power failure repair cases into samples according to the pre-constructed multi-modal instruction template, use a weighted mixed loss function of semantic loss, physical constraint loss and safety loss to optimize the parameters of the large language model after fine-tuning of field knowledge using the samples, and obtain a large language model after multi-modal task fine-tuning; A scheme intelligent generation module configured to acquire power failure repair target data, standardize the power failure repair target data according to the pre-constructed multi-modal instruction template, process the standardized data using the large language model after multi-modal task fine-tuning, and obtain a final power failure repair scheme.

Images