Power supply repair material reservation recommendation method and system based on large model agent

Abstract

The application relates to a power supply repair material reservation recommendation method and system based on a large model intelligent agent. The method comprises the following steps: a task analysis intelligent agent calls a large language model to perform semantic analysis on a repair task description, and finds similar cases from a historical repair case library; a recommendation decision intelligent agent constructs a multi-round reasoning process for interaction with the large language model through Few-shot prompt words, and outputs a repair material recommendation list through the large language model; the recommendation decision intelligent agent and a material inventory system establish an interface linkage mechanism to generate a repair material replacement list; a feedback optimization intelligent agent generates a recommendation explanation and collects user feedback data, and optimizes the large language model according to the user feedback data. Through semantic reasoning and historical case analogy by calling the large language model through the task analysis intelligent agent, a specific repair scene can be more accurately matched, the first reservation hit rate is improved, the repeated scheduling rate is reduced, and the material preparation efficiency and the decision quality are improved.

Description

Technical Field

[0001] This invention relates to the field of power supply emergency repair material recommendation technology, and in particular to a method and system for power supply emergency repair material stocking recommendation based on a large model intelligent agent. Background Technology

[0002] In the current operation and maintenance of the power system, power supply emergency repair is a crucial link in ensuring the safe and stable operation of the power grid, characterized by its suddenness, short response time, and high workload. Especially in environments with frequent natural disasters (such as typhoons, thunderstorms, high temperatures, rain, and snow) or equipment aging and malfunctions, the number of emergency repair tasks increases significantly, placing higher demands on material allocation, repair efficiency, and personnel response capabilities. In traditional workflows, the preparation of materials needed for power supply emergency repairs mainly relies on the experience and judgment of dispatchers or operation and maintenance personnel. They make preliminary estimates of the required spare parts list based on factors such as the type of accident, the task area, and historical records, and submit this list to the warehouse management system for material distribution. However, most existing material recommendation systems only provide fixed templates or simple rules, lacking a comprehensive understanding of complex semantic information such as task context, equipment status, and geographical location. Furthermore, when facing the same type of accident but occurring in different areas or at different times (such as night / rainy days), the types and quantities of materials required often differ, making it difficult for static rules to adapt to diverse situations. Existing material recommendation systems also suffer from low utilization of historical data and insufficient collaboration with the warehouse system.

[0003] With the development of artificial intelligence, especially Large Language Model (LLM) technology, AI systems have made significant breakthroughs in understanding natural language, extracting complex semantics, and performing reasoning tasks. LLMs can not only parse rich context in human language but also perform case analogies and recommendation generation driven by multiple examples (few-shot learning), demonstrating strong generalization capabilities. Meanwhile, multi-agent systems have shown good structure and scalability in complex task decomposition, asynchronous collaboration, and task division. Currently, some universities and power companies are attempting to apply knowledge graphs, rule engines, or expert-system-based decision-making solutions to power material management scenarios. However, these solutions generally suffer from problems such as "rigid rules, poor scalability, and difficulty in generalization," and have not yet truly achieved the integration of complex semantic modeling and dynamic reasoning capabilities.

[0004] Therefore, due to the large regional differences in power grids, traditional methods of recommending emergency power supply materials often cannot be deployed on demand in different power supply units, nor can they quickly adapt to different regional rules, historical data structures, and storage system interfaces. They often rely heavily on manual labor, resulting in low efficiency in material preparation and decision-making quality during the emergency power supply process, leading to excessively high costs. Summary of the Invention

[0005] In order to solve the above-mentioned technical problems, a method and system for recommending the preparation of power supply emergency repair materials based on a large model intelligent agent is provided. This method can reduce reliance on manual labor, improve the efficiency of material preparation and decision-making quality, and reduce costs.

[0006] A method for recommending the stockpiling of power supply emergency repair materials based on a large-scale intelligent agent model, the method comprising: The task parsing agent receives the emergency repair task description, calls the large language model to perform semantic parsing on the emergency repair task description, extracts multi-dimensional context information and constructs a semantic representation vector, and finds similar cases from the historical emergency repair case library based on the semantic representation vector. Based on the similar cases and multi-dimensional context information, the recommendation decision-making agent constructs a multi-round reasoning process that interacts with the large language model through Few-shot prompts, and outputs a recommended list of emergency repair materials through the large language model. The recommendation decision-making agent establishes an interface linkage mechanism with the material inventory system. When the materials in the emergency repair material recommendation list cannot be dispatched, the recommendation decision-making agent uses the big language model to perform upstream and downstream substitution analysis on the data in the material inventory system and generates an emergency repair material substitution list. The feedback optimization agent generates recommended descriptions based on the emergency repair material replacement list, collects user feedback data, and optimizes the large language model based on the user feedback data.

[0007] In one embodiment, a large language model is invoked to perform semantic parsing on the emergency repair task description, extracting multi-dimensional contextual information and constructing a semantic representation vector, including: The task parsing agent uses prompt words based on a large language model to perform structured extraction of the emergency repair task description, obtains various semantic fields, and extracts multi-dimensional contextual information based on each semantic field. The semantic fields are concatenated into a scene description text, encoded into a semantic representation vector by the large language model, and the model's reasoning dimension is enhanced by adding knowledge templates.

[0008] In one embodiment, similar cases are found from the historical emergency repair case database based on the semantic representation vector, including: The task parsing agent determines the vector database based on a historical emergency repair case library; The semantic representation vectors are semantically similar to each other in the vector database, and the matched vectors are sorted based on the semantic matching degree. Determine the filtering data, and find the corresponding similar vectors from the sorted matching vectors based on the filtering data, and find similar cases based on the similar vectors.

[0009] In one embodiment, the recommendation decision-making agent, based on the similar cases and multi-dimensional contextual information, constructs a multi-turn reasoning process that interacts with the large language model through Few-shot prompts, and outputs a recommended list of emergency repair materials through the large language model, including: The similar cases are labeled as semantic reference standards. The recommendation decision-making agent uses the semantic reference standards to make analogies to the emergency repair scenarios and obtain the analogy results. The recommendation decision-making agent inputs the category results and multi-dimensional context information into the large language model, and performs multi-round reasoning based on scene semantics through the large language model, and constructs a multi-round reasoning process; The large language model outputs a recommended list of emergency repair materials based on a multi-round reasoning process.

[0010] In one embodiment, the large language model outputs a recommended list of emergency repair materials based on a multi-turn reasoning process, including: The large language model describes the image factors of the emergency repair task based on a multi-round reasoning process; Based on the multi-round reasoning process and the influencing factors, the target procedures of the emergency repair task are determined, and corresponding emergency repair materials are matched for each target procedure. All emergency repair materials are collected into a material dataset, and the material dataset is checked to output a recommended list of emergency repair materials after the check.

[0011] In one embodiment, the recommendation decision-making agent establishes an interface linkage mechanism with the material inventory system. When materials in the emergency repair material recommendation list cannot be dispatched, the recommendation decision-making agent uses the large language model to perform upstream and downstream substitution analysis on the data in the material inventory system to generate an emergency repair material substitution list, including: The recommendation decision-making intelligent agent queries the interface and communicates with the material inventory system through the interface; The recommendation decision-making agent searches for materials in the emergency repair material recommendation list in the material inventory system through an interface; When the materials in the emergency repair material recommendation list are insufficient or cannot be dispatched, the large language model is used to perform upstream and downstream substitution analysis on the data in the material inventory system to find various alternative materials. Prioritize each of the candidate materials, generate a recommended ranking table of candidate materials, and determine a replacement list of emergency repair materials based on the recommended ranking table of candidate materials.

[0012] In one embodiment, the method further includes: The recommendation decision-making agent selects alternative materials from the candidate material recommendation ranking table and generates a recommendation report description.

[0013] In one embodiment, the feedback optimization agent generates recommended descriptions based on the emergency repair material replacement list, including: The feedback optimization agent generates recommended descriptions corresponding to the emergency repair material replacement list through the large language model. The recommendations include the recommended materials, the basis for the recommendations, the recommendation process, the recommended alternatives, and the recommendation priority.

[0014] In one embodiment, the method further includes: The feedback optimization agent determines each user role and matches the corresponding role display view according to each user role; The list of alternative repair materials and recommended instructions are displayed visually according to the view of each role.

[0015] A power supply emergency repair material warehousing recommendation system based on a large model intelligent agent, the system comprising: The task parsing agent is used to receive the emergency repair task description, call the large language model to perform semantic parsing on the emergency repair task description, extract multi-dimensional context information and construct a semantic representation vector, and find similar cases from the historical emergency repair case library based on the semantic representation vector. The recommendation decision-making agent is used to construct a multi-round reasoning process that interacts with the big language model based on the similar cases and multi-dimensional context information, and outputs a recommended list of emergency repair materials through the big language model. The recommendation decision-making agent is also used to establish an interface linkage mechanism with the material inventory system. When the materials in the emergency repair material recommendation list cannot be dispatched, the recommendation decision-making agent uses the big language model to perform upstream and downstream substitution analysis on the data in the material inventory system and generates an emergency repair material substitution list. The feedback optimization agent is used to generate recommended descriptions based on the emergency repair material replacement list, collect user feedback data, and optimize the large language model based on the user feedback data.

[0016] The aforementioned power supply emergency repair material preparation recommendation method and system based on a large model intelligent agent, through the task parsing intelligent agent calling a large language model to conduct semantic reasoning and analogy with historical cases, can more accurately match specific emergency repair scenarios, improve the initial preparation hit rate, reduce the duplicate scheduling rate, and eliminate the need for manual operation, thus reducing the burden on scheduling personnel. The recommendation decision intelligent agent generates a material list by using Few-shot analogy learning. The recommendation decision intelligent agent establishes an interface linkage mechanism with the material inventory system. During the recommendation process, factors such as the current inventory of materials, call path, scheduling time, and substitutability are fully considered, thereby improving the efficiency of material preparation and the quality of decision-making. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the application environment and system structure of a power supply emergency repair material preparation recommendation method based on a large model intelligent agent in one embodiment; Figure 2 This is a flowchart illustrating a power supply emergency repair material preparation recommendation method based on a large model intelligent agent in one embodiment; Figure 3 This is a flowchart illustrating a power supply emergency repair material preparation recommendation method based on a large model intelligent agent in another embodiment. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0019] The power supply emergency repair material stocking recommendation method based on large model intelligent agents provided in this application embodiment can be applied to, for example, Figure 1 The application environment shown. For example... Figure 1As shown, the application environment includes a task parsing agent 110, a recommendation decision agent 120, and a feedback optimization agent 130, and the three agents are interconnected and communicate with each other. The task parsing agent 110 receives the emergency repair task description, calls a large language model to perform semantic parsing on the description, extracts multi-dimensional contextual information, and constructs a semantic representation vector. Based on the semantic representation vector, it searches for similar cases in the historical emergency repair case database. The recommendation decision agent 120, based on similar cases and multi-dimensional contextual information, constructs a multi-round reasoning process that interacts with the large language model using Few-shot prompts, and outputs a recommended list of emergency repair materials through the large language model. The recommendation decision agent 120 establishes an interface linkage mechanism with the material inventory system. When materials in the recommended list cannot be dispatched, the recommendation decision agent uses the large language model to perform upstream and downstream substitution analysis on the data in the material inventory system, generating a replacement list of emergency repair materials. The feedback optimization agent 130 generates recommendation instructions based on the replacement list of emergency repair materials, collects user feedback data, and optimizes the large language model based on the user feedback data.

[0020] In one embodiment, such as Figure 2 As shown, a method for recommending the stockpiling of power supply emergency repair materials based on a large model intelligent agent is provided, including the following steps: Step 202: The task parsing agent receives the emergency repair task description, calls the large language model to perform semantic parsing on the emergency repair task description, extracts multi-dimensional context information and constructs a semantic representation vector, and finds similar cases from the historical emergency repair case library based on the semantic representation vector.

[0021] The task parsing agent is mainly responsible for receiving emergency repair work orders or task descriptions, i.e., emergency repair task descriptions. It calls a large language model to perform semantic parsing on the text, extracting multi-dimensional contextual information such as task location, fault type, equipment information, weather environment, and task urgency, and constructing a semantic representation vector. Subsequently, based on the semantic vector, it recalls typical cases with similar structures from the historical emergency repair case library to form preparation examples that can be used for reasoning reference.

[0022] In one embodiment, the power supply emergency repair material preparation recommendation method based on a large model intelligent agent may further include a semantic parsing process performed by a task parsing intelligent agent. The specific process includes: the task parsing intelligent agent extracting the emergency repair task description in a structured manner using prompt words based on a large language model to obtain various semantic fields, and extracting multi-dimensional contextual information based on each semantic field; concatenating the various semantic fields into a scene description text, encoding it into a semantic representation vector through a large language model, and enhancing the model's reasoning dimension by adding knowledge templates.

[0023] By introducing a large language model-driven parsing mechanism, combined with a task parsing agent, comprehensive scenario modeling of emergency repair tasks can be achieved. Specifically, the task parsing agent first receives text input of the emergency repair task, which may include: emergency repair work orders, such as a 10kV Ma'anshan branch line break located near Provincial Highway S303, with trees damaging the line, requiring immediate handling; speech-to-text content, such as dispatch instructions converted to text; and non-standard text content from IM / message systems.

[0024] Next, the task parsing agent can invoke a large language model for natural language understanding, using prompt words to extract structured text and extract various semantic fields. These semantic fields can include: fault type (e.g., line breakage, transformer explosion, drop fuse burnout); fault location (e.g., latitude and longitude, landmark, line name); faulty equipment (e.g., 10kV overhead line, low-voltage distribution box, pole-mounted transformer); time elements (e.g., time period, day / night, holidays); environmental context (e.g., heavy rain, strong winds, mountainous areas, difficult roads); and urgency level (e.g., whether it affects important users, whether it impacts people's livelihoods).

[0025] In this embodiment, a sample prompt snippet can be provided as follows: You are a power emergency repair expert. Please extract key information from the following text: Task Description: 10kV Ma'anshan branch line is broken on Provincial Highway S303, suspected to be due to a tree falling on it. The location is remote, and the mountain roads are slippery. Output: Fault type, faulty equipment, fault location, environmental characteristics, and whether it is an emergency task.

[0026] The structured semantic fields extracted by the task parsing agent will be further encoded into task semantic representation vectors for subsequent historical case matching and few-shot inference input. Specifically, this can include two processes: field concatenation and LLM embedding vector encoding, and knowledge-guided augmented modeling. Field concatenation and LLM embedding vector encoding concatenates multiple fields into a scene description text, which is then fed into a language model to generate embeddings. Knowledge-guided augmented modeling can add knowledge templates to enhance the inference dimension. For example, the task parsing agent can concatenate the task semantics as: broken road in mountainous area, located on S303 provincial highway, trees pressing on the road after heavy rain, slippery road, emergency task; then feed it into a large language model for embedding calculation, used for vector similarity comparison with historical tasks.

[0027] In one embodiment, the power supply emergency repair material preparation recommendation method based on a large model intelligent agent may further include a case recall process performed by a task parsing intelligent agent. The specific process includes: the task parsing intelligent agent determining a vector database based on a historical emergency repair case library; performing semantic similarity matching on the semantic representation vectors in the vector database, and sorting the matched vectors based on the semantic matching degree; determining the filtering data, and finding the corresponding similar vectors from the sorted matching vectors based on the filtering data, and finding similar cases based on the similar vectors.

[0028] The task parsing agent, based on the task semantic representation vector, recalls multiple similar work order records and their corresponding material lists from a historical emergency repair case database. The recall method employs: semantic similarity matching using vector databases (such as FAISS and Weaviate); supports Top-K retrieval + reranking schemes to optimize semantic matching ranking; and supports structural condition filtering (such as region, line type, equipment model, etc.). The recalled similar cases will serve as the Few-shot prompt input for the large language model in the subsequent recommendation inference stage, providing data support for personalized recommendations.

[0029] Step 204: The recommendation decision-making agent constructs a multi-round reasoning process that interacts with the big language model based on similar cases and multi-dimensional contextual information, and outputs a recommended list of emergency repair materials through the big language model.

[0030] The recommendation decision-making agent is responsible for integrating the current task context with historical cases, constructing a reasoning process that interacts with the large language model using Few-shot prompts, and outputting a list of highly relevant emergency repair materials. The material list includes not only the material name, specifications, and recommended quantity, but also priority, alternative solutions, and usage scenario descriptions.

[0031] In one embodiment, the power supply emergency repair material stocking recommendation method based on a large model intelligent agent may further include a large language model reasoning process. The specific process includes: similar cases are labeled as semantic reference standards, the recommendation decision intelligent agent performs emergency repair scenario analogy based on the semantic reference standards, and obtains the analogy results; the recommendation decision intelligent agent inputs the category results and multi-dimensional context information into the large language model, and performs multi-round reasoning based on the scenario semantics through the large language model, and constructs a multi-round reasoning process; the large language model outputs an emergency repair material recommendation list based on the multi-round reasoning process.

[0032] The recommendation decision agent adopts the Few-shot prompting learning technique, which guides the large language model to draw analogies with historical emergency repair scenarios to achieve reasoned recommendation under complex semantics, and constructs high-quality, structurally consistent, and scenario-matching Few-shot example inputs.

[0033] Specifically, the recommendation decision-making agent automatically selects cases from the historical work order database using a Top-K matching method based on similar semantic vectors. These cases are most closely related to the current task context, such as similar fault type, similar environment, and similar equipment. Examples are explicitly labeled with tags such as "fault category," "equipment location," "environmental conditions," and "operational restrictions" to unify semantic reference standards. Examples use a preset "task description + recommendation result" structure, concatenated as LLM prompt input, with the current task placed at the end to guide the model to focus on the current context. In this embodiment, the example prompt is: Task: Thunderstorm weather, 10kV line switch burns out, occurring in a hilly area with inconvenient transportation; Recommended materials: 1 set of outdoor high-voltage disconnect switch (GW4-10 type), 1 aerial work platform vehicle (trailer type), 2 pairs of rubber insulating boots. Task: Sudden tripping at night, possibly due to a burnt-out pole-mounted transformer fuse, located at the end of a rural line, on a muddy road; Please list the recommended materials, indicating the quantity, usage scenario, and reason for recommendation. In this embodiment, through contextual and analogy-guided prompting strategies, the Large Language Model (LLM) can infer the necessary stocking elements for the current task based on scene semantics.

[0034] To enable large language models to go beyond keyword matching or sentence generation and achieve recommendation reasoning that is "structural, logical, and interpretable," a multi-round reasoning mechanism is designed, combined with Chain-of-Thought (CoT) to guide the model's output reasoning process. In one embodiment, a power supply emergency repair material stocking recommendation method based on a large model agent may further include multi-round reasoning and inventory verification processes. Specifically, the process includes: the large language model describing the influencing factors of the emergency repair task based on the multi-round reasoning process; determining the target procedures of the emergency repair task based on the multi-round reasoning process and influencing factors, and matching corresponding emergency repair materials for each target procedure; aggregating all emergency repair materials into a material dataset, checking the material dataset, and outputting a recommended list of emergency repair materials after verification.

[0035] The multi-round reasoning path can include: The first round is task scenario analysis, where the large language model describes potential influencing factors of the current task, such as road conditions, weather impacts on repair efficiency, and operational safety requirements. The second round is work procedure breakdown, where the large language model lists the key procedures to be completed during repair (e.g., power outage—removal of old equipment—replacement of new components—repackaging and restoration—testing and power restoration), matching the required material types for each step. The third round is equipment matching and recommendation generation, outputting the name, quantity, priority, and rationale for each recommended material. The fourth round is safety / standardization review, mainly checking the compliance / standardization of the material list, such as whether it meets the State Grid material standard codes and whether it meets electrical safety requirements.

[0036] For example, the reasoning steps could be: due to a blown fuse, power needs to be cut off and components replaced; the site is muddy, and the operation is at night, requiring lighting and anti-slip equipment; the recommended list includes: one set of RW10 drop-out fuses (to replace the blown-out equipment), two pairs of rubber insulated gloves (safety requirements for high-voltage work at night), two portable lights (for nighttime illumination), and one shovel (to clear fallen branches). In other words, through multi-round, modular prompts, the large language model can form an interpretable and auditable "logically complete path" for making recommendations.

[0037] Step 206: The recommendation decision-making agent establishes an interface linkage mechanism with the material inventory system. When the materials in the emergency repair material recommendation list cannot be dispatched, the recommendation decision-making agent uses a large language model to perform upstream and downstream substitution analysis on the data in the material inventory system and generates an emergency repair material substitution list.

[0038] The recommendation decision-making intelligence is also used to connect with local warehousing systems or material ERP platforms to verify the inventory availability of recommended materials and automatically generate alternative allocation plans when there is a shortage or insufficient inventory.

[0039] In one embodiment, a power supply emergency repair material stocking recommendation method based on a large model intelligent agent may further include a process of generating an emergency repair material alternative list. The specific process includes: the recommendation decision-making intelligent agent querying an interface and communicating with the material inventory system via the interface; the recommendation decision-making intelligent agent searching for materials in the material inventory system that are on the emergency repair material recommendation list via the interface; when materials in the emergency repair material recommendation list are insufficient or cannot be scheduled, using a large language model to perform upstream and downstream substitution analysis on the data in the material inventory system to identify various alternative materials; prioritizing each alternative material to generate an alternative material recommendation ranking table, and determining the emergency repair material alternative list based on the alternative material recommendation ranking table.

[0040] In particular, if the recommendation results output by the large language model cannot form a closed loop with the actual inventory, it can easily lead to scheduling failures or duplicate order assignments. Therefore, the recommendation decision-making agent can establish an interface linkage mechanism with the enterprise's existing inventory system to ensure the feasibility of the recommendations. API integration methods can include query interfaces and returned fields.

[0041] The materials inventory system includes data such as multi-warehouse distribution, estimated transportation time, material batches, expiration dates, and allocation priorities, facilitating subsequent recommendations for alternative materials. Materials requiring recommendation can be uniformly marked with information such as "ready to be allocated immediately," "inter-warehouse transfer," "ETA: 3 hours," "no stock available," "suggested alternative," or "allocated and locked." This marking ensures that recommendations go beyond mere paper suggestions, providing execution paths, logistical loops, and delivery expectations.

[0042] When a recommended material is in short supply or cannot be dispatched, instead of relying on fixed rules, a Large Language Model (LLM) can be used to perform upstream and downstream substitution analysis to form a flexible redundancy solution. For example, the current recommendation is a GW4-10 high-voltage disconnector switch, which is out of stock in the Hefei South Warehouse. The reasoning process is as follows: search for compatible models GW4-12 and GN30-10; determine interface compatibility, installation dimensions, and voltage level; check inventory: two GN30-10 sets are available in the Chizhou East Warehouse; output suggestion; alternative recommendation: one GN30-10 disconnector switch (interface compatible, same voltage level, estimated delivery time 2.5 hours).

[0043] The Large Language Model (LLM) can also prioritize tasks based on accident risk levels. High priority includes tasks affecting mainline power supply / hospitals / government units; medium priority includes residential areas, but temporary power transfer is possible; and low priority includes non-core customers with no strict time constraints. A recommended ranking table can be generated based on three dimensions: priority level, accessibility, and substitutability, facilitating rapid decision-making by dispatchers.

[0044] In one embodiment, a power supply emergency repair material stocking recommendation method based on a large model intelligent agent may further include a process of generating a specification, specifically including: the recommendation decision intelligent agent selects alternative materials from the alternative material recommendation ranking table and generates a recommendation report specification.

[0045] The recommendation results are output as structured data by the recommendation decision agent, supporting JSON format output for inter-system integration; tables and text reports are used for manual review; and a recommendation report instruction manual is generated for approval and archiving.

[0046] Step 208: The feedback optimization agent generates recommended descriptions based on the emergency repair material replacement list, collects user feedback data, and optimizes the large language model based on the user feedback data.

[0047] In high-risk, high-responsibility scenarios such as power supply emergency repairs, in addition to providing accurate material recommendation lists, it is also necessary to possess strong interpretability and interactivity, enabling dispatchers, material handlers, and other users to quickly understand the basis for the recommendations, judge their rationality, and make fine-tuning or feedback when necessary. The feedback optimization intelligent agent is specifically responsible for the interpretation and generation of recommended content, recording user interaction feedback, and the self-optimization and continuous evolution of the large language model, constructing a supervised, adjustable, and evolvable intelligent recommendation feedback closed-loop system.

[0048] After the scheduler confirms the recommendation results, the feedback optimization agent is responsible for generating recommendation explanations, including the reasons for the recommendation, risk warnings, and recommendation logic, thereby improving the efficiency of human-machine collaboration. At the same time, it records user feedback such as modification, acceptance, and rejection, continuously optimizing the prompt word templates and model behavior to improve long-term recommendation accuracy and generalization ability.

[0049] In one embodiment, a power supply emergency repair material stocking recommendation method based on a large model agent may further include a process of generating recommendation descriptions. The specific process includes: the feedback optimization agent generating recommendation descriptions corresponding to the emergency repair material replacement list through a large language model; the recommendation descriptions include recommended materials, recommendation basis, recommendation process, recommended alternatives, and recommendation priority.

[0050] The purpose of generating recommendation descriptions is to clarify for users which resources are recommended; what the basis for recommending each resource is; whether the recommendations take into account the on-site environment, task semantics, and historical experience; whether there are alternative solutions for the recommended resources; and what the priority is.

[0051] In one embodiment, the feedback optimization agent determines each user role and matches the corresponding role display view according to each user role; the emergency repair material replacement list and recommendation instructions are visualized according to each role display view.

[0052] In the recommendation generation phase, the Large Language Model (LLM) internally maintains a logical reasoning chain; the interpreting agent can interface with the same context, allowing the LLM to regenerate from an interpretative perspective. The generated recommendation descriptions can be presented as a table summary and supplemented with natural language descriptions; they can also be generated from different role perspectives, such as a scheduler's perspective or a warehouse manager's perspective.

[0053] For example, the generated recommendation description is as follows: The current task is to repair a low-voltage distribution box in a mountainous area that has burned down. The operation is expected to be carried out at night. Based on historical cases and the current context, the following materials are recommended: distribution box door lock kit (to replace the damaged door lock and ensure safety), rubber insulating gloves (for protection against electric shock during nighttime operations), and emergency lighting (there is no natural light at night, so a portable light source is needed). Among them, the rubber gloves are in stock in Nancang, Hefei, while the emergency light needs to be transferred from Chizhou, which is expected to arrive in 2 hours. If a replacement is urgently needed, a combination of a portable flashlight and a headlamp can be selected.

[0054] To achieve continuous optimization capabilities, the entire user operation process is tracked and data is collected. This includes tracking whether users adopt recommended lists; whether resources are manually adjusted; whether the scheduling process requires adjustments due to incorrect recommendations; whether users have tagged recommendations as too many / too few, unsuitable, or unusable; and whether manually added resource list items can serve as a basis for future recommendations. Feedback data is recorded in a structured format by the feedback optimization agent for subsequent model optimization and Few-shot example expansion. The feedback records are shown in the table below:

[0055] In one embodiment, the large language model can also self-optimize, automatically building review prompts after recommendations to guide the model in answering questions such as: Is the recommendation complete? What might have been missed? Are there any redundant materials? Have inventory limitations been considered? User feedback can also be compiled into a preference dataset; supervised fine-tuning or preference optimization can be used to adjust the model response in the next round; and preference models can be customized by region or work group to form a scheduling profile.

[0056] To enhance the operability of recommendations across different user groups, the recommendations can also be visualized in different role views, including: Dispatcher's perspective: focusing on repair timeliness, dispatch accessibility, and risk control; Warehouse worker's perspective: focusing on inventory matching, outbound routes, and alternative solutions; Operations and maintenance team's perspective: focusing on on-site portability and safe operation feasibility; Project manager's perspective: summarizing recommendation quality, cost estimation, and efficiency analysis.

[0057] The recommended list can be displayed on the web in the form of an interactive table or card. Each item can be clicked to view the reasons for recommendation, historical usage records, inventory status and delivery time, improving the efficiency of manual intervention and verification.

[0058] Furthermore, to achieve quantitative evaluation and controllable optimization of the entire recommendation process, one embodiment also includes a recommendation quality scoring model, comprising: recommendation accuracy (whether the materials are ultimately used on-site), coverage (the proportion of actually used materials included in the recommendation list), redundancy (the proportion of unused materials), and user satisfaction score (obtainable through an interactive scoring module). Each round of recommendation results triggers a scoring process, generating weekly and monthly recommendation quality reports, enabling long-term intelligent recommendation quality monitoring and automatic adjustment.

[0059] This application provides a method for recommending the stockpiling of power supply emergency repair materials based on a large-scale intelligent agent model, such as... Figure 3As shown, taking power restoration tasks as input, a task parsing agent performs task parsing, semantic modeling, and historical case comparison. Then, a recommendation decision agent performs reasoning and recommendation, linking with inventory to generate a recommended material list. After confirmation by the dispatcher, a feedback optimization agent records feedback and continuously learns, achieving efficient, accurate, explainable, and implementable intelligent recommendations for restoration materials. Furthermore, this is the first time a large language model has been applied to the field of power restoration material recommendation. Through semantic understanding of unstructured restoration work orders, combined with Few-shot case reasoning, customized inventory suggestions are generated, realizing a shift from "human experience" to "model intelligence." An innovative three-agent architecture—task parsing agent, recommendation decision agent, and feedback optimization agent—is constructed. Each agent focuses on an independent task, improving modularity and scalability, and facilitating on-demand deployment and rapid implementation. Through the linkage between the agent and the warehouse system, real-time inventory queries, allocation routes, and alternative solutions are performed, achieving "recommendation is usable immediately," effectively avoiding the "recommendation out of touch" problem of traditional systems and enhancing practicality and implementability.

[0060] Compared to traditional template-based recommendation methods, this approach generates material lists through semantic reasoning and historical analogy, enabling more accurate matching of specific emergency repair scenarios, improving the initial stocking success rate, and reducing redundant dispatching. It possesses autonomous semantic understanding and reasoning capabilities, reducing the burden on dispatching personnel, and its auxiliary capabilities are particularly prominent in emergency situations such as new employees, nighttime operations, or disasters. By tracking and reflecting on user operation results through feedback optimization agents, the large language model can continuously learn from experience and dynamically adjust recommendation strategies, making the recommendation capabilities more accurate with use and forming a positive evolution mechanism.

[0061] It should be understood that although the steps in the above flowcharts are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the above flowcharts may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0062] In one embodiment, such as Figure 1 As shown, a power supply emergency repair material stocking recommendation system based on a large model intelligent agent is provided, including: a task parsing intelligent agent 110, a recommendation decision intelligent agent 120, and a feedback optimization intelligent agent 130, wherein: Task parsing agent 110 is used to receive the emergency repair task description, call the large language model to perform semantic parsing on the emergency repair task description, extract multi-dimensional context information and construct semantic representation vector, and find similar cases from the historical emergency repair case library based on the semantic representation vector. Recommendation decision-making agent 120 is used to construct a multi-round reasoning process that interacts with a large language model based on similar cases and multi-dimensional context information, using Few-shot prompts, and outputs a recommended list of emergency repair materials through the large language model. The recommendation decision-making agent 120 is also used to establish an interface linkage mechanism with the material inventory system. When the materials in the emergency repair material recommendation list cannot be dispatched, the recommendation decision-making agent uses a large language model to perform upstream and downstream substitution analysis on the data in the material inventory system and generates an emergency repair material substitution list. The feedback optimization agent 130 is used to generate recommended descriptions based on the emergency repair material replacement list, collect user feedback data, and optimize the large language model based on the user feedback data.

[0063] In one embodiment, the task parsing agent 110 is also used to perform structured extraction of the emergency repair task description using prompt words based on a large language model to obtain various semantic fields, and extract multi-dimensional contextual information based on each semantic field; concatenate the various semantic fields into a scene description text, encode it into a semantic representation vector through a large language model, and enhance the model's reasoning dimension by adding knowledge templates.

[0064] In one embodiment, the task parsing agent 110 is further configured to determine a vector database based on a historical emergency repair case library; perform semantic similarity matching on the semantic representation vectors in the vector database, and sort the matched vectors based on the semantic matching degree; determine the filtering data, and find the corresponding similar vectors from the sorted matching vectors based on the filtering data, and find similar cases based on the similar vectors.

[0065] In one embodiment, similar cases are labeled as semantic reference standards. The recommendation decision agent 120 is also used to make analogies of emergency repair scenarios based on semantic reference standards to obtain analogy results. The category results and multi-dimensional context information are input into the large language model. The large language model performs multi-round reasoning based on scenario semantics and constructs a multi-round reasoning process. The large language model outputs a recommended list of emergency repair materials based on the multi-round reasoning process.

[0066] In one embodiment, the recommendation decision agent 120 is also used to describe the image factors of the emergency repair task based on the multi-round reasoning process of the large language model; determine each target process of the emergency repair task according to the multi-round reasoning process and influencing factors, and match the corresponding emergency repair materials for each target process; collect each emergency repair material into a material dataset, check the material dataset, and output the checked emergency repair material recommendation list.

[0067] In one embodiment, the recommendation decision-making agent 120 is also used to query the interface and communicate with the material inventory system through the interface; to search for materials in the emergency repair material recommendation list in the material inventory system through the interface; when the materials in the emergency repair material recommendation list are insufficient or cannot be dispatched, a large language model is used to perform upstream and downstream substitution analysis on the data in the material inventory system to find various alternative materials; the alternative materials are prioritized and sorted to generate an alternative material recommendation ranking table, and an emergency repair material replacement list is determined based on the alternative material recommendation ranking table.

[0068] In one embodiment, the recommendation decision-making agent 120 is further configured to select alternative materials from the alternative material recommendation ranking table and generate a recommendation report description.

[0069] In one embodiment, the feedback optimization agent 130 is also used to generate recommendation descriptions corresponding to the emergency repair material replacement list through a large language model; the recommendation descriptions include recommended materials, recommendation basis, recommendation process, recommended alternatives, and recommendation priority.

[0070] In one embodiment, the feedback optimization agent 130 is also used to determine each user role and match the corresponding role display view according to each user role; and to visualize the emergency repair material replacement list and recommendation instructions according to each role display view.

[0071] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0072] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0073] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for recommending and stocking emergency power supply repair materials based on a large-scale intelligent agent model, characterized in that, The method includes: The task parsing agent receives the emergency repair task description, calls the large language model to perform semantic parsing on the emergency repair task description, extracts multi-dimensional context information and constructs a semantic representation vector, and finds similar cases from the historical emergency repair case library based on the semantic representation vector. Based on the similar cases and multi-dimensional context information, the recommendation decision-making agent constructs a multi-round reasoning process that interacts with the large language model through Few-shot prompts, and outputs a recommended list of emergency repair materials through the large language model. The recommendation decision-making agent establishes an interface linkage mechanism with the material inventory system. When the materials in the emergency repair material recommendation list cannot be dispatched, the recommendation decision-making agent uses the big language model to perform upstream and downstream substitution analysis on the data in the material inventory system to generate an emergency repair material substitution list. The feedback optimization agent generates recommended descriptions based on the emergency repair material replacement list, collects user feedback data, and optimizes the large language model based on the user feedback data.

2. The method for recommending and stocking power supply emergency repair materials based on a large-scale intelligent agent as described in claim 1, characterized in that, The large language model is invoked to perform semantic parsing on the description of the emergency repair task, extracting multi-dimensional contextual information and constructing a semantic representation vector, including: The task parsing agent uses prompt words based on a large language model to perform structured extraction of the emergency repair task description, obtains various semantic fields, and extracts multi-dimensional contextual information based on each semantic field. The semantic fields are concatenated into a scene description text, encoded into a semantic representation vector by the large language model, and the model's reasoning dimension is enhanced by adding knowledge templates.

3. The method for recommending and stocking power supply emergency repair materials based on a large-scale intelligent agent according to claim 1, characterized in that, Similar cases are found from the historical emergency repair case database based on the semantic representation vector, including: The task parsing agent determines the vector database based on a historical emergency repair case library; The semantic representation vectors are semantically similar to each other in the vector database, and the matched vectors are sorted based on the semantic matching degree. Determine the filtering data, and find the corresponding similar vectors from the sorted matching vectors based on the filtering data, and find similar cases based on the similar vectors.

4. The method for recommending and stocking power supply emergency repair materials based on a large-scale intelligent agent according to claim 1, characterized in that, Based on the aforementioned similar cases and multi-dimensional contextual information, the recommendation decision-making agent constructs a multi-turn reasoning process that interacts with the large language model through Few-shot prompts, and outputs a recommended list of emergency repair materials through the large language model, including: The similar cases are labeled as semantic reference standards. The recommendation decision-making agent uses the semantic reference standards to make analogies to the emergency repair scenarios and obtain the analogy results. The recommendation decision-making agent inputs the category results and multi-dimensional context information into the large language model, and performs multi-round reasoning based on scene semantics through the large language model, and constructs a multi-round reasoning process; The large language model outputs a recommended list of emergency repair materials based on a multi-round reasoning process.

5. The method for recommending and stocking power supply emergency repair materials based on a large-scale intelligent agent according to claim 4, characterized in that, The large language model outputs a recommended list of emergency repair materials based on a multi-round reasoning process, including: The large language model describes the image factors of the emergency repair task based on a multi-round reasoning process; Based on the multi-round reasoning process and the influencing factors, the target procedures of the emergency repair task are determined, and corresponding emergency repair materials are matched for each target procedure. All emergency repair materials are collected into a material dataset, and the material dataset is checked to output a recommended list of emergency repair materials after the check.

6. The method for recommending and stocking power supply emergency repair materials based on a large-scale intelligent agent according to claim 1, characterized in that, The recommendation decision-making agent establishes an interface linkage mechanism with the material inventory system. When materials in the emergency repair material recommendation list cannot be dispatched, the recommendation decision-making agent uses the large language model to perform upstream and downstream substitution analysis on the data in the material inventory system, generating an emergency repair material substitution list, including: The recommendation decision-making intelligent agent queries the interface and communicates with the material inventory system through the interface; The recommendation decision-making agent searches for materials in the emergency repair material recommendation list in the material inventory system through an interface; When the materials in the emergency repair material recommendation list are insufficient or cannot be dispatched, the large language model is used to perform upstream and downstream substitution analysis on the data in the material inventory system to find various alternative materials. Prioritize each of the candidate materials, generate a recommended ranking table of candidate materials, and determine a replacement list of emergency repair materials based on the recommended ranking table of candidate materials.

7. The method for recommending and stocking power supply emergency repair materials based on a large-scale intelligent agent according to claim 6, characterized in that, The method further includes: The recommendation decision-making agent selects alternative materials from the candidate material recommendation ranking table and generates a recommendation report description.

8. The method for recommending and stocking power supply emergency repair materials based on a large-scale intelligent agent according to claim 1, characterized in that, The feedback optimization agent generates recommendations based on the emergency repair material replacement list, including: The feedback optimization agent generates recommended descriptions corresponding to the emergency repair material replacement list through the large language model. The recommendations include the recommended materials, the basis for the recommendations, the recommendation process, the recommended alternatives, and the recommendation priority.

9. The method for recommending and stocking power supply emergency repair materials based on a large-scale intelligent agent according to claim 1, characterized in that, The method further includes: The feedback optimization agent determines each user role and matches the corresponding role display view according to each user role; The list of alternative repair materials and recommended instructions are displayed visually according to the view of each role.

10. A power supply emergency repair material preparation and recommendation system based on a large-scale intelligent agent model, characterized in that, The system includes: The task parsing agent is used to receive the emergency repair task description, call the large language model to perform semantic parsing on the emergency repair task description, extract multi-dimensional context information and construct a semantic representation vector, and find similar cases from the historical emergency repair case library based on the semantic representation vector. The recommendation decision-making agent is used to construct a multi-round reasoning process that interacts with the big language model based on the similar cases and multi-dimensional context information, and outputs a recommended list of emergency repair materials through the big language model. The recommendation decision-making agent is also used to establish an interface linkage mechanism with the material inventory system. When the materials in the emergency repair material recommendation list cannot be dispatched, the recommendation decision-making agent uses the big language model to perform upstream and downstream substitution analysis on the data in the material inventory system and generates an emergency repair material substitution list. The feedback optimization agent is used to generate recommended descriptions based on the emergency repair material replacement list, collect user feedback data, and optimize the large language model based on the user feedback data.

Images