Power grid outage notification optimization generation method and system based on large language model

Abstract

The present application relates to a kind of power grid outage notice optimization generation method and system based on large language model.The method includes: collecting power grid outage data for processing, obtain structured power grid outage data input into large language model, based on Prompt template generation strategy, Prompt diversity parameter control mechanism carries out semantic abstraction and intent modeling, extracts semantic elements;Using semantic elements for each different preset role perspective generates corresponding outage notification content, generates notification draft to carry out scene adaptation and style optimization, language proofreading and logic optimization, obtains each version of final power grid outage notice;And according to artificial feedback, large language model is optimized.Through generating outage notification content for different preset role perspective, the readability of notification content and user acceptance can be improved;Since scene adaptation, style optimization, logic optimization and other processing are carried out, the content accuracy, logic and language quality can be improved, and artificial intervention can be reduced.

Description

Technical Field

[0001] This invention relates to the field of power grid notification technology, and in particular to an optimized generation method and system for power grid outage notifications based on a large language model. Background Technology

[0002] With the continuous expansion and intelligentization of my country's power system, planned power outages for distribution network maintenance, load adjustment, and new installations are becoming increasingly frequent. Against this backdrop, timely, accurate, and user-friendly power outage notifications have become crucial for power companies to fulfill their service commitments, maintain customer relationships, and ensure stable social operation. Traditionally, power outage notifications are manually drafted by grid dispatchers, marketing personnel, or public relations staff. The typical process includes: the dispatch system issuing the power outage plan → a designated person extracting key fields (time, location, scope of impact, etc.) → manually drafting the announcement using a notification template → publishing it through SMS platforms, WeChat official accounts, official websites, and other channels. Although some power companies have recently attempted to use rule engines or template engines for semi-automatic generation of notification content, significant technical bottlenecks remain, making it difficult to meet the new demands for "intelligent, personalized, and diversified" power outage notifications.

[0003] Existing methods generally rely on static templates and field replacement logic, resulting in rigid language and a lack of semantic understanding and text optimization capabilities. Common problems include stiff content, incomplete information, and failure to cover public concerns, leading to increased user complaints and poor dissemination effectiveness. For example, notices like "Due to line maintenance, there will be a power outage in XX area during a certain period" fail to effectively explain the construction background, details of the affected area, or reminder measures, hindering the building of user trust. On the other hand, current systems mostly use a "single notification version, multiple channels reuse" approach for announcements, lacking the ability to adapt to the reading habits and information concerns of different audiences. This often leads to a chain of problems in government services, such as misunderstandings and untimely responses.

[0004] Therefore, traditional power outage notices require manual editing, which is labor-intensive and often results in low quality of expression and low user satisfaction. Summary of the Invention

[0005] In order to solve the above-mentioned technical problems, this paper provides a method and system for optimizing the generation of power grid outage notices based on a large language model, which can improve the expression quality of power grid outage notices and meet the needs of different audiences.

[0006] A method for optimizing the generation of power grid outage notifications based on a large language model, the method comprising:

[0007] Power outage data is collected from various data systems, and the power outage data is standardized and converted into unified structured power outage data.

[0008] Structured power grid outage data is input into a large language model. Based on the Prompt template generation strategy and the Prompt diversity parameter control mechanism, semantic abstraction and intent modeling are performed on the structured power grid outage data to extract the semantic elements required for power outage notification.

[0009] The semantic elements are used to generate corresponding power outage notification content for each different preset role perspective, and a notification draft is generated through a content fusion strategy.

[0010] The draft notification is adapted to the target distribution channel and its style is optimized. The language is reviewed and the logic is optimized to obtain the final power grid outage notification for each version.

[0011] Obtain revision instructions for the final power grid outage notification, receive human feedback based on the revision instructions, and optimize the large language model based on the human feedback.

[0012] In one embodiment, power outage data is collected from various data systems, and the power outage data is standardized to convert it into unified structured power outage data, including:

[0013] Collect various types of power grid outage data from various data source systems;

[0014] The power grid outage data is processed according to a standardized field format to construct a unified structured power grid outage data.

[0015] The data source system includes a power grid dispatching system, a marketing service system, and a manual supplementation system.

[0016] The structured power grid outage data includes outage plans, outage equipment, user impact range, construction unit, responsible unit information, and natural language description fields.

[0017] In one embodiment, semantic abstraction and intent modeling are performed on the structured power grid outage data based on a Prompt template generation strategy and a Prompt diversity parameter control mechanism to extract the semantic elements required for the power outage notification, including:

[0018] Based on the Prompt template generation strategy, the structured power grid outage data is converted into a structured semantic representation for notification generation using a large language model and meaning reconstruction technology.

[0019] The Prompt diversity parameter control mechanism employs a Few-shot example combination strategy to enhance the accuracy of intent understanding, performs intent modeling based on the structured semantic representation, and extracts the semantic elements required for power outage notification.

[0020] The semantic elements include the background and cause of the power outage, the scope of impact and user classification prompts, the expected time for power restoration, risk warnings and external environment linkages, and safe electricity use and precautions.

[0021] In one embodiment, the method further includes:

[0022] Based on the Prompt diversity parameter control mechanism, the semantic expression of the structured semantic representation is transformed, and the parameters of the large language model are controlled to generate text expressions of different styles.

[0023] Generate different versions of notification content for various text styles;

[0024] External environment information is acquired, and the external environment information is fused with the notification content based on a context fusion mechanism to achieve dynamic enhancement of the Prompt.

[0025] In one embodiment, the semantic elements are used to generate corresponding power outage notification content for each different preset role's perspective, and a notification draft is generated through a content fusion strategy, including:

[0026] Based on the role simulation and language generation capabilities of the large language model, various preset roles are designed, and Prompt templates from different role perspectives are designed.

[0027] Using the semantic elements, power outage notification content corresponding to different preset role perspectives is generated according to each of the Prompt templates;

[0028] Calculate the evaluation metrics for each version of the power outage notification content and determine the role weights for each preset role.

[0029] The context summaries of the power outage notifications from each version are fused, and a draft notification is output based on the evaluation metrics and role weights.

[0030] In one embodiment, the method includes:

[0031] The language style of the power outage notification content is adjusted using the Prompt template, and keyword filtering rules are introduced to standardize the terminology of the power outage notification content. The adjusted power outage notification content is then fused with the summary.

[0032] In one embodiment, the method further includes:

[0033] Based on the aforementioned large language model, a new role-view Prompt template is added to achieve role expansion.

[0034] In one embodiment, the draft notification is adapted to the target distribution channel for specific scenarios and style optimization, and language and logic are reviewed and optimized to obtain various versions of the final power outage notification, including:

[0035] Identify each target publishing channel and obtain the style characteristics and format restrictions corresponding to each target publishing channel;

[0036] Each of the aforementioned notification drafts is displayed according to its respective style characteristics and format, and then adapted to the scene and optimized in style to obtain each preliminary notification.

[0037] Each of the preliminary notices was reviewed for language and optimized for logic to obtain the final power outage notices for each version.

[0038] In one embodiment, a revision instruction for the final power outage notification is obtained, human feedback is received based on the revision instruction, and the large language model is optimized based on the human feedback, including:

[0039] Output and display the final power outage notification, and obtain revision instructions based on the final power outage notification;

[0040] Based on the revision instructions, user behavior and preference information is collected from manual feedback and stored in the feedback database;

[0041] The large language model is optimized using the feedback database.

[0042] A power grid outage notification optimization generation system based on a large language model, the system comprising:

[0043] The data processing module is used to collect power grid outage data from various data systems, standardize the power grid outage data, and convert it into unified structured power grid outage data.

[0044] The semantic element extraction module is used to input structured power grid outage data into a large language model, and perform semantic abstraction and intent modeling on the structured power grid outage data based on the Prompt template generation strategy and the Prompt diversity parameter control mechanism to extract the semantic elements required for the power outage notification.

[0045] The draft generation module is used to generate corresponding power outage notification content for each different preset role perspective using the semantic elements, and to generate a notification draft through a content fusion strategy.

[0046] The notification optimization module is used to adapt the draft notification to the target release channel, optimize its style and language, and perform language review and logic optimization to obtain the final power grid outage notification for each version.

[0047] The model optimization module is used to obtain revision instructions for the final power grid outage notification, receive human feedback according to the revision instructions, and optimize the large language model according to the human feedback.

[0048] The aforementioned method and system for optimizing the generation of power grid outage notifications based on a large language model employs a Prompt template generation strategy and a Prompt diversity parameter control mechanism for semantic abstraction and intent modeling to extract semantic elements. This allows for the rapid extraction of key element information, saving manual resources. By generating outage notification content for different preset role perspectives, the readability and user acceptance of the notification content are improved. Furthermore, due to scene adaptation, style optimization, and logic optimization, the accuracy, logic, and language quality of the content are enhanced, reducing manual intervention. Additionally, a human feedback channel is provided, allowing for continuous optimization of the notification and ultimately improving the expression quality of power grid outage notifications to meet the needs of different audiences. Attached Figure Description

[0049] Figure 1 This is an application environment diagram of a power grid outage notification optimization generation method based on a large language model in one embodiment.

[0050] Figure 2 This is a flowchart illustrating an optimized generation method for power grid outage notifications based on a large language model in one embodiment.

[0051] Figure 3 This is a flowchart illustrating the optimized generation method for power grid outage notifications based on a large language model in another embodiment.

[0052] Figure 4 This is a block diagram of a power grid outage notification optimization generation system based on a large language model in one embodiment.

[0053] Figure 5 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0054] 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.

[0055] The optimized generation method for power outage notifications based on a large language model provided in this application can be applied to, for example... Figure 1The application environment shown. For example... Figure 1 As shown, the application environment includes computer device 110. Computer device 110 can collect power outage data from various data systems, standardize the power outage data, and convert it into unified structured power outage data. Computer device 110 can input the structured power outage data into a large language model, and based on a Prompt template generation strategy and a Prompt diversity parameter control mechanism, perform semantic abstraction and intent modeling on the structured power outage data to extract the semantic elements required for the power outage notification. Computer device 110 can use these semantic elements to generate corresponding power outage notification content for different preset role perspectives, and generate a notification draft through a content fusion strategy. Computer device 110 can adapt the notification draft to the target distribution channel, optimize its style, and perform language review and logic optimization to obtain various versions of the final power outage notification. Computer device 110 can obtain revision instructions for the final power outage notification, receive human feedback based on the revision instructions, and optimize the large language model based on the human feedback. The computer device 110 can be, but is not limited to, various personal computers, laptops, smartphones, robots, drones, tablets, and other devices.

[0056] In one embodiment, such as Figure 2 As shown, an optimized generation method for power grid outage notifications based on a large language model is provided, including the following steps:

[0057] Step 202: Collect power grid outage data from various data systems, standardize the power grid outage data, and convert it into unified structured power grid outage data.

[0058] Computer devices can perceive key information about a task from multi-source data and organize it into an input format that can be recognized by a large language model, laying a semantic foundation for subsequent content generation and style control.

[0059] In one embodiment, the proposed method for optimizing the generation of power grid outage notifications based on a large language model may further include a data processing process. This process includes: collecting various types of power grid outage data from various data source systems; processing the power grid outage data according to standardized field formats to construct unified structured power grid outage data; wherein the data source systems include a power grid dispatching system, a marketing service system, and a manual supplementation system; and the structured power grid outage data includes outage plans, outage equipment, user impact range, construction unit, responsible unit information, and natural language description fields.

[0060] Computer equipment can extract structured or semi-structured task data from source systems such as dispatching systems and operation and maintenance systems. This data includes: power outage plans (such as task number, planned time, and plan type), power outage equipment (such as transformer substation name, power supply line, and distribution transformer), user impact range (residential / enterprise unit classification and power supply radius), construction unit and responsible unit information, and natural language description fields (such as remarks and brief description of reasons). It can also automatically standardize field formats and build a unified data input structure for subsequent processing.

[0061] Specifically, each source system can support multiple data input types from various sources, as shown in the table below:

[0062]

[0063] After standardization processing, the collected power outage data can be converted into unified structured power outage data.

[0064] Step 204: Use semantic elements to generate corresponding power outage notification content for each different preset role perspective, and generate a notification draft through content fusion strategy.

[0065] Computer devices can use large language models to perform Prompt-enhanced intent understanding on the input structured power outage data, abstracting the semantic elements required for the notification.

[0066] In one embodiment, the proposed method for optimizing the generation of power grid outage notifications based on a large language model may further include a process for extracting semantic elements. Specifically, this process includes: using a Prompt template generation strategy, transforming structured power grid outage data into a structured semantic representation for notification generation through meaning reconstruction technology using a large language model; employing a Few-shot example combination strategy to enhance the accuracy of intent understanding through a Prompt diversity parameter control mechanism; performing intent modeling based on the structured semantic representation; and extracting the semantic elements required for the outage notification; the semantic elements include the outage background and cause description, the scope of impact and user classification prompts, the expected time for power restoration, risk warnings and external environment linkages, and safe electricity use and precautions.

[0067] By setting a Prompt template generation strategy, the large language model can be guided to extract semantic elements. Specifically, the five extracted semantic elements are explained and illustrated in the table below:

[0068]

[0069] In this embodiment, a specific Prompt example is provided as follows:

[0070] Input via computer device: You are an assistant at an electric power company, responsible for generating a summary of a power outage notice for the public based on the following task information. Please extract five core parts. Task information: Reason for outage: Planned maintenance on the 10kV XX line; Outage time: July 2nd, 08:00 to 17:00; Affected area: XX residential area, XX industrial park; User type: Residential users, enterprises and institutions; Supplementary notes: This is an annual routine maintenance; External environment: Orange high temperature warning. The computer device can output five parts in concise and natural language using a large language model: Explanation of the reason for outage; Affected area; Expected restoration time; Safety reminders; and helpful tips and suggestions.

[0071] In one embodiment, the proposed method for optimizing and generating power grid outage notifications based on a large language model may further include a process of diversifying expressions. Specifically, this process includes: transforming the semantic expression of the structured semantic representation based on a Prompt diversity parameter control mechanism, and controlling the parameters of the large language model to generate text expressions of different styles; generating different versions of notification content for each different style of text expression; acquiring external environment information, and fusing the external environment information with the notification content based on a context fusion mechanism to achieve dynamic enhancement of the Prompt.

[0072] To prevent the generated content from being monotonous, this embodiment can introduce a Prompt diversity parameter control mechanism, such as: changing the expression (e.g., "to ensure equipment operation → due to the annual inspection of power grid equipment"); controlling the generation parameters such as temperature / Top-k to sample different expressions; and generating multiple versions for subsequent fusion and filtering.

[0073] In this embodiment, the computer device also supports integrating external environmental information, such as weather and holidays, to dynamically enhance the prompt. For example, if the weather is hot, the message "It's hot, please take precautions against heatstroke" is automatically added; if it is the eve of a holiday, the message "Because the holiday is approaching, please make arrangements for your electricity usage in advance" is displayed.

[0074] Step 206: Use semantic elements to generate corresponding power outage notification content for each different preset role perspective, and generate a notification draft through content fusion strategy.

[0075] Computer equipment can design various role-playing prompts, such as: the perspective of a power dispatching expert (ensuring accurate terminology and rigorous task description); the perspective of an ordinary resident (concise and easy to understand, highlighting key information); the perspective of government and enterprise users (emphasizing service impact and the rationality of explanation); and the perspective of a publicity editor (fluent language and user-friendly dissemination). Each role-playing prompt, in conjunction with master data, triggers a large language model (LLM) to generate content, and uses content fusion strategies (such as priority weighting, sentence selection, and automatic summarization) to generate a final draft notification.

[0076] In one embodiment, the proposed method for optimizing the generation of power outage notifications based on a large language model may further include a process for generating a notification draft. This process includes: designing various preset roles based on the role simulation and language generation capabilities of the large language model, and designing Prompt templates for different role perspectives; using semantic elements, generating power outage notification content corresponding to different preset role perspectives based on each Prompt template; calculating evaluation metrics for each version of the power outage notification content and determining the role weights corresponding to each preset role; performing contextual summarization fusion on each version of the power outage notification content, and outputting a notification draft based on the evaluation metrics and role weights.

[0077] Traditional notification content generation often relies on a single language style, failing to balance power grid professionalism with public readability. Large-scale language models, however, possess the ability to simulate character expression styles, control tone, and emphasize key points. In this embodiment, the computer device can leverage the powerful character simulation and language generation capabilities of the large-scale model to design Prompt templates from multiple perspectives, generating diverse, focused, and adaptable power outage notification drafts. A fusion strategy is then used to select the optimal version. Specifically, the preferences and styles corresponding to the generated content drafts from different perspectives are shown in the table below:

[0078]

[0079] In this embodiment, by equipping each role with a dedicated Prompt template and using Few-shot examples or instruction annotations, the output style and focus of the model can be controlled. Specifically, when the role is a power dispatching expert, a residential user representative, a government and enterprise user representative, a publicity editor, or a safety management personnel, the dedicated Prompt templates for several randomly selected roles can be represented as follows:

[0080] Power Dispatch Specialist: Based on the following planned maintenance information, draft an internal power outage notice. Please highlight the reason for the outage, the line name, the construction time, and the affected area. Use precise language and standardized terminology. The planned maintenance information is as follows: Line under maintenance: 10kV XX line; Planned time: July 2, 2025, 08:00-17:00; Construction unit: State Grid XX Transmission Operation and Maintenance Center; Affected area: XX residential area, XX industrial park; Number of users: approximately 274 households.

[0081] Residential User Representatives: Explain the upcoming power outage to residents in a friendly and natural manner, highlighting the reason, time, affected area, and helpful tips. The information is as follows: Power outage time: July 2nd, 08:00 to 17:00; Reason: Equipment maintenance (annual maintenance); Affected area: XX residential area and surrounding area; External environment: High temperature red alert.

[0082] For government and enterprise clients: Draft a formal notification letter to inform enterprise clients of the planned power outage. The letter should be well-written, logically clear, and include risk warnings and suggestions. The information provided is as follows: Task No.: GZ-20250702001; Affected Area: XX Industrial Park; Suggested Measures: Please arrange emergency power supply in advance or adjust production plans.

[0083] In one embodiment, the provided method for optimizing the generation of power outage notifications based on a large language model may further include the process of implementing a summation optimization strategy. The specific process includes: adjusting the language style of the outage notification content through a Prompt template, introducing keyword filtering rules to standardize the terminology of the outage notification content, and then performing upper-level summary fusion on the adjusted outage notification content.

[0084] Computer devices can simultaneously generate multiple notification versions based on different roles' prompts and execute fusion strategies. Specifically, this can include several parts: comparing evaluation metrics, contextual summary fusion, voting-weighted fusion, and generating manual revision suggestions. The computer device can use metrics such as BLEU, Readability, sentence length, and keyword coverage to evaluate version quality; extract clear and readable segments from each version and combine them into a summary notification draft; assign weights to the output of each role and select the final main output version through aggregation model logic; if conflicts or redundant content are detected, prompts are automatically generated for quick manual confirmation or rewriting.

[0085] Specifically, computer devices can also control sentence type (declarative / suggestive / explanatory), tone (neutral / mild / formal), and word preference through Prompt; introduce keyword filtering rules to ensure terminology consistency (such as uniformly using "planned maintenance" instead of "routine maintenance"); and set content layout parameters such as word limit, number of paragraphs, and whether to list items in bullet points, based on the target publishing scenario.

[0086] In one embodiment, the power grid outage notification optimization generation method based on a large language model may further include a role expansion process, specifically including: adding a new role perspective Prompt template based on the large language model to achieve role expansion.

[0087] Computer equipment can support flexible expansion of role perspectives based on the dynamic development of the power industry. For example: Media Release Assistant: used to generate graphic and text content for government new media; Emergency Warning Broadcaster: used for automatic generation of broadcast scripts; Legal Compliance Reviewer: used to check whether the generated content contains risky statements.

[0088] Step 208: Adapt the draft notification to the target distribution channel, optimize its style and context, and review and optimize its language and logic to obtain the final power outage notification for each version.

[0089] In practical applications, power grid companies need to deliver power outage notices to different user groups through various channels, including WeChat official accounts, SMS platforms, government and enterprise portals, customer service hotlines, and newspaper announcements. Different channels have different requirements for the style, format, and information density of the notice content. If a single document is submitted to multiple platforms with a uniform style, problems such as being verbose and difficult to read, overly simplified, or inconsistent in style often arise, affecting the communication effectiveness and user satisfaction. Therefore, by designing content adaptation and optimization tailored to different publishing scenarios, and combining the characteristics of the publishing platform, the preferences of the target audience, and language control strategies, the large language model is guided to output diverse versions with strong adaptability.

[0090] Computer devices can perform customized generation for different notification distribution channels, such as: for SMS platforms: concise content highlighting key time and regional information; for WeChat official account platforms: rich in graphics and text, clearly segmented, and with a friendly language; for government and enterprise special reports: standardized format, rigorous logic, and complete appendix structure. Specifically, different Prompt style configurations and word count control strategies can be automatically switched according to the channel.

[0091] Next, the computer equipment can perform secondary language optimization on the initial generated results, including: word refinement and sentence fluency adjustment; reinforcement of ambiguous logic (e.g., replacing "restore as soon as possible" with "power is expected to be restored at xx time"); regulatory compliance review and standardization of safety terminology (e.g., "temporary power outage for construction"); and support for an LLM-based "self-checking Prompt chain" to achieve automatic diagnosis and repair of the model-generated content.

[0092] In one embodiment, the method for optimizing and generating power grid outage notices based on a large language model may further include a process of optimizing the notices. The specific process includes: determining each target publishing channel and obtaining the style features and format restrictions corresponding to each target publishing channel; displaying each draft notice according to each style feature and format for scenario adaptation and style optimization to obtain each preliminary notice; and performing language review and logic optimization on each preliminary notice to obtain each version of the final power grid outage notice.

[0093] The preset typical notification publishing scenario types are shown in the table below:

[0094]

[0095] Computer devices control content generation strategies by customizing Prompt templates for each publishing scenario. A specific example is shown below:

[0096] SMS Notification: Please generate a short message of no more than 70 characters based on the following task information. The message should highlight the time, location, reason, and expected power restoration. Time: July 2nd, 08:00–17:00; Location: XX Community; Reason: Equipment maintenance; Restoration expected by 17:00 on the same day. Example: "During the period from 08:00 to 17:00 on July 2nd, XX Community experienced a temporary power outage due to equipment maintenance. Power is expected to be restored on the same day. We apologize for the inconvenience."

[0097] Prompt for WeChat Official Account Long Article: Please generate a notice suitable for WeChat Official Account publication, with friendly and natural language, and content divided into paragraphs, including the reason for the power outage, schedule, affected area, friendly reminders, and safety tips, and keep it within 500 words. The generated paragraph structure is as follows: Power Outage Notice Explanation, Schedule, Affected Area, Friendly Reminders, and Safety Tips.

[0098] Government and Enterprise Notification Prompt: Please draft a formal notification suitable for sending to government and enterprise users. Use accurate language, emphasize the scope of impact and recommended measures, and retain contact information. Generated result structure: Notification title, salutation and background information, plan details (table or paragraph), risk warnings and suggestions, contact information and closing.

[0099] In this embodiment, the computer device can perform structured control on the LLM-generated content, mainly including word count control: using a "describe within n words..." type of prompt to guide the discussion; post-processing to prune the word count of large model outputs; structural format control: removing redundant punctuation and pause words for SMS scenarios; adding titles, paragraph breaks, emoticons, or tags for public accounts; automatically inserting HTML structure tags for web page announcements; and output post-processing: performing semantic polishing such as sentence correction, typo detection, and punctuation optimization on the generated content.

[0100] In this embodiment, the computer device also supports generating multiple scenario versions in a single task, simultaneously outputting: a simplified SMS version; a long article version for a public WeChat account; a government / enterprise notification letter; and a structured webpage announcement. These versions can be called by the front-end interface to achieve automatic distribution and intelligent adaptation, and support interface parameter control style.

[0101] When external events such as high temperatures, rainy seasons, and holidays are detected, enhanced prompts are automatically generated, such as: "It will be hot during the power outage; please take precautions against heatstroke in advance"; "Please pay attention to pedestrian safety and stay away from the work area during construction." The tone also changes depending on the user type (government / enterprise, residential). The residential user version uses a friendly prompt, such as "Thank you for your understanding and cooperation"; the government / enterprise version uses a neutral, professional expression, such as "Please arrange your production plan accordingly."

[0102] After completing the initial multi-role content generation and scenario adaptation, the generated power outage notice drafts often still have the following problems: the sentences are not fluent or contain repetitions; the information logic is not complete, such as the "reason-time-impact-reminder" chain being broken; vague, redundant, and non-standard terms are used; and the formatting is inconsistent, affecting the standardization of the release.

[0103] Therefore, a language polishing and self-review mechanism based on a large language model can be proposed to reconstruct the initial draft of the large model. The notification text can be deeply optimized and corrected from multiple dimensions such as language quality, expression logic, and security compliance to ensure that the final output version has the characteristics of semantic clarity, standardized format, logical closed loop, and easy understanding by users.

[0104] In this embodiment, the computer device uses LLM to construct a set of polishing Prompt templates to guide the model in optimizing the initial draft. The optimization results are shown in the table below:

[0105]

[0106] For example, please optimize the language of the following notification to make it more fluent, friendly, and professional, while keeping it under 400 words and preserving all original information:

[0107] Original message: XX residential area will experience a power outage on the morning of July 2nd. This is for equipment maintenance. Your cooperation and understanding are appreciated.

[0108] Output Example: Due to power equipment maintenance, XX residential area will experience a temporary power outage from 08:00 to 17:00 on July 2nd. Please make necessary power arrangements in advance. We apologize for any inconvenience caused during the maintenance period and thank you for your support and cooperation!

[0109] To ensure clear logical expression and structural integrity, a review strategy can be introduced in this embodiment:

[0110] Logical chain verification: Are the following complete: "Cause → Time → Location → Impact → Reminder"?

[0111] Ambiguity identification prompts: Identify vague expressions such as "recovering quickly" or "partial area" and suggest replacing them with more specific terms;

[0112] Conflict detection: If the body text is "July 2nd" but the title is "July 3rd", the model can automatically point it out and suggest a correction.

[0113] Example illustration of a logical integrity check Prompt:

[0114] Please check the following notification for semantic ambiguity, unclear logic, or missing information. If any issues are found, please indicate the cause and provide optimization suggestions:

[0115] Original text: "Line maintenance will be carried out tomorrow, which will affect some areas. We ask for your understanding."

[0116] Output Examples: Problem 1: The specific date of "tomorrow" is not specified; Suggestion: Please replace "tomorrow" with a specific date, such as "July 2nd"; Problem 2: It is not specified which areas are included in "some areas"; Suggestion: Add a description of the affected areas, such as "XX residential area and XX industrial park will be affected".

[0117] In this embodiment, the computer equipment certificate can have a built-in terminology dictionary and safety compliance rules. For example, non-standard terms such as "power outage" and "shutdown" should be replaced with formal terms such as "planned power outage" and "equipment maintenance". It emphasizes that safety reminders must use standardized terms, such as "please stay away from the construction area" and "do not touch the equipment". It also marks and explains suspected risk statements, such as "expected restoration" should be clearly stated as "power is expected to be restored at XX time".

[0118] In this embodiment, a "self-proofing + generation" Prompt chain can also be introduced, whereby the model first performs a self-evaluation as a "reviewer" and then rewrites the notification content based on the role of a "rewriter".

[0119] Example instructions, self-check: Please determine if there are any logical errors, language problems, or missing information in the following notice, and list your suggestions for improvement; rewrite output: Please regenerate a complete and natural version of the notice content based on the above suggestions.

[0120] This chained prompt structure can significantly improve the self-consistency and accuracy of the model-generated content, and is particularly suitable for high-risk, formal publishing channels.

[0121] Step 210: Obtain the revision instruction for the final power grid outage notification, receive human feedback based on the revision instruction, and optimize the large language model based on the human feedback.

[0122] The computer equipment can support the generation of multiple notification style versions at once; allowing manual selection, revision, or quick confirmation and release; feeding back user modification records to the local database for subsequent model optimization and style fine-tuning; and finally, the output content can be synchronized to multiple terminals such as WeChat official accounts, mini-program notification pages, SMS gateways, and web page announcements.

[0123] In one embodiment, the provided method for optimizing and generating power grid outage notifications based on a large language model may further include a model optimization and notification optimization process. The specific process includes: outputting and displaying the final power grid outage notification, and obtaining revision instructions based on the final power grid outage notification; collecting user behavior and preference information based on the revision instructions and storing the user behavior and preference information in a feedback database; and optimizing the large language model through the feedback database.

[0124] To adapt to the multi-dimensional, multi-user, and multi-platform notification publishing needs in power grid operations, and to continuously improve the quality and personalization of notification content generation, this embodiment also designs a multi-version output and user feedback closed-loop mechanism to ensure that the generated content not only covers a wide range of usage scenarios, but also has the ability to continuously evolve, transfer styles, and accurately adapt.

[0125] After semantic parsing, role-driven generation, scenario adaptation, and polishing / proofreading, the computer device can support the output of multiple content versions at once, each targeting different distribution channels or user groups, as shown in the table below:

[0126]

[0127] Each version is generated based on an independent Prompt process and then output to the release interface after content fusion and language review optimization.

[0128] In this embodiment, considering that some notifications involve sensitive times (high temperatures, holidays, before major events) or important areas (hospitals, power supply to government agencies), manual content confirmation and revision are allowed after multiple versions are output: the backend UI provides multiple versions of content and supports quick preview and comparison; light editing is possible (changing titles, fine-tuning tone, inserting notes); all editing actions are automatically recorded and form a version difference log (diff); users can choose to separate "released version" and "archived version" to retain all historical drafts.

[0129] To enable dynamic and personalized evolution of notification content generation, computer devices can also support feedback collection mechanisms: click-through rate and reading time feedback (channels such as WeChat official accounts); SMS receipts and complaint analysis (used to identify errors or ambiguities); collection of manual revision records (edited content as a reference sample for fine-tuning); collection of user-initiated feedback (e.g., evaluations and scores such as "whether the notification is clear" and "whether it affects the schedule"). This feedback data will be uniformly stored in the "Notification Optimization Feedback Database" for training lightweight prompt word control modules (such as Prompt Selector) or building preference models (such as LoRA adapter).

[0130] In one embodiment, the proposed method for optimizing the generation of power grid outage notifications based on a large language model may further include an automatic optimization process, specifically including: Prompt memory construction: saving the best-performing Prompt structures from different regions, units, and time periods as Prompt Memory for reuse in future similar tasks; Fine-tuning adapter training (LoRA / Adapter): training the LoRA module based on manually revised samples and reading preference labels to improve generation quality; Automatic strategy adjustment: automatically switching the model generation strategy based on platform feedback data (such as switching to a more concise style, adjusting the expression of safety prompts, etc.); Dynamic learning of customized content templates: after accumulating sufficient samples, the system can automatically form a "content expression preference profile" for a region / line / unit, and intelligently select the most suitable expression method in future tasks.

[0131] This application provides an optimized generation method for power grid outage notifications based on a large language model. The specific process is as follows: Figure 3 As shown, the process includes: data source for the dispatch system; task information collection and parsing; obtaining power outage plan data, equipment information, and impact range; semantic abstraction and intent modeling through a large language model; extraction of five major semantic elements: power outage background, impact range, recovery time, risk warning, and precautions; then, content generation driven by multiple roles, including the perspectives of dispatch experts, ordinary residents, government and enterprise users, and publicity editors; followed by scenario adaptation and style optimization; publication to SMS platforms, official accounts, government and enterprise special reports, etc.; language proofreading and logic optimization; multiple versions output, published to various terminals such as WeChat / mini-programs / SMS / webpages; and finally, feedback loop optimization through user feedback.

[0132] This application is the first to introduce a large language model into the power industry's power outage notification generation scenario, realizing the semantic mapping of structured task information to natural language text, breaking through the limitations of traditional template-replacement notification generation methods. It constructs a multi-role collaborative Prompt generation mechanism to enhance content diversity and user-friendliness, introducing multiple perspectives such as "power dispatcher," "residential user representative," and "government and enterprise customer representative." Through chained Prompts or Agent methods, it guides the generation of content text with differentiated styles, authentic user perspectives, and interpretability, significantly enhancing the user experience. It also addresses various communication methods including SMS, official accounts, web pages, and government / enterprise special reports. Different Prompt chains with varying language styles, length limits, and structural requirements are designed for different publishing channels to ensure that the generated content not only meets the requirements of the target platform but also enhances the dissemination effect. By incorporating external information such as weather, holidays, and event background, the model can automatically supplement user-conceived content such as high-temperature operation risk reminders and holiday duty instructions, making the notifications more humane and practical. By introducing a self-review and user feedback mechanism, a closed-loop optimization capability is formed. A large-scale model "self-checking Prompt chain" is designed to realize the logical verification and polishing reconstruction of the content, and the generation strategy is continuously optimized based on user feedback, which can significantly improve the accuracy and stability of the output content.

[0133] Furthermore, the power outage notification optimization generation method based on a large language model provided in this application replaces manual writing with automatic system generation, reducing manual operations by more than 90% and effectively improving work efficiency. It is particularly suitable for issuing multiple notifications and peak operation periods, and can significantly improve the efficiency of power outage notification content generation. The language of the content is clearer, more natural, and more approachable, effectively conveying key information such as the cause of the power outage, the scope of impact, and the restoration time, reducing the risk of user misunderstanding and complaints, and enhancing the expression quality and user acceptance of the notification content. It improves the professionalism and humanistic care of the notification information, strengthens the enterprise's service capabilities, and helps to build a modern power distribution service system that emphasizes both "digitalization and humanization". It is not only applicable to power outage notification scenarios, but can also be extended to other power information release tasks such as equipment alarms, planned maintenance, load adjustment, and energy management, and has broad industry application value.

[0134] It should be understood that although the steps in the flowchart above are shown sequentially as indicated by 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 flowchart above 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.

[0135] In one embodiment, such as Figure 4 As shown, a power grid outage notification optimization generation system based on a large language model is provided, including: a data processing module 410, a semantic element extraction module 420, a draft generation module 430, a notification optimization module 440, and a model optimization module 450, wherein:

[0136] Data processing module 410 is used to collect power grid outage data from various data systems, standardize the power grid outage data, and convert it into unified structured power grid outage data;

[0137] The semantic element extraction module 420 is used to input structured power grid outage data into the large language model, and perform semantic abstraction and intent modeling on the structured power grid outage data based on the Prompt template generation strategy and the Prompt diversity parameter control mechanism to extract the semantic elements required for the power outage notification.

[0138] The draft generation module 430 is used to generate corresponding power outage notification content for each different preset role perspective using semantic elements, and to generate notification drafts through content fusion strategies.

[0139] The notification optimization module 440 is used to adapt the notification draft to the target release channel, optimize the style, and perform language review and logic optimization to obtain the final power grid outage notification for each version.

[0140] The model optimization module 450 is used to obtain revision instructions for the final power grid outage notification, receive human feedback based on the revision instructions, and optimize the large language model based on the human feedback.

[0141] In one embodiment, the data processing module 410 is further configured to collect various types of power grid outage data from various data source systems; process the power grid outage data according to standardized field formats to construct unified structured power grid outage data; wherein, the data source systems include power grid dispatching systems, marketing service systems, and manual supplementation systems; and the structured power grid outage data includes outage plans, outage equipment, user impact range, construction units, responsible unit information, and natural language description fields.

[0142] In one embodiment, the semantic element extraction module 420 is further used to convert structured power grid outage data into a structured semantic representation for notification generation using a large language model and meaning reconstruction technology based on the Prompt template generation strategy. The Prompt diversity parameter control mechanism adopts a Few-shot example combination strategy to enhance the accuracy of intent understanding, performs intent modeling based on the structured semantic representation, and extracts the semantic elements required for the power outage notification. The semantic elements include the background and reason for the power outage, the scope of impact and user classification prompts, the expected time of power restoration, risk warnings and external environment linkages, and safe electricity use and precautions.

[0143] In one embodiment, the semantic element extraction module 420 is also used to transform the semantic expression of the structured semantic representation based on the Prompt diversity parameter control mechanism, and control the parameters of the large language model to generate text expressions of different styles; generate different versions of notification content for each different style of text expression; obtain external environment information, and fuse the external environment information with the notification content based on the context fusion mechanism to achieve dynamic enhancement of the Prompt.

[0144] In one embodiment, the draft generation module 430 is further used to design various preset roles based on the role simulation and language generation capabilities of a large language model, and to design Prompt templates for different role perspectives; using semantic elements, to generate power outage notification content corresponding to different preset role perspectives based on each Prompt template; to calculate the evaluation index of each version of the power outage notification content, and to determine the role weight corresponding to each preset role; to perform contextual summary fusion on each version of the power outage notification content, and to output a notification draft based on the evaluation index and role weight.

[0145] In one embodiment, the notification optimization module 440 is further configured to adjust the language style of the power outage notification content using a Prompt template, and introduce keyword filtering rules to standardize the terminology of the power outage notification content, thereby obtaining the adjusted power outage notification content for summary fusion.

[0146] In one embodiment, the notification optimization module 440 is also used to add new role perspective Prompt templates based on a large language model to achieve role expansion.

[0147] In one embodiment, the notification optimization module 440 is further configured to determine each target release channel and obtain the style features and format restrictions corresponding to each target release channel; display each notification draft according to each style feature and format for scenario adaptation and style optimization to obtain each preliminary notification; and perform language review and logic optimization on each preliminary notification to obtain each version of the final power grid outage notification.

[0148] In one embodiment, the model optimization module 450 is also used to output and display the final power grid outage notification, and obtain revision instructions based on the final power grid outage notification; collect user behavior and preference information based on the revision instructions, and store the user behavior and preference information in the feedback database; and optimize the large language model through the feedback database.

[0149] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 5 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When executed by the processor, the computer program implements an optimized generation method for power grid outage notifications based on a large language model. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0150] Those skilled in the art will understand that Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0151] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of an optimized generation method for power grid outage notifications based on a large language model.

[0152] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of an optimized generation method for power grid outage notifications based on a large language model.

[0153] 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. When executed, the computer program 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 may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of 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.

[0154] 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.

[0155] 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 optimizing the generation of power grid outage notifications based on a large language model, characterized in that, The method includes: Power outage data is collected from various data systems, and the power outage data is standardized and converted into unified structured power outage data. Structured power grid outage data is input into a large language model. Based on the Prompt template generation strategy and the Prompt diversity parameter control mechanism, semantic abstraction and intent modeling are performed on the structured power grid outage data to extract the semantic elements required for power outage notification. The semantic elements are used to generate corresponding power outage notification content for each different preset role perspective, and a notification draft is generated through a content fusion strategy. The draft notification is adapted to the target distribution channel and its style is optimized. The language is reviewed and the logic is optimized to obtain the final power grid outage notification for each version. Obtain revision instructions for the final power grid outage notification, receive human feedback based on the revision instructions, and optimize the large language model based on the human feedback.

2. The optimized generation method for power grid outage notifications based on a large language model according to claim 1, characterized in that, Power outage data is collected from various data systems, and this data is standardized and converted into unified structured power outage data, including: Collect various types of power grid outage data from various data source systems; The power grid outage data is processed according to a standardized field format to construct a unified structured power grid outage data. The data source system includes a power grid dispatching system, a marketing service system, and a manual supplementation system. The structured power grid outage data includes outage plans, outage equipment, user impact range, construction unit, responsible unit information, and natural language description fields.

3. The optimized generation method for power grid outage notifications based on a large language model according to claim 1, characterized in that, Based on the Prompt template generation strategy and the Prompt diversity parameter control mechanism, semantic abstraction and intent modeling are performed on the structured power grid outage data to extract the semantic elements required for the outage notification, including: Based on the Prompt template generation strategy, the structured power grid outage data is converted into a structured semantic representation for notification generation using a large language model and meaning reconstruction technology. The Prompt diversity parameter control mechanism employs a Few-shot example combination strategy to enhance the accuracy of intent understanding, performs intent modeling based on the structured semantic representation, and extracts the semantic elements required for power outage notification. The semantic elements include the background and cause of the power outage, the scope of impact and user classification prompts, the expected time for power restoration, risk warnings and external environment linkages, and safe electricity use and precautions.

4. The optimized generation method for power grid outage notifications based on a large language model according to claim 3, characterized in that, The method further includes: Based on the Prompt diversity parameter control mechanism, the semantic expression of the structured semantic representation is transformed, and the parameters of the large language model are controlled to generate text expressions of different styles. Generate different versions of notification content for various text styles; External environment information is acquired, and the external environment information is fused with the notification content based on a context fusion mechanism to achieve dynamic enhancement of the Prompt.

5. The optimized generation method for power grid outage notifications based on a large language model according to claim 1, characterized in that, The semantic elements are used to generate corresponding power outage notification content for each different preset role's perspective, and a notification draft is generated through a content fusion strategy, including: Based on the role simulation and language generation capabilities of the large language model, various preset roles are designed, and Prompt templates from different role perspectives are designed. Using the semantic elements, power outage notification content corresponding to different preset role perspectives is generated according to each of the Prompt templates; Calculate the evaluation metrics for each version of the power outage notification content and determine the role weights for each preset role. The context summaries of the power outage notifications from each version are fused, and a draft notification is output based on the evaluation metrics and role weights.

6. The optimized generation method for power grid outage notifications based on a large language model according to claim 5, characterized in that, The method includes: The language style of the power outage notification content is adjusted using the Prompt template, and keyword filtering rules are introduced to standardize the terminology of the power outage notification content. The adjusted power outage notification content is then fused with the summary.

7. The optimized generation method for power grid outage notifications based on a large language model according to claim 5, characterized in that, The method further includes: Based on the large language model, a new role-view Prompt template is added to achieve role expansion.

8. The optimized generation method for power grid outage notifications based on a large language model according to claim 1, characterized in that, The draft notification is adapted to the target distribution channel, its style is optimized, and its language and logic are reviewed and optimized to obtain the final power outage notifications for various versions, including: Identify each target publishing channel and obtain the style characteristics and format restrictions corresponding to each target publishing channel; Each of the aforementioned notification drafts is displayed according to its respective style characteristics and format, and then adapted to the scene and optimized in style to obtain each preliminary notification. Each of the preliminary notices was reviewed for language and optimized for logic to obtain the final power outage notices for each version.

9. The optimized generation method for power grid outage notifications based on a large language model according to claim 1, characterized in that, Obtain revision instructions for the final power grid outage notification, receive human feedback based on the revision instructions, and optimize the large language model based on the human feedback, including: Output and display the final power outage notification, and obtain revision instructions based on the final power outage notification; Based on the revision instructions, user behavior and preference information is collected from manual feedback and stored in the feedback database; The large language model is optimized using the feedback database.

10. A power grid outage notification optimization generation system based on a large language model, characterized in that, The system includes: The data processing module is used to collect power grid outage data from various data systems, standardize the power grid outage data, and convert it into unified structured power grid outage data. The semantic element extraction module is used to input structured power grid outage data into a large language model, and perform semantic abstraction and intent modeling on the structured power grid outage data based on the Prompt template generation strategy and the Prompt diversity parameter control mechanism to extract the semantic elements required for the power outage notification. The draft generation module is used to generate corresponding power outage notification content for each different preset role perspective using the semantic elements, and to generate a notification draft through a content fusion strategy. The notification optimization module is used to adapt the draft notification to the target release channel, optimize its style and language, and perform language review and logic optimization to obtain the final power grid outage notification for each version. The model optimization module is used to obtain revision instructions for the final power grid outage notification, receive human feedback according to the revision instructions, and optimize the large language model according to the human feedback.

Images