Power grid fault report automatic generation method combining rule driving and retrieval enhancement generation

Abstract

The invention discloses a power grid fault report automatic generation method combining rule driving and retrieval enhancement generation. According to the method, aiming at the problems of low labor efficiency, single content, insufficient template rigidity, specialty and accuracy and the like in traditional fault report generation, a high-quality power grid fault report is cooperatively generated on the basis of multi-source heterogeneous data automatic acquisition and structuralization by utilizing rule-driven information screening and combining a knowledge base and historical case retrieval enhancement. The method comprises the steps that multi-source data of scheduling, protection, operation and maintenance and the like are automatically collected and subjected to structured processing; key information elements are extracted through industry rules and expert experience; retrieving and complementing background knowledge and typical cases based on a knowledge base and a historical case base; the main paragraph of the report is automatically generated in combination with rules and retrieval content, and the smoothness and normalization of report expression are improved through an intelligent language model; and finally, quality detection and a manual auditing mechanism are integrated, so that the accuracy and specialty of the report are guaranteed. The method effectively improves the automatic generation efficiency and professional level of the power grid fault report, and is suitable for scenes such as intelligent power grid operation and maintenance management.

Description

Technical Field

[0001] This invention relates to the field of power system automation and intelligent operation and maintenance technology, specifically to an automatic generation method for power grid fault reports based on multi-source heterogeneous data fusion, rule-driven approaches, and enhanced knowledge retrieval. This method integrates power system operation data acquisition, information extraction, natural language processing, and artificial intelligence generation technologies, aiming to improve the automation level and professional expression capabilities of power grid fault reports. It is particularly suitable for application scenarios such as smart grid operation and maintenance management, fault diagnosis, emergency response, and technical archiving, meeting the needs of modern power systems for efficient, accurate, and standardized fault information processing. Background Technology

[0002] In recent years, smart grids, as an important development direction of modern power systems, rely on advanced information and communication technologies and automation technologies to achieve real-time monitoring and dynamic control of grid operation status. With the explosive growth of grid equipment and operational data, the diagnosis, analysis, and emergency response to grid faults have become increasingly complex. Fault reports, as key documents in grid operation and maintenance, not only record the fault process, cause analysis, and handling measures, but also provide a basis for subsequent experience accumulation and technological improvement, possessing significant technical and management value.

[0003] Existing fault report writing largely relies on manual work, which is inefficient and limited by the professional knowledge and writing skills of the writers, resulting in inconsistent report completeness, accuracy, and professionalism. Some power companies use template-based autofill methods, which improve the standardization of report writing to some extent, but due to the fixed template design, it is difficult to meet the diversity and complexity of fault scenarios, and the generated content generally lacks depth and personalization.

[0004] With the rapid development of artificial intelligence, especially natural language processing technology, rule-based or generative model-based automatic text generation methods have begun to be applied to the automated writing of power grid fault reports. However, current technologies still have significant limitations: on the one hand, purely rule-driven systems lack flexibility and are difficult to adapt to complex and ever-changing fault information and diverse expression needs; on the other hand, while generative AI-based solutions can produce fluent text, they are prone to content deviations due to a lack of accurate citation of professional knowledge and fact-checking, and it is difficult to ensure the standardized use of industry terminology, resulting in insufficient reliability and professionalism of automatically generated reports.

[0005] Furthermore, existing solutions generally lack transparency and interpretability in the automated report generation process, making it difficult for maintenance personnel to fully trust the generated results and increasing the burden and workload of manual review. The power industry has extremely high requirements for the accuracy, consistency, and standardization of fault reports; if automated solutions cannot adequately address these issues, large-scale deployment and application will be difficult.

[0006] Therefore, there is an urgent need for a comprehensive method that integrates intelligent collection of multi-source data, rule-driven professional information screening, semantic retrieval enhancement based on knowledge base and historical cases, and intelligent text generation and quality control to achieve automated, efficient, and high-quality generation of power grid fault reports. This would meet the power industry's demand for accurate, professional, and verifiable report content and promote technological progress in smart grid operation and maintenance management. Summary of the Invention

[0007] To address the technical bottlenecks of existing automatic power grid fault report generation methods, such as unverifiable and uninterpretable nature, low efficiency of manual verification, and lack of professional depth and flexibility in report content, this invention provides an automatic power grid fault report generation method that combines rule-driven and retrieval-enhanced generation. This method solves the problems of monotonous content and rigid expression in traditional fixed-template-based generation, overcomes the shortcomings of pure generation models in terms of factual accuracy and industry standard adaptation, while improving the transparency and traceability of the report generation process and effectively reducing the burden of manual review.

[0008] This invention proposes an automatic power grid fault report generation method that combines rule-driven and retrieval-enhanced generation, comprising the following steps:

[0009] 1) Automatically collect and structure multi-source power grid fault-related data, including dispatch records, protection actions, operation and maintenance logs, and historical cases;

[0010] 2) Based on power grid industry rules and expert knowledge base, information is filtered from structured data, and key information elements are extracted and standardized;

[0011] 3) Use semantic retrieval technology to supplement background knowledge and industry standards from the knowledge base and historical case database;

[0012] 4) By combining rule-driven paragraph templates with enhanced search content, the main parts of the fault report are automatically generated, including an event summary, cause analysis, process description, and handling measures;

[0013] 5) Utilize natural language processing and intelligent generation models to polish and verify the consistency of the report text, thereby enhancing the fluency and professionalism of the expression;

[0014] 6) The generated reports undergo automatic quality checks and support manual review and correction to ensure accuracy and completeness.

[0015] Preferably, the multi-source data is structured as follows:

[0016]

[0017] In the formula, For the kth piece of information, For timestamps, For equipment identification, For location, Fault type For signals or values, For context description.

[0018] Preferably, the rule-driven information filtering function is:

[0019]

[0020] In the formula, A structured collection of key information. For the original information set, For a set of rules, Select operators for the rules.

[0021] Preferably, the similarity calculation formula for the enhanced retrieval is:

[0022]

[0023] In the formula, and These are the semantic vectors of structured elements and knowledge base entries, respectively. This represents the cosine similarity.

[0024] Preferably, the report consistency quality scoring formula is:

[0025]

[0026] In the formula, The number of key information points accurately reproduced in the report. The total number of structured input elements.

[0027] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0028] This invention significantly enhances the richness and professionalism of power grid fault reports by integrating intelligent acquisition and structuring of multi-source heterogeneous data, combined with industry rule-driven information filtering and semantic retrieval enhancement based on knowledge bases and historical cases. The generation strategy, employing a template-driven approach combined with an intelligent language model, ensures standardized report structure, fluent expression, and high professional scalability. Simultaneously, the integration of automatic quality detection and manual review mechanisms enhances report accuracy and credibility, significantly reduces manual review workload, and improves the efficiency and consistency of fault report generation. Overall, this invention addresses multiple bottlenecks in flexibility, accuracy, verifiability, and professionalism inherent in traditional automatic generation methods, meeting the urgent need for efficient and reliable fault report generation in smart grid operation and maintenance management, and promoting the automation and intelligence levels of the power industry. Attached Figure Description

[0029] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0030] Figure 1 This is a flowchart of the technical solution of the present invention. Detailed Implementation

[0031] This invention addresses the problems of rigid templates, limited content, and insufficient knowledge utilization in existing automatic power grid fault report generation methods. It proposes a new method that combines rule-driven and retrieval-enhanced generation. This method comprehensively utilizes professional rules, knowledge base retrieval, and intelligent generation algorithms to achieve multi-source information integration, in-depth content mining, and enhanced professionalism in fault reports. The specific implementation process of this method will be described in detail below.

[0032] S1. Multi-source fault information acquisition and structuring

[0033] This invention first establishes an automatic collection and structuring system for multi-source heterogeneous data, focusing on the scenario of automatic generation of power grid fault reports. When a power grid fault occurs, relevant information is often distributed across multiple data sources, including dispatch automation systems, substation protection devices, online monitoring terminals, field operation platforms, historical fault case databases, and operation and maintenance log systems. These data vary significantly in format, structure, granularity, and update frequency. Therefore, a highly reliable data acquisition interface needs to be designed to support automatic capture and real-time synchronization of data from different types and sources.

[0034] In practical implementation, the raw data from various data sources is first adapted according to power grid information standards (such as CIM, IEC61850, etc.). Data streams from SCADA systems, DCS systems, protection and control devices, including telemetry and telecontrol data, action sequence records, equipment alarms, and operation logs, are uniformly collected into a centralized data processing module. For text-based information (such as maintenance logs, dispatch instruction records, and inspection work orders), natural language processing (NLP) technology is used for preliminary preprocessing, including text segmentation, regular expression matching, named entity recognition, and keyword extraction, to extract structured fields such as time, equipment name, geographical location, fault type, and on-site handling process.

[0035] Furthermore, for each fault-related information, a unified data entry is established, described using the following structure:

[0036]

[0037] in, This refers to the information collected for the kth item; For timestamps (the time when the event occurred or was recorded); This is a unique identifier or name for the equipment (e.g., "XX Transformer #3 Main Transformer"). Location (e.g., substation, line, tower number, etc.); The fault type (e.g., tripping, protection action, telemetry over-limit, etc.); For signals or values ​​(such as relay protection action type, telemetry value, etc.); This refers to either a contextual description or the original text.

[0038] The set of data entries can be expressed as:

[0039]

[0040] This step also involves data cleaning and normalization. For example, redundant events reported from multiple sources within the same time window need to be deduplicated and merged; for abnormal or missing data, interpolation, inference, and other methods are used to complete them. If the original data contains natural language text, key information needs to be extracted through part-of-speech tagging, phrase recognition, and other means, and stored in structured fields. In addition, fault information from the historical case database needs to be incorporated into a unified data structure for subsequent retrieval and comparison.

[0041] S2. Rule-driven information filtering and structured element extraction

[0042] After collecting and structuring multi-source fault information, the key to automatically generating high-quality fault reports lies in efficiently extracting elements with decision-making value and professional guidance from the complex and massive amount of raw data. This invention introduces a rule-driven mechanism at this stage, designing a multi-level, scalable information filtering and structured extraction strategy, enabling subsequent generation stages to work efficiently based on accurate and standardized key information sets.

[0043] Specifically, based on industry standards for power grid operation and maintenance management, and combined with multi-source knowledge such as State Grid Corporation regulations, historical fault case analysis, and expert experience databases, a "standardized dictionary of fault report elements" and a "fault type-element mapping rule set" were developed. These rules describe the filtering logic of key information in formal language. For example, for all events tagged with "protection action" or "tripping," the occurrence time, substation, name of protection device, equipment number, and action type must be extracted first. If an event involves multiple levels of equipment (such as main transformers, lines, and busbars), it is prioritized to be sorted from largest to smallest in terms of the scope of the accident's impact and written into the main report in sequence. For concurrent reports of the same event (such as remote signaling, telemetry, and operation logs), they are sorted by event time and merged into the same main report line.

[0044] This step can be formally described as follows:

[0045]

[0046] in, A structured collection of key information; The set of original information entries that were input; It is a collection of multiple domain rules; Operators for rule-based filtering and extraction.

[0047] Taking a specific implementation as an example, consider the following collected data:

[0048] "At 16:03 on May 12, 2025, the differential protection of the low-voltage side bus of the #2 main transformer of XX substation activated, the device issued a trip command, and the equipment was taken out of operation."

[0049] • "At 16:03 on May 12, 2025, the remote signaling signal of substation XX showed that main transformer #2 was disconnected and the bus voltage dropped." The system automatically applied time sorting, keyword priority, and event merging rules to extract standardized elements:

[0050] Time: 16:03, May 12, 2025

[0051] Equipment: XX Substation #2 main transformer, low-voltage side busbar

[0052] • Type of protection action: Differential protection action

[0053] Event Outcome: Tripping, equipment out of service, bus voltage drop

[0054] • Signal type: telesignaling, telemetry

[0055] Meanwhile, by setting up an "industry-specific dictionary" and a "semantic synonym unification rule", synonyms with different expressions (such as "disconnect" and "trip", "action" and "send signal") are grouped into unified element labels, thereby improving the robustness and accuracy of information extraction.

[0056] This step not only significantly improves the standardization and consistency of information screening, but also provides a high-quality, structured input foundation for subsequent search enhancement and automated generation, ensuring that the generated report covers all key events and decision-making elements, and strictly adheres to the reporting standards and professional logic of the power industry.

[0057] S3. Enhanced Retrieval Based on Knowledge Base and Historical Cases

[0058] After completing the structured extraction of key fault information, this invention introduces a retrieval enhancement mechanism based on a knowledge base and historical cases to further improve the professionalism, standardization, and relevance of the generated reports. By integrating professional knowledge, historical experience, and industry standards, the automatically generated fault reports can meet the requirements of personalized scenarios while ensuring the accuracy and industry consistency of the content.

[0059] First, build a power grid fault knowledge base. With historical case library .knowledge base Includes text entries such as national standards, company regulations, typical failure causes and handling measures, and a historical case library. It includes reports of various fault events that have actually occurred in the power grid in the past, expert debriefings, dispatch logs, and handling suggestions. All of this data has been standardized and semantically tagged for easy retrieval.

[0060] In practical implementation, the structured element set output by S2 is... A multi-level semantic retrieval algorithm is employed for relevant content matching. First, keyword indexing is used to quickly filter knowledge entries and historical cases highly relevant to the current fault type, equipment category, and regional characteristics. Then, semantic vector matching techniques (such as Sentence-BERT and WordEmbedding) are applied to structured elements. Vectorize each entry in the knowledge base and case base, and calculate the cosine similarity:

[0061]

[0062] in, For structured elements With the knowledge base entry similarity, and Its semantic vector is used. Based on a similarity threshold, the knowledge content and historical cases most relevant to the current fault scenario are selected to form a retrieval enhancement content set. :

[0063]

[0064] in, For the retrieval operator, This knowledge supplement includes standardized paragraphs, typical expressions, causal analysis, and expert advice.

[0065] Enhanced search capabilities not only fill in gaps in structured elements but also provide textual materials such as technical terms, handling procedures, and similar case experiences. For example, when a structured element indicates that "the protection action type is differential protection," the system will prioritize retrieving authoritative entries and relevant cases on "typical faults of transformer differential protection" to enrich the causal analysis and recommendations in the generated paragraphs.

[0066] In addition, to ensure real-time performance and flexibility, the search engine supports multi-dimensional filtering (such as time, device, region, fault type, etc.) and dynamic knowledge base updates, so that each report generated can incorporate the latest industry knowledge and case experience.

[0067] S4. Automatic generation combining rules and enhanced retrieval

[0068] After completing the extraction of key information and knowledge / case retrieval, this invention enters the core stage of automatic report generation. This stage adopts a collaborative strategy of "rule-driven + retrieval enhancement + intelligent generation," which ensures that the fault report has a standardized structure and accurate content, while also possessing personalized expression and professional expansion capabilities.

[0069] First, the system incorporates multiple report template sets based on power grid industry standards and company reporting specifications. Each template is divided into multiple paragraphs according to the logical structure of a standard report, such as "Event Summary", "Cause Analysis", "Process Description", "Handling Measures", and "Recommendations and Outlook". Each paragraph has several structured slots reserved for filling in key elements and supplementing search content.

[0070] The automatic generation process can be formally expressed as:

[0071]

[0072] in, For automatic report generation algorithms, This is the set of structured elements output by S2. To enhance the content set for S3 retrieval, This is a collection of paragraph templates. The specific implementation is as follows:

[0073] 1) Backbone generation (rule-driven)

[0074] The system first uses a rule-based framework to determine the overall structure of the report and the main idea of ​​each paragraph. Pre-defined fill points are set within each paragraph's framework, formed by structured elements. Prioritize data entry. For example, the "Event Summary" automatically combines elements such as fault time, equipment, and type to generate a standard header paragraph; the "Cause Analysis" prioritizes the inclusion of information such as protection action signals, fault sequence, and equipment status.

[0075] 2) Content expansion (search enhancement)

[0076] For content requiring additional background information, technical terms, processing suggestions, or similar cases, the search results obtained from S3 are retrieved based on the paragraph topic. The system automatically selects suitable text based on similarity and inserts it into the corresponding slot, dynamically enhancing both professionalism and richness. For example, if the equipment is described as "transformer differential protection operation," standard paragraphs on differential protection principles and common malfunction analyses will be automatically inserted.

[0077] 3) Smooth transition (intelligent generation)

[0078] To further enhance the overall fluency, coherence, and personalization of the report, a generative language model (such as a pre-trained language model based on the Transformer architecture, typically a GPT-like model) is used to intelligently polish and logically correct the assembled text. The system automatically identifies issues such as information duplication and incoherent sentences, ensuring that the overall style of the report meets the requirements of a professional report.

[0079] The generation process for each segment can be further refined as follows:

[0080]

[0081] in, For the first Section content, To organically fill templates and structured information with search content. This completes and refines the details of the language model.

[0082] During the output phase, the system will also automatically insert standard header and closing statements, format time and device names, automatically number, add tags and search indexes according to the company's personalized needs, thereby improving the professionalism and standardization of the report and its archival value.

[0083] S5. Report quality inspection and manual controllable audit

[0084] After automatically generating a fault report, this invention comprehensively ensures the accuracy and professionalism of the report by integrating quality inspection and manually controlled review mechanisms. The system first performs data consistency and fact-checking on the report text, comparing key information such as time, equipment, actions, and fault types in the report item by item with structured data. Its consistency score can be expressed by the following formula:

[0085]

[0086] in, For consistency score, The number of key information items accurately reproduced in the report. The total number of structured input elements.

[0087] Meanwhile, the system employs natural language processing technology to analyze the overall logic and context of the report, detecting causal relationships and the coherence of chronological descriptions to avoid factual errors or contradictions. Furthermore, the system automatically verifies whether technical terms and expressions conform to industry standards; when it detects non-standard or ambiguous expressions, it generates correction suggestions or automatically replaces them with standard terminology. Regarding report structure and readability, the system ensures that all standard paragraphs are covered, the content is complete, and the expression is fluent, and it assesses text quality using quality evaluation metrics such as BLEU and ROUGE.

[0088] If all automatic checks pass, the report can be directly archived and pushed. If key information is missing or quality risks are found, the system will mark the report as "pending manual review" and generate detailed correction suggestions to facilitate timely supplementation and improvement by maintenance personnel. Through the above mechanism, this invention not only improves the efficiency of automatic fault report generation but also ensures the accuracy and standardization of the final report, effectively supporting the high standards of operation and maintenance and archiving management in the power industry.

[0089] This invention constructs an intelligent and standardized method for automatically generating power grid fault reports through five steps: multi-source fault information collection and structuring, rule-driven information filtering, enhanced retrieval based on knowledge base and historical cases, automatic generation combining rules and retrieval, and report quality detection and manual review. This method not only achieves comprehensive collection and element extraction of various types of power grid fault information, but also integrates rich professional knowledge and historical experience. Through multi-technology collaboration, it automatically generates fault reports that are standardized in content, logically clear, and highly professional. Before output, automatic and manual quality control ensures the accuracy and practicality of the reports, providing efficient and reliable support for intelligent operation and maintenance of the power grid.

Claims

1. A method for automatically generating power grid fault reports by combining rule-driven and retrieval-enhanced generation, characterized in that, Includes the following steps: Step 1: Automatically collect and structure multi-source power grid fault-related data, including dispatch records, protection actions, operation and maintenance logs, and historical cases; Step 2: Based on power grid industry rules and expert knowledge base, the structured data is screened, and key information elements are extracted and standardized. Step 3: Perform semantic retrieval using the knowledge base and historical case database to supplement the background knowledge and industry standards required for the report; Step 4: Combining rule-driven paragraph templates with enhanced retrieval content, automatically generate the main parts of the fault report, including event summary, cause analysis, process description, and handling measures; Step 5: Utilize natural language processing techniques and generative models to optimize the language and verify consistency of the report, thereby improving the fluency and professionalism of the report's expression; Step 6: Perform automatic quality checks on the generated report, and support manual review and correction, ultimately outputting a qualified power grid fault report.

2. The method for automatically generating power grid fault reports according to claim 1, characterized in that, The multi-source data acquisition includes data from SCADA systems, protection and control devices, dispatch automation systems, operation and maintenance work orders, and historical case databases that are automatically accessed and integrated.

3. The method for automatically generating power grid fault reports according to claim 1, characterized in that, The rule-driven information filtering includes setting multi-level element extraction rules based on industry standards and expert experience to achieve automatic extraction and normalization of key information such as event type, device name, and action sequence.

4. The method for automatically generating power grid fault reports according to claim 1, characterized in that, The enhanced retrieval employs semantic retrieval and keyword matching to filter entries related to the current fault scenario from the knowledge base and historical case database, and automatically completes the report content.

5. The method for automatically generating power grid fault reports according to claim 1, characterized in that, The automatic generation step combines preset templates with intelligent language models, ensuring both a standardized report structure and natural, professional expression.

6. The method for automatically generating power grid fault reports according to claim 1, characterized in that, The quality inspection includes automatically checking the consistency of key information, the accuracy of facts, the standardization of professional terminology, and the completeness of the text in the report, and generating correction suggestions.

Images