Automatic auditing and checking system and method for operation safety of transformer substation

Abstract

The invention discloses an automatic auditing and checking system and method for substation operation safety. The system comprises an automatic data acquisition module, a data semantic fusion analysis module, a dynamic auditing rule generation module, a risk collaborative prediction module, an automatic information pushing module and a visual interaction configuration module. All the modules are connected through an intranet data interface and a system internal communication protocol, aiming at the prominent problems that the auditing process depends on manpower, the information system is split, the system semantics is difficult to understand, the risk identification is not comprehensive, the rule adaptability is poor, the response and push chain is broken and the like commonly existing in the existing substation operation auditing work. The invention provides an automatic auditing and checking system and method for substation operation safety, and aims to construct a full-process closed-loop intelligent auditing platform with semantic comprehension capability, multi-source data fusion capability, risk collaborative prediction capability and dynamic rule evolution capability.

Description

Technical Field

[0001] This invention relates to an automatic verification system and method for substation operation safety. The secondary structure of the primary permanent magnet linear motor with a segmented secondary design is simple and suitable for long-distance rail transit systems that require high reliability and low cost. Background Technology

[0002] With the deepening implementation of the national energy strategy and the continuous improvement of the digitalization and intelligence level of the power system, substations, as key nodes of the power grid, have seen their operation and maintenance operations become a crucial factor restricting the safe and stable operation of the overall power grid, particularly in the current context of "reducing manpower and increasing efficiency" and building "inherently safe work teams." How to achieve intelligent auxiliary decision-making and dynamic safety verification throughout the entire operation process has become a core issue for power companies' substation operation and maintenance management.

[0003] Substation on-site operations involve multiple dimensions of data and regulations, including work permit approval, personnel qualification review, equipment status confirmation, work environment assessment, and safety measure implementation. In the traditional model, these review and verification processes heavily rely on human experience and manual operation, resulting in low efficiency, information lag, incomplete coverage, and weak risk identification capabilities. For example, on-site auditors must manually review work permits, personnel qualifications, defect records, and risk data from multiple systems, checking each item against regulations before work commences. This process is complex, cumbersome, and prone to safety hazards due to omissions or misjudgments.

[0004] In recent years, with the gradual introduction of technologies such as artificial intelligence (AI), natural language processing (NLP), big data analytics, and RPA (Robotic Process Automation), the power industry has made positive progress in intelligent operation and maintenance. For example, some enterprises have explored applications such as RPA-based report generation, NLP-based Q&A for regulations and rules, and knowledge graph-based experience retrieval. However, most systems are still in the initial stage of "assisting decision-making" rather than "automatic review," lacking a full-process semantic understanding of the operational scenario and a systematic closed-loop review process.

[0005] Furthermore, current power grid companies have multiple internal business systems, such as work order management platforms, personnel qualification databases, defect and risk management systems, and equipment inspection systems. The lack of unified data standards and limited interface openness between these systems results in lengthy information acquisition paths and difficulties in data fusion, making it impossible to support the need for automated review and dynamic early warning throughout the entire process from pre-operation to operation and post-operation. While some units use manual rule compilation to create review item lists, they lack intelligent update mechanisms, resulting in weak adaptability and poor timeliness of the review rules, making them unable to adapt to complex and ever-changing on-site operating environments.

[0006] Furthermore, with the rapid growth of job data and the increasing number of hidden risks, relying solely on static rules is insufficient to accurately identify potential risks. How to leverage advanced technologies such as semantic recognition, reinforcement learning, and federated collaborative modeling to achieve intelligent evolution of job review strategies while ensuring data security is a key direction for current technological development.

[0007] Therefore, there is an urgent need to build an intelligent auditing system for substation operation scenarios, which has the capabilities of automatic data collection, semantic analysis, adaptive rule generation, risk prediction and linkage, intelligent information push and interface visualization configuration. This system will truly realize the automation, intelligence and sustainable optimization of safety auditing throughout the entire operation process, and promote the transformation of substation operation and maintenance work towards a new model of "less manned operation, intelligent collaboration and safety closed loop".

[0008] Currently, for safety audits before substation operations, most organizations still use a traditional "semi-automatic + manual confirmation" approach. This primarily relies on multiple independent information systems, such as work order management systems, personnel qualification management platforms, and defect and risk management systems, for manual querying and data verification. The core logic of this approach is that the system provides the relevant data query interface or export function, while the audit responsibility still rests with on-site or back-end maintenance personnel.

[0009] Specifically, in the work plan approval process, the reviewers usually need to log in to the work order management system to view the work task content, work time, workers, and configured safety measures; then switch to the personnel qualification management platform to confirm whether the personnel involved have the corresponding professional category and valid work qualifications; then enter the defect and risk management system to manually search for whether there are any unclosed defects or high-risk hazards in the relevant equipment; finally, some units also require manual review of auxiliary information such as recent equipment inspection records and work weather conditions.

[0010] Taking a substation operation and maintenance auxiliary platform deployed by a provincial power grid company as an example, its work ticket management subsystem provides functions for ticket generation, process approval, and ticket archiving, but lacks intelligent data verification capabilities. The personnel qualification system can export lists of personnel certificates and violation records, but lacks the ability to link and verify these records with the personnel field of the work ticket. The equipment defect ledger and risk ledger systems provide equipment-level defect history, but the information granularity is coarse and the classification standards are inconsistent, requiring manual identification and integration. Furthermore, some units have configured manually created audit rule lists to guide on-duty personnel or managers in their checks, but these rule lists need to be updated regularly, and their execution efficiency is limited by individual experience and execution ability.

[0011] It is worth mentioning that some organizations have tried to use customized Excel or auto-fill templates to achieve semi-automatic data alignment. For example, they can preset personnel lists and qualification expiration dates, or define key equipment and high-risk keywords to achieve simple condition matching prompts. However, these methods lack system-level linkage and semantic recognition capabilities, and cannot cope with complex situations such as diverse job descriptions, inconsistent system interfaces, and multi-dimensional information fusion.

[0012] At the data application level, existing technologies mostly remain at the stage of information display and manual comparison, and have not yet formed a closed-loop system of semantic understanding and rule-driven review. Moreover, there are large differences in data structure, interface standards and calling methods among various data systems, which leads to difficulties in integration, low efficiency and difficulty in meeting the needs of "full process and full elements" automatic review.

[0013] Although power grid companies have established various business systems such as work permit management systems, personnel qualification systems, and equipment defect and risk management platforms, these existing technical solutions still have many prominent problems in substation operation review practices, mainly in the following aspects: 1. The review process relies on human experience, which is inefficient and prone to omissions. Various information systems have relatively independent functions, requiring auditors to switch between multiple systems, manually search for relevant information, and compare audit elements. This makes them highly susceptible to oversights due to incomplete information retrieval, identification errors, or auditor fatigue. Especially during periods of high-frequency operation, the audit workload is heavy and the risks are multiplied.

[0014] 2. Data is fragmented between systems, making information integration difficult. The data structures, calling interfaces, and naming conventions of different systems vary greatly, and there is a lack of a unified semantic parsing and standardized processing mechanism. This makes it difficult to achieve semantic-level fusion and linkage analysis of multi-dimensional data such as work tickets, personnel qualifications, equipment defects, and risk levels, which affects the systematicness and accuracy of the review and judgment.

[0015] 3. Lacks intelligent semantic analysis capabilities and cannot adapt to diverse scenarios. The current review process is mainly based on field comparison and static rule lists, which cannot understand the implicit meanings in unstructured texts such as job descriptions, personnel history, and equipment risk contexts, and makes it difficult to identify potential risks such as vague descriptions, alternative expressions, or missing key conditions.

[0016] 4. Rigid rule configuration, unable to dynamically adapt to changes in risk scenarios. Traditional rule lists require regular maintenance by administrators and cannot be automatically optimized or evolved based on audit results and risk events reported during actual operation. This leads to a disconnect between audit strategies and actual on-site conditions, resulting in situations where "rules are not applicable" and "problems cannot be identified".

[0017] 5. Unable to achieve proactive early warning and closed-loop tracking. Existing systems mostly provide passive information display and lack closed-loop capabilities such as alarm generation, risk push notification, personnel confirmation, response recording and traceability. This makes it difficult to transmit high-risk operation information to key positions in a timely manner, resulting in potential risks such as delayed early warning and unclear responsibilities.

[0018] 6. Unable to support unified portal and visual operation The current user interface is diverse and the data is presented in a fragmented manner. Users need to have strong information technology skills and system usage experience, which makes it difficult to promote and train users. In particular, the user experience is poor in grassroots stations or high-voltage field operations.

[0019] In summary, existing technologies suffer from severe information silos, rigid rules, weak analytical capabilities, and broken response chains when dealing with complex, dynamic, and multi-source integrated substation operation scenarios. These issues make it difficult to meet the high standards and intelligent requirements of the new generation of power grids for "intrinsically safe operation audit systems." There is an urgent need to build a closed-loop, controllable, and dynamically optimized operation safety audit system by integrating technologies such as semantic analysis, fusion modeling, automatic rule generation, and intelligent push.

[0020] In addition to the traditional approach of combining manual review with work order systems, some power grid companies have introduced auxiliary operation and maintenance management systems, equipment operation and maintenance ledger platforms, dispatch support systems, and team evaluation systems in recent years to support the management, recording, and evaluation of daily substation operations. The core of these systems is "replacing some manual supervision with data statistics," primarily focusing on management needs such as business process documentation, data ledger improvement, and performance evaluation. However, they have not truly achieved dynamic, automated review and risk assessment capabilities before, during, and after operations.

[0021] Taking the widely deployed "Substation Operation Management System" of a certain power grid company as an example, the system has functions such as work order process management, team member attendance records, equipment inspection data uploading, and work task planning. Users can enter and retrieve information such as work tasks, work orders, defect statistics, and inspection conclusions by module in the system, realizing information archiving and tracking of the entire operation process. At the same time, the system also supports the output of key business indicators such as "two-ticket execution rate," "defect closure rate," and "inspection completion rate" for managers to evaluate the station's operational quality and personnel performance.

[0022] Another type of system, such as Equipment Lifecycle Management (EAM) systems or integrated equipment ledger platforms, focuses on establishing asset ledgers and managing maintenance plans. Users can register equipment operating status, historical maintenance records, defect work orders, risk assessment reports, etc., online, providing basic support for long-term equipment health management. Some organizations have also developed independent "risk assessment subsystems" or "operational anomaly recording systems" to record past operational cases involving violations, hazardous environments, or unresolved potential hazards, for use in back-end approvals and as a reference for experience.

[0023] Furthermore, in practical applications, some organizations have attempted to use a "fixed template + automatic reporting" approach, inputting elements such as inspection records, equipment defect photos, and meteorological data through system-defined fields, and then generating reports or warning messages to be pushed to WeChat groups, station administrator emails, and other channels. While this mechanism can assist in information integration during the operational preparation phase to some extent, the information display still primarily relies on raw data stacking and lacks multi-dimensional fusion and analysis capabilities.

[0024] Overall, these existing technologies focus more on "data storage and statistical management" rather than "semantic understanding and audit decision-making" in terms of functionality; they adopt field-based and form-based models in data relationship processing and have not yet formed a unified problem-driven knowledge structure; in terms of system responsiveness, they also lack the ability to automatically call rules, make semantic judgments and provide intelligent feedback, making it difficult to meet the needs of high-intensity and high-complexity operational security audit scenarios.

[0025] Although some units have already implemented various substation operation management systems and equipment ledger systems to improve the process and data-driven level of operation management, these systems still expose a series of unavoidable technical shortcomings in actual operation, mainly in the following aspects: 1. The system's functionality is heavily focused on management, lacking sufficient auditing capabilities. Existing systems are mostly used for management tasks such as recording historical data, tracking work order processes, and outputting statistical indicators. They do not provide support functions for key aspects such as pre-operation review, risk identification, and dynamic strategy intervention, making it difficult to meet the review requirements such as "whether the operation is compliant and whether the risk is hidden".

[0026] 2. The data structure is closed, and the system and data are not integrated. Various institutional documents (such as operating procedures, qualification requirements, and safety measures standards) are not included in the structured system and exist only as attachments or materials for manual review. As a result, although the system has data, it cannot establish a semantic mapping between "institutions - tasks - behaviors", and the review lacks logical support.

[0027] 3. Lack of problem-driven entry points; users are forced to passively search for information. The system interface and logic mostly organize information in a "module-field-directory" manner, which cannot automatically extract key information to form auxiliary decision-making suggestions based on the problems or scenarios faced by users. It heavily relies on users' experience in using the system and their ability to understand the scenario, resulting in a steep learning curve.

[0028] 4. Lack of intelligent linkage and semantic response mechanisms The lack of cross-data access paths between subsystems prevents the system from automatically performing interconnected analysis and judgment of relevant data when frontline personnel raise questions or submit tasks. For example, the system cannot identify scenario-based questions such as "Does this task involve defective equipment?" or "Has the licensor violated regulations before?", lacking the ability to respond to audit objectives.

[0029] 5. Significant interface limitations make it difficult to integrate into intelligent review processes. Most management systems are deployed in a closed manner, with inconsistent interface standards and limited data openness, making it difficult to seamlessly connect with next-generation intelligent systems, especially lacking the technical foundation in areas such as RPA automatic retrieval, AI model linkage, and closed-loop message push.

[0030] 6. Low data utilization rate, and no knowledge feedback and optimization loop has been formed. Information such as risk characteristics, approval defects, and accident cases generated during the operation process has not been converted into structured knowledge, and cannot be used to feed back into subsequent review strategy optimization or rule evolution. The system lacks self-learning and self-enhancing capabilities and relies on human experience accumulation and regular template updates in the long term.

[0031] In summary, while existing management platforms and ledger systems have promoted the digitization of data and the standardization of processes in substation operations to some extent, there are significant gaps in the intelligent audit chain of "semantic understanding - risk identification - strategy generation - result push - closed-loop feedback," making it difficult to build the "self-sensing, self-analysis, and self-decision-making" operational safety assurance capabilities required for the future smart grid. Summary of the Invention

[0032] To address the prominent problems commonly found in existing substation operation review processes, such as reliance on manual review processes, fragmented information systems, difficulty in understanding the semantics of regulations, incomplete risk identification, poor rule adaptability, and broken response and push chains, this invention proposes an automatic review and verification system and method for substation operation safety. The aim is to build a closed-loop intelligent review platform with semantic understanding capabilities, multi-source data fusion capabilities, risk collaborative prediction capabilities, and dynamic rule evolution capabilities.

[0033] To achieve the above-mentioned objectives, the present invention employs the following technical solution: An automatic verification and validation system for substation operation safety, the system comprising an automatic data acquisition module, a data semantic fusion and analysis module, a dynamic verification rule generation module, a risk collaborative prediction module, an automatic information push module, and a visual interactive configuration module; The modules are connected via an intranet data interface and an internal system communication protocol. The automatic data acquisition module is used to acquire various types of substation production data; The data semantic fusion analysis module extracts the data, and after semantic feature extraction and fusion analysis, identifies risk characteristics; The dynamic audit rule generation module calls upon the aforementioned risk features to perform reinforcement learning training and rule optimization. The risk collaborative prediction module generates a global risk prediction model based on the local prediction model and the federated learning collaborative mechanism, and feeds back the prediction results; The automatic information push module receives the output of the global risk prediction model, generates alarm information, and pushes it. The visual interactive configuration module provides an interactive interface for all modules, enabling parameter configuration and monitoring management.

[0034] Furthermore, the automatic data acquisition module includes an interface calling unit, a data crawling unit, a data cleaning unit, and a structured storage unit, wherein... The interface calling unit is used to call the data interfaces of the power grid management platform and the substation operation support system to obtain various important production data involved in the substation operation process; The data crawling unit is used to automatically obtain various important production data involved in the substation operation process through intranet RPA technology or web crawlers when there is no interface data or interface calls are restricted. The data cleaning unit is used to perform data denoising, outlier removal, null value processing, and field standardization on the raw data of various important production processes involved in the substation operation, generating a unified standard data format. Structured storage units are used to classify and store cleaned data in a relational database, forming a standardized job safety audit database.

[0035] Furthermore, the data semantic fusion analysis module includes a semantic feature extraction unit, a data fusion unit, and a risk feature identification unit, wherein, The semantic feature extraction unit is fine-tuned based on the Transformer large language model to automatically extract key semantic features from text data such as work task descriptions, safety measure descriptions, personnel qualifications, and defect information. The data fusion unit intelligently matches and merges various important production data from different sources according to their semantic features, generating a unified semantic feature database and clarifying the relationship between related work tasks, personnel qualifications, equipment defects, and risk information. The risk feature identification unit automatically identifies potential safety risk features in the operation based on the fused semantic feature database and clearly marks the risk level of the data.

[0036] Furthermore, the dynamic review rule generation module includes a reinforcement learning training unit, a rule optimization unit, and a rule base maintenance unit, wherein, The reinforcement learning training unit constructs a deep reinforcement learning environment based on historical audit data and risk identification results, trains an automated audit decision-making model, and dynamically generates optimal rules. The rule optimization unit automatically adjusts the rule priority and weight based on feedback; The rule base maintenance unit updates the optimized audit rules to the rule base in real time to ensure the continuous effectiveness and dynamic adaptability of the rule base.

[0037] Furthermore, the risk collaborative prediction module includes a local prediction unit, a federated learning collaboration unit, and a global model application unit, wherein, The local prediction unit trains a safety risk prediction model based on local audit and risk data to assess potential safety risks at the local substation. The federated learning collaborative unit uses a federated learning algorithm to safely aggregate the local model parameters of multiple substations and construct a global risk prediction model. The global model application unit applies the global model to each substation to provide guidance for risk prediction.

[0038] Furthermore, the automatic information push module includes an alarm generation unit, a communication interface unit, and a message push unit, wherein, The alarm generation unit automatically generates standardized alarm information and audit reports based on the audit and prediction results; The communication interface unit is used to connect to the intranet Dify platform, elink data interface or mobile SMS interface to prepare the push channel. The message push unit is used to call the communication interface unit to quickly push alarm information to the corresponding operation and management personnel via elink or SMS.

[0039] Based on the aforementioned automatic verification and validation system for substation operation safety, this application also provides an automatic verification method for substation operation safety, the method comprising the following steps: Step 1: Obtain and integrate various important production data involved in substation operations in real time from the power grid production system and operation and maintenance support system; Step 2: Based on the pre-trained large language model, perform deep semantic analysis, intelligent fusion, and risk identification on the various important production data obtained in Step 1; Step 3: Intelligently train historical task data using reinforcement learning technology to automatically generate an audit rule base that can adapt to different task scenarios; Step 4: By constructing a collaborative risk prediction mechanism among substations, we can achieve early detection and warning of potential safety hazards. Step 5: Implement an intelligent alarm mechanism to accurately notify and promptly report on the results of work review and potential risks; Step 6: Configure the system, adjust rules, edit risk information, and monitor the operational status through a visual interface.

[0040] Furthermore, step 1 specifically includes: Step 1.1: The automatic data acquisition module, based on intranet data interface technology, automatically establishes a secure communication connection with the work order management module of the power grid management platform, and automatically obtains detailed information on all work orders currently being executed at all substations in real time, including the specific content of the work task, the work location, the start and end times of the work plan, the personnel information in the approval process, and detailed fields of the implemented safety measures. Step 1.2: The automatic data acquisition module automatically calls the personnel qualification module interface in the power grid management platform to quickly and efficiently query the qualification validity of the relevant personnel corresponding to each work order, including the specific type of qualification, validity period and historical safety violation records. The system automatically identifies and filters abnormal qualification information. Step 1.3: Call the interfaces of the integrated power outage management module and the defect management module in the power grid management platform to accurately obtain the maintenance content and status requirements of the equipment involved in the operation, as well as the latest defect information of the relevant equipment, and automatically form a detailed equipment status database; Step 1.4: The automatic data acquisition module calls the data interface of the power grid risk management module to automatically obtain the latest power grid risk data related to the equipment of the operation task, including the potential risks, risk levels and their specific descriptions for the equipment, so as to form an effective risk data source. Step 1.5: Connect to the data interface of the substation operation support system to automatically acquire on-site and equipment inspection data in real time, including environmental meteorological data, recent inspection photos of equipment, and temperature detection data; Step 1.6: The data cleaning and structured storage unit automatically receives data streams from multiple channels, standardizes the raw collected data, automatically removes outliers and invalid data, unifies the data format, and finally stores it in the structured database.

[0041] Furthermore, step 2 specifically includes: Step 2.1: The semantic feature extraction unit calls a large language model that has been fine-tuned based on the corpus in the field of substation operation and maintenance to perform semantic vectorization encoding on the textual descriptions of work tasks and safety measures in the work order, and extract high-level semantic features of key equipment, operation behavior, work type and execution conditions. Step 2.2: Using semantic vector matching and entity alignment algorithms, the semantic features of the operation extracted in Step 1 are semantically fused with heterogeneous data such as personnel qualifications, equipment defects, power outage maintenance orders, and risk records to form a unified semantic information graph of the operation. Step 2.3: Based on the fused semantic graph, the risk identification model is invoked to identify and mark potential violations, conflicts or omissions in the operation, including: personnel qualifications not matching, operation scope exceeding, equipment risks not closed-loop, and safety measures inconsistent with tasks. The identification results are accompanied by a credibility score and risk level label.

[0042] Furthermore, step 3 specifically includes: Step 3.1: Use deep reinforcement learning algorithm to establish a learning model with the audit behavior as the action space and the risk identification result as the state space. Use past actual operation data as training samples to simulate the risk control effect brought about by different rule combinations in the audit process. Use the indicators of reduced operation accident rate and improved audit consistency as reward function to iteratively optimize the audit strategy. Step 3.2: The reinforcement learning model will evaluate the effectiveness of each review rule in actual operation in real time and dynamically adjust the execution weight of the rule; Step 3.3: Establish a rule base maintenance mechanism to periodically evaluate and iterate the review rules; Furthermore, step 4 specifically includes: Step 4.1: Deploy a lightweight risk prediction model locally at each substation. Based on the historical work tasks, personnel qualification fluctuations, equipment defect recurrence rate, and correlation data between meteorological conditions and accidents at each station, train a risk assessment model with local characteristics.

[0043] Step 4.2: Without transmitting the original data, the model parameters trained locally at multiple substations are aggregated to the central node using federated learning technology and merged under a secure computing mechanism to form a global risk prediction model. Step 4.3: The global model optimized by federated learning will be automatically distributed to each site on a daily basis or triggered by major events, for real-time updates to the local prediction engine.

[0044] The present invention has the following beneficial effects: This invention integrates multiple key technologies, including Natural Language Processing (NLP), Large Language Modeling (LLM), RPA process automation, reinforcement learning, federated learning, and visual configuration, to connect various heterogeneous data sources such as the job management system, personnel qualification database, equipment defect and risk system, and inspection environment data platform, thereby achieving the following technical objectives: 1. Achieve automatic collection and standardized integration of multi-source operation data. It automatically connects to the data interfaces of multiple business systems such as the power grid management platform and the substation operation support system, or collects data through RPA crawler technology, cleans, standardizes and stores it in a structured manner, and establishes a high-quality data foundation for operation review.

[0045] 2. Possesses deep semantic recognition and information fusion capabilities. Based on the fine-tuned large language model, semantic features are extracted from unstructured text content such as job tasks, safety measures, and equipment information. These features are then integrated with structured data such as qualifications, defects, and risks to establish a semantic graph of "task-equipment-personnel-risk".

[0046] 3. Construct a dynamically evolving mechanism for generating review rules. By employing reinforcement learning algorithms, the review strategy is dynamically trained based on data such as historical review results and risk response effects, continuously optimizing the priority, judgment logic, and weight of review rules, and realizing an intelligent review decision-making system that is "self-learning and self-optimizing".

[0047] 4. Achieve collaborative prediction and closed-loop notification of operational risks. By combining federated learning technology, local risk models from various sites are aggregated to achieve high-precision global risk prediction. Alarm information is pushed to responsible personnel through multiple channels (such as Elink, SMS, and mobile APP), and response behavior is recorded to ensure information reachability and result traceability.

[0048] 5. Provides visual interaction and configuration capabilities. It allows users to flexibly configure and adjust data interface call strategies, audit rule parameters, risk alarm thresholds, etc., and monitor the system's operating status, push records, risk distribution, etc. in real time, thereby improving the system's usability, maintainability, and adaptability for promotion.

[0049] By integrating and implementing the above key technical solutions, this invention can realize a complete closed-loop process of "automatic perception, intelligent analysis, strategy matching, alarm push, and feedback optimization" for substation operation review, significantly improving the efficiency and accuracy of pre-operation safety review, promoting the transformation from "human-based problem finding" to "system-based problem finding", and ensuring the inherent safety and standardization level of substation operation and maintenance work. Attached Figure Description

[0050] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is an example diagram of the overall architecture of an automatic verification and validation system for substation operation safety provided in an embodiment of the present invention; Figure 2 This is a flowchart of an automatic substation operation safety verification method based on the above-mentioned automatic substation operation safety verification system, provided by an embodiment of the present invention. Detailed Implementation

[0051] The technical solutions of the invention will now be described in detail with reference to the accompanying drawings, further clarifying and completely describing the technical solutions in the embodiments of the invention. It should be noted that the described embodiments are merely a part of the implementation examples of the invention, not all of them. All other implementation examples obtained by those skilled in the art based on the implementation examples of the invention without inventive effort are within the scope of protection of the invention.

[0052] Example 1:

[0053] The first aspect of this invention proposes an automatic verification and validation system for substation operation safety. The system includes an automatic data acquisition module, a data semantic fusion and analysis module, a dynamic verification rule generation module, a risk collaborative prediction module, an automatic information push module, and a visual interactive configuration module. The modules are connected via an intranet data interface and an internal system communication protocol, ensuring clear and unambiguous data flow. Specific definitions and functional descriptions are as follows: (1) Automatic data acquisition module: The automatic data acquisition module includes an interface calling unit, a data crawling unit, a data cleaning unit, and a structured storage unit, and its functions are defined as follows: 1) Interface calling unit: used to call the data interfaces of the power grid management platform and the substation operation support system to obtain work order information, personnel qualification information, power outage maintenance order information, defect management information, risk management information and equipment inspection environment information, etc.

[0054] 2) Data crawling unit: When there is no interface data or the interface call is restricted, the relevant data in the above production system is automatically obtained through intranet RPA technology or web crawler.

[0055] 3) Data cleaning unit: Performs data denoising, outlier removal, null value processing, and field standardization on the collected raw data to generate a unified standard data format.

[0056] 4) Structured storage unit: The cleaned data is classified and stored in a relational database to form a standardized work safety audit database.

[0057] (2) Data semantic fusion analysis module: The data semantic fusion analysis module includes a semantic feature extraction unit, a data fusion unit, and a risk feature identification unit, and its functions are defined as follows: 1) Semantic feature extraction unit: Based on the Transformer large language model, it is fine-tuned to automatically extract key semantic features from text data such as job task descriptions, safety measure descriptions, personnel qualifications, and defect information.

[0058] 2) Data fusion unit: Intelligently matches and merges data from different sources according to their semantic features to generate a unified semantic feature database, clarifying the relationships between related work tasks, personnel qualifications, equipment defects and risk information.

[0059] 3) Risk Feature Identification Unit: Based on the fused semantic feature database, it automatically identifies potential safety risk features in the operation, clearly marks the risk level of the data, and provides a basis for subsequent review.

[0060] (3) Dynamic review rule generation module: The dynamic review rule generation module includes a reinforcement learning training unit, a rule optimization unit, and a rule base maintenance unit, and its functions are defined as follows: 1) Reinforcement learning training unit: Based on historical audit data and risk identification results, a deep reinforcement learning environment is built to train an automated audit decision model and dynamically generate the optimal audit strategy.

[0061] 2) Rule Optimization Unit: Continuously evaluates the accuracy and risk capture rate of audit rules in actual application, and automatically adjusts rule priority and weight based on feedback.

[0062] 3) Rule base maintenance unit: Updates the optimized audit rules to the rule base in real time to ensure the continuous effectiveness and dynamic adaptability of the rule base.

[0063] (4) Risk Collaborative Prediction Module: The risk collaborative prediction module includes a local prediction unit, a federated learning collaboration unit, and a global model application unit, and its functions are defined as follows: 1) Local Prediction Unit: This unit trains a safety risk prediction model using local audit and risk data to assess potential safety risks at local substations. In this embodiment, "local audit" refers to each substation itself, such as the audit of work at substation A. The local prediction unit uses the risk data from substation A itself to train the safety risk prediction model, independent of other substations. The model parameters are called the local model parameters of substation A. Therefore, federated learning refers to aggregating the local model parameters of multiple substations to form a global risk prediction model. This model can be understood as knowing the risk correlation points between different substations, such as whether work at substation A will affect substation B, etc.

[0064] 2) Federated Learning Collaborative Unit: Under the premise of ensuring data privacy, the federated learning algorithm is used to achieve secure aggregation of local model parameters of multiple substations and build a high-precision global risk prediction model.

[0065] 3) Global Model Application Unit: The global model formed by federated learning is applied to each substation to provide risk prediction guidance and realize cross-site risk linkage and collaborative early warning.

[0066] (5) Automatic Information Push Module: The automatic information push module includes an alarm generation unit, a communication interface unit, and a message push unit, and its functions are defined as follows: 1) Alarm generation unit: Based on the audit and prediction results, automatically generate standardized alarm information and audit reports, including site name, work order number, task, person in charge information and specific risks or audit issues.

[0067] 2) Communication Interface Unit: Responsible for connecting to the intranet Dify platform, elink data interface or mobile SMS interface, and preparing the push channel.

[0068] 3) Message push unit: Automatically calls the communication interface unit to quickly push alarm information to the corresponding operation and management personnel via elink or SMS, ensuring timely handling of alarms and risks.

[0069] (6) Visual Interactive Configuration Module: The visual interactive configuration module includes a parameter configuration unit, a rule adjustment unit, a risk warning configuration unit, and a system monitoring unit, the functions of which are defined as follows: 1) Parameter configuration unit: Provides an interface for configuring data interface, acquisition frequency and storage period, and supports flexible customization of the data acquisition process.

[0070] 2) Rule Adjustment Unit: Provides a manual adjustment interface for dynamic review rules of reinforcement learning, allowing users to intervene and modify the weight and priority of review rules.

[0071] 3) Risk warning configuration unit: Provides a customizable editing interface for risk warning information templates, allowing users to freely set the thresholds and conditions for triggering alarms and flexibly control risk warning strategies.

[0072] 4) System monitoring unit: Displays real-time system operating status, data collection status of each module, audit results, alarm push records, and other information, so that users can understand the system operating status in a timely manner and quickly troubleshoot abnormal problems.

[0073] (7) Explanation of data flow between the above modules: 1) The automatic data acquisition module obtains various substation production data through the intranet interface or crawler. After the data is cleaned by the data cleaning unit, it is stored in the database. The data flow is: interface call unit / data crawling unit → data cleaning unit → structured storage unit.

[0074] 2) The data semantic fusion analysis module extracts data from the structured storage unit, and after semantic feature extraction and fusion analysis, it flows to the risk feature identification unit: structured storage unit → semantic feature extraction unit → data fusion unit → risk feature identification unit.

[0075] 3) The dynamic audit rule generation module calls the output data of the risk feature identification unit to perform reinforcement learning training and rule optimization, and updates the rule base: Risk Feature Identification Unit → Reinforcement Learning Training Unit → Rule Optimization Unit → Rule Base Maintenance Unit.

[0076] 4) The risk collaborative prediction module generates a global risk prediction model based on the local prediction model and the federated learning collaborative mechanism, and feeds back the prediction results: local prediction unit → federated learning collaborative unit → global model application unit.

[0077] 5) The automatic information push module receives the outputs of the rule base maintenance unit and the risk collaborative prediction module, generates alarm information and pushes it: rule base maintenance unit / global model application unit → alarm generation unit → communication interface unit → message push unit.

[0078] 6) The visual interactive configuration module provides an interactive interface for all the above modules, enabling parameter configuration and monitoring management, with bidirectional flow of data and instructions: User interface operation ←→ Parameter configuration unit / rule adjustment unit / risk warning configuration unit / system monitoring unit ←→ Each functional module.

[0079] Example 2:

[0080] A second aspect of the present invention provides an automatic verification method for substation operation safety based on the above-mentioned automatic verification and validation system for substation operation safety, the method comprising the following steps: Step 1, Automatic Data Acquisition, Cleaning, and Storage: This step automatically acquires and integrates various important production data related to substation operations in real time from the power grid production system and operation and maintenance support system to ensure the efficiency and accuracy of the audit process. The specific implementation steps are as follows: Step 1.1, Automatic Extraction of Work Permit Information: The automatic data acquisition module, based on intranet data interface technology, automatically establishes a secure communication connection with the work permit management module of the power grid management platform, and automatically obtains detailed information on all work permits currently being executed at all substations in real time, including detailed fields such as the specific content of the work task, the work location, the start and end times of the work plan, the personnel status in the approval process, and the implemented safety measures, ensuring the completeness and timely acquisition of data.

[0081] Implementation Case: For example, this implementation plan takes the work order numbered 2025-03-C1-003 carried out on March 10, 2025 at a certain 110kV substation as an example. The system can automatically obtain in real time the task content, which is clearly defined as "high voltage test of switch 1178 and phase A TYD of a certain 110kV line", and includes Yu Moumou as the person in charge of the work, Zhong Ming as the on-site permitter, and specific safety measures information, such as disconnecting switch 1178 and opening disconnectors 11781 and 11784, to ensure the detail and availability of the data.

[0082] Step 1.2, Automatic Personnel Qualification Verification: The automatic data acquisition module automatically calls the personnel qualification module interface in the power grid management platform to quickly and efficiently query the qualification validity of relevant personnel corresponding to each work order, including the specific type of qualification, validity period, and historical safety violation records. The system automatically identifies and filters abnormal qualification information to ensure the accuracy and usability of personnel qualification data.

[0083] Implementation Case: For example, regarding Yu, the person in charge of the work in this embodiment, the system automatically confirms that his "high voltage test supervisor qualification" is valid until December 31, 2026 and has no historical violation records; at the same time, the licensor Zhong's qualification is valid, but the system automatically reminds him that he had a "safety measure omission" deduction record in November 2024, which should be given special attention during the audit.

[0084] Step 1.3, Synchronous Integration of Maintenance Tasks and Defect Information: In addition, the system automatically calls the interfaces of the integrated power outage management module and defect management module in the power grid management platform to accurately obtain the maintenance content and status requirements of the equipment involved in the operation, as well as the latest defect information of the relevant equipment, and automatically forms a detailed equipment status database.

[0085] Implementation Case: In this embodiment, the system automatically detects that the 1178 switch on a certain 110kV line has an unresolved defect of "abnormal contact temperature rise", and its status is marked as "pending processing". At the same time, the maintenance task clearly states that the equipment needs to be in a de-energized state, and the system automatically marks the equipment as having potential risks that require special attention.

[0086] Step 1.4, Data Integration for Operational Equipment Risks: The automatic data acquisition module further calls the data interface of the power grid risk management module to automatically obtain the latest power grid risk data related to the equipment involved in the operation, especially the potential risks, risk levels and their specific descriptions of the equipment, in order to form an effective risk data source.

[0087] Implementation Case: In this case, the system automatically confirmed that there was a risk of a weak regional power grid when the 1178 switch of a certain 110kV line was de-energized. The risk level was defined as "high". The system automatically highlighted this risk in the data to facilitate subsequent risk warning and review work.

[0088] Step 1.5, Linked Collection of Environmental and Equipment Inspection Information: At the same time, the system automatically connects to the data interface of the substation operation support system to automatically acquire on-site and equipment inspection data in real time, including environmental meteorological data (such as wind speed, temperature, and weather conditions) as well as recent inspection photos and temperature detection data of equipment (inspection photos and temperature sensor data collected by drones, robots, or cameras, etc.), to ensure that on-site environmental data is included in the review and consideration in a timely and accurate manner.

[0089] Implementation Case: For example, the system automatically collected environmental data at 9:00 AM on the day of the operation, including a wind speed of 4.2 m / s, a temperature of 28.3℃, and clear weather. It also automatically collected the most recent inspection photos and temperature data of switch 1178 to build a complete environmental risk database.

[0090] Step 1.6, Data Standardization Cleaning and Structured Storage: The data cleaning and structured storage unit automatically receives the data streams from the above multiple channels, standardizes the raw collected data, automatically removes outliers and invalid data, unifies the data format, and finally stores it securely and stably in the structured database to ensure efficient data retrieval and utilization.

[0091] Implementation Case: The system automatically integrates the detailed work order information, personnel qualification information, equipment maintenance orders, defect records, risk levels and environmental information obtained above into a standardized database, ensuring the automation, integrity and real-time nature of data processing, and providing efficient and reliable data support for subsequent audit and analysis.

[0092] Step 2, Multi-source Data Semantic Fusion Analysis: Based on a pre-trained large language model, deep semantic analysis, intelligent fusion, and risk identification are performed on the various structured and unstructured data obtained in Step 1 to improve the data understanding capability and risk prediction accuracy of the job review. This step specifically includes: Step 2.1, Semantic Feature Extraction: The semantic feature extraction unit in the system of the present invention calls a large language model (such as a BERT model based on the Transformer structure) that has been fine-tuned based on the corpus in the field of substation operation and maintenance to perform semantic vectorization encoding on the textual descriptions of work tasks and safety measures in the work order, and extract high-level semantic features such as key equipment, operation behavior, work type, and execution conditions.

[0093] Implementation Case: Taking "1178 switch and A-phase TYD high voltage test on a 110kV line" as the task description, the large language model can accurately extract "1178 switch" as the core equipment, "high voltage test" as a high-risk operation, and further capture "A-phase TYD" to represent the precise operation location, providing a basic semantic vector for subsequent operation matching and risk assessment.

[0094] Step 2.2, Multi-source information fusion: This invention uses semantic vector matching and entity alignment algorithms to semantically fuse the operation semantic features extracted in Step 1 with heterogeneous data such as personnel qualifications, equipment defects, power outage maintenance orders, and risk records to form a unified operation semantic information graph.

[0095] Implementation Case: The system can automatically associate the "1178 switch" mentioned in the work task with the "abnormal temperature rise of 1178 switch contacts" in the equipment defect record. At the same time, it can bind the personnel "Yu Moumou" with their "high voltage test supervisor qualification" and build a semantic graph of "equipment-task-personnel-risk" to achieve deep data integration.

[0096] Step 2.3, Risk Feature Identification: Based on the fused semantic graph, the system calls the risk identification model to identify and mark potential violations, conflicts or omissions in the operation, including: personnel qualifications not matching, operation scope exceeding, equipment risks not closed-loop, safety measures inconsistent with tasks, etc. The identification results are accompanied by a credibility score and risk level label.

[0097] Implementation Case: In this embodiment, the system identified that the licensor, Zhong Ming, had a history of "omission of safety measures" and that the safety measures listed on the work order "did not specify the location of the grounding point on the 1178 line side". Based on this, the risk model gave a medium-to-high level risk label of "there may be omissions in the implementation of safety measures".

[0098] Step 3: Dynamic Review Rule Generation: By using reinforcement learning technology to intelligently train historical task data, a review rule base adaptable to different task scenarios is automatically generated, enabling continuous evolution and self-optimization of review strategies. This step specifically includes: Step 3.1, Reinforcement Learning Model Training: This invention uses deep reinforcement learning algorithms (such as DQN or Proximal Policy Optimization) to establish a learning model with the audit behavior as the action space and the risk identification result as the state space. The system uses past actual operation data as training samples to simulate the risk control effect brought about by different rule combinations in the audit process, and uses indicators such as reduced operation accident rate and improved audit consistency as reward functions to continuously iterate and optimize the audit strategy.

[0099] Implementation Case: Using 5,000 work records from a 220kV substation over the past 12 months as training data, the system simulates the impact of different rule sets (such as qualification verification priority, equipment defect closed-loop verification logic, etc.) on the audit results, and automatically optimizes the rule path that best balances audit efficiency and safety assurance.

[0100] Step 3.2, Dynamic Optimization of Rule Priority: The reinforcement learning model will evaluate the effectiveness of each review rule in actual operation in real time and dynamically adjust the execution weight of the rules. Inefficient rules will be downgraded or eliminated, while rules that perform well in practice and accurately identify high-risk situations will have their weight and priority increased.

[0101] Implementation examples: For instance, the rule "Prohibit high-pressure testing only for equipment with unclosed-loop oil leakage" has shown good accident prevention effects in multiple reviews, and the system automatically increases its priority in the rule base; while the rule "Non-keyword errors such as 'signage' in safety measures are considered low risk" has been manually intervened multiple times due to inaccurate judgments, and the system automatically lowers its trust weight.

[0102] Step 3.3, Continuous Maintenance and Iteration of the Rule Base: This invention establishes a rule base maintenance mechanism to periodically evaluate and iterate the review rules. New rules after each round of reinforcement learning training will be archived and stored with version numbers, while older versions are retained for easy backtracking and comparison. Simultaneously, it supports manual policy intervention and scenario-based configuration, achieving human-machine collaborative optimization.

[0103] Implementation Case: After adding the logic of "abnormal information of inspection equipment participating in the judgment of suspicious operation", the current version of the rule base v3.2 has identified 6 hidden risk operations in advance in a pilot application at a certain site. The system automatically records the validity of the new version of the rule and updates it to the main system.

[0104] Step 4, Collaborative Safety Risk Prediction: By constructing a collaborative risk prediction mechanism among substations, early detection and warning of potential safety hazards can be achieved. This step fully utilizes a federated learning architecture to achieve knowledge sharing and collective enhancement of risk identification capabilities while protecting data privacy. Specifically, it includes the following sub-steps: Step 4.1, Local Risk Prediction Model Construction: The system deploys a lightweight risk prediction model (such as LSTM or GRU model) locally in each substation. Based on the historical work tasks, personnel qualification fluctuations, equipment defect recurrence rate, and correlation data between meteorological conditions and accidents at each station, a risk assessment model with local characteristics is trained.

[0105] Implementation Case: For example, a certain station used high-temperature operation data from the past year to train a local model and found that there was a historical correlation of more than 80% of operations with wind speeds exceeding 6 m / s and involving high-altitude switch operations. Based on this, the system established a corresponding prediction model and output a visual risk score.

[0106] Step 4.2, Federated Learning Collaborative Modeling: Without transmitting the original data, the system uses federated learning technology to aggregate the model parameters (such as weights, gradients, etc.) trained locally at multiple substations to the central node, and merges them into a global risk prediction model under a secure computing mechanism, thereby realizing experience sharing among multiple sites and collaborative improvement of model accuracy.

[0107] Implementation Case: The system extracts and aggregates parameters from the local models collected by several sites and finds that the combination of "frequent changes in personnel qualifications + unclosed-loop defects + nighttime construction" has universal high-risk characteristics. It is eventually incorporated into the global risk prediction rules for general reference by all sites.

[0108] Step 4.3, Global Model Application and Dynamic Distribution: The global model optimized by federated learning will be automatically distributed to each site in a "daily update" or "major event trigger" manner to update the local prediction engine in real time, ensuring that each substation can obtain continuously evolving risk assessment capabilities without relying on external data.

[0109] Implementation Case: When a certain station plans to carry out a work order related to "unclosed-loop defective equipment + construction by external contractors", the local model, after applying the latest global model, automatically raises the risk level of the operation to "high" and sends an early warning reminder to the station manager and dispatching personnel 24 hours in advance.

[0110] Step 5, Automatic Push of Audit Results and Risk Alerts: An intelligent alert mechanism ensures accurate notification and timely feedback of audit results and potential risks, guaranteeing key personnel receive system identification results immediately, enabling them to organize on-site response and intervention. This includes the following sub-steps: Step 5.1, Automatic generation of alarm information: Based on the results of semantic analysis of the operation, the judgment conclusion of the review rules, and the results of risk collaborative prediction, the system automatically generates alarm information with standardized content and complete information, including the substation name, work order number, work task summary, responsible person information, triggered risk points and corresponding suggestions.

[0111] Implementation Case: For work permit 2025-03-C1-003 at a certain 110kV substation, the system automatically identified three major risk points: "Permitter has a violation record," "Equipment defects not mentioned in closed-loop management," and "High-temperature operation + abnormal wind speed," generating the following alarm information: [Warning] High-risk operations at a certain station. ●Ticket Number: 2025-03-C1-003 ●Task: High voltage test of phase A TYD switch 1178 ●Risk points: Unreviewed licensor's violation records, unresolved defects, abnormal weather. ●Recommendation: Exercise caution when granting licenses, provide supplementary information, and thoroughly verify the equipment status. Step 5.2, Multi-channel information push: The system connects to the Dify platform Elink interface on the intranet, SMS gateway, desktop pop-up or mobile office App, and pushes information according to different alarm levels and receiving roles to ensure that key personnel receive accurate alarms within the operable time window.

[0112] Implementation Case: The station manager received the official push notification via Elink; the patrol and maintenance shift leader received an SMS reminder; and the main station control backend interface simultaneously displayed a pop-up message: "High-risk operation at Jiangnan Station requires manual confirmation."

[0113] Step 5.3, Push Effect Feedback and Closed-Loop Recording: The system records the reception status and personnel response of each alarm push, and includes the feedback information in the closed-loop document of the work order to ensure that every risk alarm has traceable records and accountability.

[0114] Implementation Case: The system recorded that Mr. Yu confirmed receiving the alarm through Elink at 08:41 on March 10, 2025. The system automatically marked "Alarm Confirmation Completed" in the background. If there is no response within 30 minutes, the system will remind him again and escalate the alarm to the station management personnel.

[0115] Step 6, Visual Interface Interaction Configuration: This step utilizes a visual interface to configure the system, adjust rules, edit risk information, and monitor operational status. This lowers the barrier to entry for system use and improves its adaptability for wider adoption. Specifically, it includes the following sub-steps: Step 6.1, Data Interface and Acquisition Strategy Configuration: Users can select which production system data interfaces to enable in the interface, customize the acquisition frequency (such as every 15 minutes, hour, or day) and data retention period, and switch the interface call mode (real-time / scheduled / manual) based on the actual station requirements.

[0116] Implementation Case: In the initial stage of the launch of a certain station, the duty officer manually set it to collect only three types of data: "work tickets, personnel qualifications, and environmental information", and set the data retention period to 6 months; in the future, it can be expanded to a full interface collection mode according to the usage.

[0117] Step 6.2, Rule and Model Parameter Adjustment: The interface supports manual fine-tuning of the activation status, risk threshold, audit triggering conditions, and federated learning model update cycle of some audit rules; it provides slider-style settings for key strategies (such as "personnel violation tolerance value", "weather risk coefficient", etc.) and supports one-click rollback to historical versions.

[0118] Implementation Case: During the rainy season when operations are frequent, the system administrator lowered the "weather coefficient threshold" from the default 5.0m / s to 3.5m / s to improve the sensitivity of early warnings for thunderstorms and high winds; patrol and maintenance personnel can quickly switch to the "relaxed emergency operation review mode" to deal with emergencies.

[0119] Step 6.3, System Operation Status and Visual Monitoring of Audits: Users can view graphical indicators such as data collection status, rule execution frequency, number of risk identifications, and push response rate in real time; it supports exporting audit logs, generating monthly security reports, and statistical analysis of violation tendencies and distribution of high-frequency alarm devices.

[0120] Implementation Case: On the management backend page, the site administrator can view the number of work tickets approved this week, the number of high-risk alarms, the distribution of unresponsive alarms, and the early warning closure rate with one click, which facilitates overall control of the site's operational safety status.

[0121] Explanation of the beneficial effects of the technical solution of this invention: 1. Describe the solution in conjunction with the technical specifications, ensuring it is logical and well-supported by evidence; 2. It can be described in accordance with the technical problems to be solved in the aforementioned sections.

[0122] This invention constructs an automated verification system and method for substation operation scenarios, connecting key aspects such as operation data collection, semantic recognition, rule execution, risk prediction, information push, and configuration management, and possesses the following significant technical effects and application value: 1. Improve the efficiency and accuracy of work review and reduce the rate of human error in review. This invention automatically collects work orders, personnel qualifications, defect records, risk information, and on-site environmental data, and automatically integrates all the information required for the audit, replacing the traditional manual comparison method. This greatly improves the response speed and accuracy of the audit process and effectively avoids audit blind spots and safety hazards caused by insufficient experience or oversight.

[0123] 2. Achieve semantic-level task understanding and risk identification to enhance intelligent judgment capabilities. The system employs a large language model to perform deep semantic modeling on textual information such as work tasks and safety measures. It also establishes a semantic fusion graph by combining heterogeneous data such as equipment status, qualification requirements, and environmental weather. The system is capable of understanding complex work scenarios and can accurately identify hidden risks such as work type not matching qualifications, omission of safety measures, and failure to include unclosed defects.

[0124] 3. Construct a rule self-evolution mechanism to achieve dynamic optimization of the review strategy. This invention introduces reinforcement learning technology to train the audit strategy model through historical operation data, automatically generate rules and continuously iterate and optimize them, realizing the transformation of the audit strategy from a static list to a dynamic evolution. The system has adaptive capabilities and can continuously improve the accuracy of risk identification and the rationality of the audit strategy in actual operation.

[0125] 4. Break down system barriers to form a unified audit view and data loop. The system enables data fusion and semantic linkage between multiple power business systems, solving the problems of data fragmentation, closed interfaces, and difficulty in information flow between traditional systems. It truly realizes the system upgrade from "data silos" to "intelligent integration" and builds a unified audit and verification logical view and a traceable closed-loop chain of results.

[0126] 5. Establish a collaborative risk prediction system to achieve early warning and closed-loop response. By combining local models with federated learning technology, the system achieves risk prediction capabilities across sites and time scales. It can identify and warn of high-risk tasks before operations commence, and ensure key personnel receive decision support in a timely manner through an automatic information push module, realizing closed-loop control throughout the entire process of "early risk perception—real-time early warning—closed-loop response."

[0127] 6. Provides a flexible and configurable management interface to enhance the adaptability of promotion. The system supports visual configuration of parameters such as data collection strategies, audit rules, and alarm thresholds, which facilitates personalized deployment and strategy adjustment according to different site needs and operation and maintenance strategies, reduces the threshold for system use, and improves the breadth and sustainability of applications.

[0128] 7. Support the digital transformation of substation operation and maintenance, and empower the construction of inherently safe work teams. This invention integrates intelligent technologies such as AI, knowledge graphs, and reinforcement learning, transforming safety audits from "human control" to "intelligent control." This not only ensures the depth of policy implementation but also promotes the refinement and scientification of audit logic, providing strong digital support for building inherently safe work teams that are "risk-visible, audit-controllable, and operation-traceable."

[0129] In summary, this invention significantly improves the intelligence, automation, and closed-loop management of pre-operation safety audits in substations, effectively solving the problems of inefficiency, fragmentation, and reliance on manual labor in existing technologies, and has good practical application value and promotion prospects.

[0130] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. An automatic verification and validation system for substation operation safety, characterized in that, The system includes an automatic data acquisition module, a data semantic fusion and analysis module, a dynamic audit rule generation module, a risk collaborative prediction module, an automatic information push module, and a visual interactive configuration module. The modules are connected via an intranet data interface and an internal system communication protocol. The automatic data acquisition module is used to acquire various types of substation production data; The data semantic fusion analysis module extracts the data, and after semantic feature extraction and fusion analysis, identifies risk characteristics; The dynamic audit rule generation module calls upon the aforementioned risk features to perform reinforcement learning training and rule optimization. The risk collaborative prediction module generates a global risk prediction model based on the local prediction model and the federated learning collaborative mechanism, and feeds back the prediction results; The automatic information push module receives the output of the global risk prediction model, generates alarm information, and pushes it. The visual interactive configuration module provides an interactive interface for all modules, enabling parameter configuration and monitoring management.

2. The automatic verification system for substation operation safety according to claim 1, characterized in that, The automatic data acquisition module includes an interface calling unit, a data crawling unit, a data cleaning unit, and a structured storage unit. The interface calling unit is used to call the data interfaces of the power grid management platform and the substation operation support system to obtain various important production data involved in the substation operation process; The data crawling unit is used to automatically obtain various important production data involved in the substation operation process through intranet RPA technology or web crawlers when there is no interface data or interface calls are restricted. The data cleaning unit is used to perform data denoising, outlier removal, null value processing, and field standardization on the raw data of various important production processes involved in the substation operation, generating a unified standard data format. Structured storage units are used to classify and store cleaned data in a relational database, forming a standardized job safety audit database.

3. The automatic verification and validation system for substation operation safety according to claim 1, characterized in that, The data semantic fusion analysis module includes a semantic feature extraction unit, a data fusion unit, and a risk feature identification unit, wherein... The semantic feature extraction unit is fine-tuned based on the Transformer large language model to automatically extract key semantic features from text data such as work task descriptions, safety measure descriptions, personnel qualifications, and defect information. The data fusion unit intelligently matches and merges various important production data from different sources according to their semantic features, generating a unified semantic feature database and clarifying the relationship between related work tasks, personnel qualifications, equipment defects, and risk information. The risk feature identification unit automatically identifies potential safety risk features in the operation based on the fused semantic feature database and clearly marks the risk level of the data.

4. The automatic verification system for substation operation safety according to claim 1, characterized in that, The dynamic review rule generation module includes a reinforcement learning training unit, a rule optimization unit, and a rule base maintenance unit, wherein... The reinforcement learning training unit constructs a deep reinforcement learning environment based on historical audit data and risk identification results, trains an automated audit decision-making model, and dynamically generates optimal rules. The rule optimization unit automatically adjusts the rule priority and weight based on feedback; The rule base maintenance unit updates the optimized audit rules to the rule base in real time to ensure the continuous effectiveness and dynamic adaptability of the rule base.

5. The automatic verification and validation system for substation operation safety according to claim 1, characterized in that, The risk collaborative prediction module includes a local prediction unit, a federated learning collaborative unit, and a global model application unit, wherein... The local prediction unit trains a safety risk prediction model based on local audit and risk data to assess potential safety risks at the local substation. The federated learning collaborative unit uses a federated learning algorithm to safely aggregate the local model parameters of multiple substations and construct a global risk prediction model. The global model application unit applies the global model to each substation to provide guidance for risk prediction.

6. The automatic verification system for substation operation safety according to claim 1, characterized in that, The automatic information push module includes an alarm generation unit, a communication interface unit, and a message push unit, wherein... The alarm generation unit automatically generates standardized alarm information and audit reports based on the audit and prediction results; The communication interface unit is used to connect to the intranet Dify platform, elink data interface or mobile SMS interface to prepare the push channel. The message push unit is used to call the communication interface unit to quickly push alarm information to the corresponding operation and management personnel via elink or SMS.

7. The automatic verification method for substation operation safety based on the automatic verification and validation system for substation operation safety as described in claim 1, characterized in that, The method includes the following steps: Step 1: Obtain and integrate various important production data involved in substation operations in real time from the power grid production system and operation and maintenance support system; Step 2: Based on the pre-trained large language model, perform deep semantic analysis, intelligent fusion, and risk identification on the various important production data obtained in Step 1; Step 3: Intelligently train historical task data using reinforcement learning technology to automatically generate an audit rule base that can adapt to different task scenarios; Step 4: By constructing a collaborative risk prediction mechanism among substations, we can achieve early detection and warning of potential safety hazards. Step 5: Implement an intelligent alarm mechanism to accurately notify and promptly report on the results of work review and potential risks; Step 6: Configure the system, adjust rules, edit risk information, and monitor the operational status through a visual interface.

8. The automatic verification method for substation operation safety according to claim 7, characterized in that, Step 1 specifically includes: Step 1.1: The automatic data acquisition module, based on intranet data interface technology, automatically establishes a secure communication connection with the work order management module of the power grid management platform, and automatically obtains detailed information on all work orders currently being executed at all substations in real time, including the specific content of the work task, the work location, the start and end times of the work plan, the personnel information in the approval process, and detailed fields of the implemented safety measures. Step 1.2: The automatic data acquisition module automatically calls the personnel qualification module interface in the power grid management platform to quickly and efficiently query the qualification validity of the relevant personnel corresponding to each work order, including the specific type of qualification, validity period and historical safety violation records. The system automatically identifies and filters abnormal qualification information. Step 1.3: Call the interfaces of the integrated power outage management module and the defect management module in the power grid management platform to accurately obtain the maintenance content and status requirements of the equipment involved in the operation, as well as the latest defect information of the relevant equipment, and automatically form a detailed equipment status database; Step 1.4: The automatic data acquisition module calls the data interface of the power grid risk management module to automatically obtain the latest power grid risk data related to the equipment of the operation task, including the potential risks, risk levels and their specific descriptions for the equipment, so as to form an effective risk data source. Step 1.5: Connect to the data interface of the substation operation support system to automatically acquire on-site and equipment inspection data in real time, including environmental meteorological data, recent inspection photos of equipment, and temperature detection data; Step 1.6: The data cleaning and structured storage unit automatically receives data streams from multiple channels, standardizes the raw collected data, automatically removes outliers and invalid data, unifies the data format, and finally stores it in the structured database.

9. The automatic verification method for substation operation safety according to claim 7, characterized in that, Step 2 specifically includes: Step 2.1: The semantic feature extraction unit calls a large language model that has been fine-tuned based on the corpus in the field of substation operation and maintenance to perform semantic vectorization encoding on the textual descriptions of work tasks and safety measures in the work order, and extract high-level semantic features of key equipment, operation behavior, work type and execution conditions. Step 2.2: Using semantic vector matching and entity alignment algorithms, the semantic features of the operation extracted in Step 1 are semantically fused with heterogeneous data such as personnel qualifications, equipment defects, power outage maintenance orders, and risk records to form a unified semantic information graph of the operation. Step 2.3: Based on the fused semantic graph, the risk identification model is invoked to identify and mark potential violations, conflicts or omissions in the operation, including: personnel qualifications not matching, operation scope exceeding, equipment risks not closed-loop, and safety measures inconsistent with tasks. The identification results are accompanied by a credibility score and risk level label.

10. The automatic verification method for substation operation safety according to claim 7, characterized in that, Step 3 specifically includes: Step 3.1: Use deep reinforcement learning algorithm to establish a learning model with the audit behavior as the action space and the risk identification result as the state space. Use past actual operation data as training samples to simulate the risk control effect brought about by different rule combinations in the audit process. Use the indicators of reduced operation accident rate and improved audit consistency as reward function to iteratively optimize the audit strategy. Step 3.2: The reinforcement learning model will evaluate the effectiveness of each review rule in actual operation in real time and dynamically adjust the execution weight of the rule; Step 3.3: Establish a rule base maintenance mechanism to conduct periodic evaluations and version iterations of the review rules.

Images