Power grid operation decision management system based on artificial intelligence

Abstract

The invention discloses a power grid operation decision management system based on artificial intelligence, and the system comprises an application service layer which is used for receiving a power grid operation financial problem text transmitted by a user through a user side, and transmitting the power grid operation financial problem text to a semantic analysis module; the semantic analysis module is used for carrying out natural language understanding on the power grid operation financial problem text to obtain a standardized problem text; the knowledge retrieval module is used for performing vector similarity retrieval in the data resource layer based on the standardized problem text and recalling candidate fragments; the intelligent question and answer module is used for obtaining a dynamic recall threshold value, filtering the candidate segments according to the dynamic recall threshold value, generating decision answers corresponding to the power grid operation financial question text according to the reserved candidate segments through a pre-trained AI model, and returning the decision answers to the user side; the data resource layer is used for dynamically constructing a power grid exclusive knowledge base; and the model optimization layer is used for optimizing the AI model based on the feedback of the decision answer.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, in particular to an electric power grid operation decision management system based on artificial intelligence. BACKGROUND

[0002] With the rapid development of the power industry and the continuous expansion of the power grid scale, the operation of the power grid is facing increasingly complex decision-making challenges. The operation of the power grid not only involves power production, transmission, distribution and consumption, etc., but also is closely related to finance, market, policy, etc. In the process of power grid operation, financial decision-making is a crucial link, which is directly related to the economic benefits and sustainable development of power grid enterprises. However, the traditional financial decision-making management mode of power grid operation has many defects and cannot meet the needs of modern power grid operation.

[0003] Traditional financial decision-making mainly relies on the experience and professional knowledge of financial personnel, and lacks intelligent decision support tools. When facing complex financial problems, financial personnel often have difficulty in comprehensively and accurately analyzing various factors to make the best decisions. For example, when making investment decisions, factors such as project cost, benefit, risk, etc. need to be considered, and traditional decision-making methods are difficult to quantitatively analyze and comprehensively evaluate these factors, which can easily lead to decision-making errors. SUMMARY

[0004] Therefore, the embodiments of the present application provide an electric power grid operation decision management system based on artificial intelligence.

[0005] According to one aspect of the present application, an electric power grid operation decision management system based on artificial intelligence is provided, which comprises an operation decision platform, the operation decision platform being in communication connection with a data resource layer, a model optimization layer and an application service layer, the application service layer being in communication connection with a semantic analysis module, a knowledge retrieval module and an intelligent question and answer module;

[0006] The application service layer is configured to receive an electric power grid operation financial problem text sent by a user through a user terminal, and transmit the electric power grid operation financial problem text to the semantic analysis module.

[0007] The semantic analysis module is configured to perform natural language understanding on the electric power grid operation financial problem text to obtain a standardized problem text.

[0008] The knowledge retrieval module is configured to perform vector similarity retrieval in the data resource layer based on the standardized problem text, and recall candidate segments.

[0009] The intelligent question answering module is used to obtain a dynamic recall threshold, filter the candidate segments according to the dynamic recall threshold, filter out candidate segments whose topic relevance to the normalized question text is less than the dynamic recall threshold, generate the decision answer corresponding to the power grid operation financial question text through a pre-trained AI model based on the retained candidate segments, and return the decision answer to the user terminal.

[0010] The data resource layer is used to dynamically build a power grid-specific knowledge base, supporting the uploading of various power grid knowledge data, including financial statements, equipment maintenance records, market transaction data, policy documents, and equipment fault records, while also supporting dynamic updates and cleaning of internal data.

[0011] The model optimization layer is used to optimize the AI ​​model based on feedback from the decision answer.

[0012] By employing the above technical solutions, this application provides an AI-based power grid operation decision management system. With an operation decision platform at its core, it connects a data resource layer, a model optimization layer, and an application service layer. The application service layer receives user text containing power grid operation financial questions, converts it into standardized text via a semantic parsing module, retrieves candidate fragments based on vector similarity using a knowledge retrieval module, and generates and returns answers using a pre-trained AI model after filtering based on a dynamic recall threshold. The data resource layer dynamically constructs a knowledge base and supports data updates and cleaning. The model optimization layer optimizes the AI ​​model based on feedback. This application helps improve the efficiency and accuracy of power grid operation financial decision-making, enhances knowledge sharing and inheritance, and improves system adaptability and flexibility.

[0013] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0014] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0015] Figure 1 A schematic diagram of the structure of an artificial intelligence-based power grid operation decision management system provided in an embodiment of this application is shown;

[0016] Figure 2 This paper illustrates a schematic diagram of another AI-based power grid operation decision management system provided in an embodiment of this application. Detailed Implementation

[0017] Hereinafter, the present application will be described in detail with reference to the accompanying drawings and in conjunction with embodiments. It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0018] In the present embodiment, an artificial intelligence-based power grid operation decision management system is provided, as shown in Figure 1 The operation decision platform is communicatively connected with a data resource layer, a model optimization layer, and an application service layer. The application service layer is communicatively connected with a semantic analysis module, a knowledge retrieval module, and an intelligent question and answer module.

[0019] The application service layer is configured to receive a power grid operation financial problem text sent by a user through a user terminal, and transmit the power grid operation financial problem text to the semantic analysis module.

[0020] The semantic analysis module is configured to perform natural language understanding on the power grid operation financial problem text to obtain a standardized problem text.

[0021] The knowledge retrieval module is configured to perform vector similarity retrieval in the data resource layer based on the standardized problem text, and recall candidate segments.

[0022] The intelligent question and answer module is configured to obtain a dynamic recall threshold, filter the candidate segments according to the dynamic recall threshold to filter out candidate segments with a theme relevance less than the dynamic recall threshold from the standardized problem text, generate a decision answer corresponding to the power grid operation financial problem text according to the retained candidate segments through a pre-trained AI model, and return the decision answer to the user terminal.

[0023] The data resource layer is configured to dynamically construct a power grid exclusive knowledge base, support uploading various power grid knowledge data including financial statements, equipment maintenance records, market transaction data, policy documents, and equipment failure records, and support dynamic updating and cleaning of internal data.

[0024] The model optimization layer is configured to optimize the AI model based on feedback on the decision answer.

[0025] The power grid operation decision management system based on artificial intelligence in the embodiments of the present application mainly consists of an operation decision platform, a data resource layer, a model optimization layer, and an application service layer. The operation decision platform coordinates and manages the communication and interaction between the data resource layer, the model optimization layer, and the application service layer, ensuring that each module can operate efficiently and orderly. The data resource layer dynamically constructs a power grid dedicated knowledge base, which can support the uploading of various types of power grid knowledge data, including financial statements, equipment maintenance records, market transaction data, policy documents, and equipment failure records. These data cover all aspects of power grid operation, providing a rich knowledge base for the system. And support dynamic updating and cleaning of internal data, ensure that the data in the knowledge base always remain up-to-date and accurate, providing a reliable basis for subsequent decision-making. The model optimization layer optimizes the pre-trained AI model based on the feedback of the decision answers generated by the intelligent question and answer module. By continuously adjusting the model parameters, the accuracy and reliability of the model are improved, so that the system can better adapt to the changing power grid operation environment and user needs. The application service layer receives the power grid operation financial problem text sent by the user through the user terminal and passes it to the semantic analysis module. The semantic analysis module performs natural language understanding on the received power grid operation financial problem text and converts it into a standardized problem text, eliminating ambiguity and ambiguity in natural language, making subsequent knowledge retrieval and answer generation more accurate. The knowledge retrieval module performs vector similarity retrieval in the data resource layer based on the standardized problem text, recalling candidate segments related to the problem. Through vector similarity calculation, the most relevant knowledge content to the problem can be quickly and accurately found. The intelligent question and answer module obtains a dynamic recall threshold, filters the candidate segments based on the threshold, and filters out the candidate segments whose theme relevance to the standardized problem text is less than the dynamic recall threshold. Then, the pre-trained AI model is used to generate the decision answer corresponding to the power grid operation financial problem text according to the retained candidate segments, and the answer is returned to the user terminal. The setting of the dynamic recall threshold can flexibly adjust the screening standard of the candidate segments, improving the quality and relevance of the answer.

[0026] By applying the technical solutions of the embodiments, the operation decision platform is taken as the core, connecting the data resource layer, the model optimization layer, and the application service layer. The application service layer receives the user's power grid operation financial problem text, which is converted into a standardized text by the semantic analysis module. The knowledge retrieval module performs vector similarity retrieval based on it to recall candidate segments. The intelligent question and answer module generates an answer using a pre-trained AI model after filtering based on a dynamic recall threshold and returns it. The data resource layer dynamically constructs a knowledge base and supports data updating and cleaning. The model optimization layer optimizes the AI model based on feedback. The embodiments of the present application help to improve the efficiency and accuracy of power grid operation financial decision-making, enhance knowledge sharing and inheritance, and improve system adaptability and flexibility.

[0027] Optionally, the knowledge retrieval module is further configured to: convert the normalized question text into a question vector, retrieve all power grid knowledge data segment vectors in all AIGC knowledge bases in the data resource layer, calculate similarity scores between the question vector and each power grid knowledge data segment vector, and recall the top n candidate segments in descending order of the similarity scores, where n is a preset positive integer.

[0028] In this embodiment, the knowledge retrieval module converts the normalized question text obtained from the semantic analysis module into a vector form, i.e., a question vector. Subsequently, the knowledge retrieval module retrieves vectors corresponding to power grid knowledge data segments stored in all AIGC (artificial intelligence generated content) knowledge bases included in the data resource layer. Then, the knowledge retrieval module calculates similarity scores between the question vector and each power grid knowledge data segment vector by using a specific algorithm (e.g., cosine similarity). Finally, the knowledge retrieval module sorts these segments in descending order of the similarity scores and recalls the top n candidate segments in the sorting, where n is a preset positive integer representing an upper limit of the number of recalled candidate segments.

[0029] Optionally, the semantic analysis module is further configured to: perform question type identification and preset keyword identification on the power grid operation and finance question text, wherein the preset keywords include numerical condition keywords and confidential keywords; determine a baseline recall threshold value according to the question type corresponding to the power grid operation and finance question text; if only the numerical condition keywords are identified, determine a recall threshold value decrease value according to a mapping relationship between the number of identified numerical condition keywords and a preset numerical condition keyword number interval and the recall threshold value decrease value, and adjust the baseline recall threshold value according to the determined recall threshold value decrease value to obtain the dynamic recall threshold value; if only the confidential keywords are identified, determine a recall threshold value increase value according to a mapping relationship between the number of identified confidential keywords and a preset confidential keyword number interval and the recall threshold value increase value, and adjust the baseline recall threshold value according to the determined recall threshold value increase value to obtain the dynamic recall threshold value; if both the numerical condition keywords and the confidential keywords are identified, determine a recall threshold value decrease value according to a mapping relationship between the number of identified numerical condition keywords and a preset numerical condition keyword number interval and the recall threshold value decrease value, and determine a recall threshold value increase value according to a mapping relationship between the number of identified confidential keywords and a preset confidential keyword number interval and the recall threshold value increase value; determine a recall threshold value adjustment value according to the determined recall threshold value decrease value and the recall threshold value increase value, and adjust the baseline recall threshold value according to the determined recall threshold value adjustment value to obtain the dynamic recall threshold value; otherwise, determine the baseline recall threshold value as the dynamic recall threshold value.

[0030] In this embodiment, the semantic analysis module performs question type recognition and preset keyword recognition. Among them, the preset keywords are explicitly divided into numerical condition keywords (such as "amount", "cost", "yield" and other value-related words) and confidential keywords (such as "confidential", "top secret", "internal data" and other words related to sensitive information). According to the identified question type, the system determines a baseline recall threshold. Then, according to different keyword recognition situations, the recall threshold is dynamically adjusted: only numerical condition keywords are identified: according to the number of identified numerical condition keywords, combined with the mapping relationship between the preset numerical condition keyword quantity interval and the recall threshold drop value, the corresponding recall threshold drop value is determined, and the baseline recall threshold is adjusted by the value to obtain the dynamic recall threshold. For example, if the question contains multiple numerical condition keywords, the recall threshold may be appropriately lowered to recall more value-related segments. Only confidential keywords are identified: according to the number of identified confidential keywords, combined with the mapping relationship between the preset confidential keyword quantity interval and the recall threshold rise value, the corresponding recall threshold rise value is determined, and the baseline recall threshold is adjusted by the value to obtain the dynamic recall threshold. For example, if the question involves sensitive information, the recall threshold may be increased to ensure that the recalled segments are more accurate and meet security requirements. Both numerical condition keywords and confidential keywords are identified: according to the number of numerical condition keywords and confidential keywords, the corresponding recall threshold drop value and rise value are determined. Then, the two adjustment values are combined to determine the final recall threshold adjustment value, and the baseline recall threshold is adjusted by the value to obtain the dynamic recall threshold. No preset keywords are identified: the baseline recall threshold is directly used as the dynamic recall threshold. Thus, by dynamically adjusting the recall threshold, the flexibility and adaptability of the system are improved.

[0031] In the embodiments of the present application, as shown in Figure 2 The data resource layer includes a structured data processing unit, an unstructured text processing unit, and an external data processing unit; the AIGC knowledge base includes a relational database and a vector database; the structured data processing unit is configured to establish an independent index for table data in the power grid knowledge data, extract key fields, and construct a relational database; the unstructured text processing unit is configured to use a file slicing technology to divide text data in the power grid knowledge data into semantic paragraphs, convert the semantic paragraphs into vectors through a word embedding model, and store the vectors in a vector database; and the external data processing unit is configured to access meteorological warning information in real time through an API interface, obtain related knowledge in the knowledge base according to the meteorological warning information, and generate meteorological warning decision information according to the related knowledge and send the meteorological warning decision information to a management terminal.

[0032] In this embodiment, the technical data resource layer is composed of a structured data processing unit, an unstructured text processing unit and an external data processing unit, and the AIGC knowledge base includes a relational database and a vector database. The structured data processing unit establishes an independent index for table data in power grid knowledge data, extracts key fields, and constructs a relational database from these key fields, facilitating efficient storage and query of structured data. The unstructured text processing unit uses file slicing technology to segment text data in power grid knowledge data into semantic paragraphs, and then uses a word embedding model to convert these semantic paragraphs into vectors, which are stored in a vector database for subsequent vector similarity retrieval operations. The external data processing unit accesses meteorological warning information in real time through an API interface, finds related knowledge in the knowledge base according to the meteorological warning information, generates meteorological warning decision information according to the obtained related knowledge, and sends it to the management terminal, realizing timely response and decision support for external meteorological information. For example, the structured unit quickly locates the equipment ID with maintenance cost over 50%, the unstructured text processing unit extracts the "frequent insulation aging" semantic from the fault log, and the meteorological warning real-time data associates the future humidity > 90% warning, at which time the decision suggestion such as adding a moisture-proof coating is output. Thus, through the division of labor and cooperation of different units of the data resource layer, structured and unstructured power grid knowledge data and external meteorological warning information are efficiently processed, a perfect knowledge base system is constructed, precise storage, retrieval and utilization of various data are realized, comprehensive, timely and accurate information support for power grid operation decision-making is provided, and the scientificity and effectiveness of decision-making are improved.

[0033] In the embodiments of the present application, the data resource layer is optionally provided with a security control mechanism, which includes a permission grading mechanism, a field-level encryption mechanism and an operation audit mechanism. The permission grading mechanism is used to mark permission tags in the knowledge fragments in the AIGC knowledge base, and the permission tags include public mode tags and private mode tags. The knowledge fragments marked with the private mode tags are only visible to the creators and authorized persons, and the knowledge retrieval module retrieves in the knowledge fragments corresponding to the public mode tags and the private mode tags matching the user. The field-level encryption mechanism is used to encrypt sensitive fields in the knowledge fragments. The operation audit mechanism is used to record various operation behaviors on the AIGC knowledge base.

[0034] In this embodiment, the security management mechanism set by the data resource layer includes a permission grading mechanism, a field-level encryption mechanism, and an operation auditing mechanism. The permission grading mechanism: mark the knowledge fragments in the AIGC knowledge base with permission tags, divided into public mode tags and private mode tags. The knowledge fragments marked with private mode tags are only visible to the creator and authorized users. When the knowledge retrieval module retrieves, it only searches in the knowledge fragments corresponding to the public mode tags and the private mode tags matched with the user, so as to control the access rights of different users to the knowledge fragments. The field-level encryption mechanism: sensitive fields in the knowledge fragments are encrypted to prevent sensitive information from being illegally obtained or tampered with during storage and transmission, and to ensure the security of the data. The operation auditing mechanism: detailed records of various operation behaviors on the AIGC knowledge base, such as data access, modification, deletion, upload, etc., are recorded, so that in the case of security problems or data abnormalities, the operation process can be traced back to find the root cause, and at the same time, the user operation can be supervised and managed. Thus, the security and confidentiality of the power grid knowledge data in the AIGC knowledge base are ensured, and only authorized users can access the corresponding data to prevent sensitive information from being leaked. At the same time, the operation auditing function provides strong support for data security management and problem troubleshooting, and improves the security and reliability of the entire power grid operation decision management system.

[0035] In the embodiments of the present application, optionally, the model optimization layer includes a model training unit, configured to: obtain an initial large model, insert a lightweight adaptation matrix in the power grid field into the initial large model to obtain an initial AI model, train the initial AI model according to power grid field problem samples to optimize parameters of the lightweight adaptation matrix, and obtain the AI model. In the training process, the power grid field problem samples are used to generate decisions by the initial AI model, expert scores of the generated decisions are obtained, a reward model is trained using the expert scores to distinguish the quality of the decisions, and the quality evaluation information of the decisions is used to optimize the parameters of the lightweight adaptation matrix using the reward model.

[0036] In this embodiment, the model training unit in the model optimization layer conducts AI model construction and optimization. First, an initial large model is obtained, and a lightweight adaptation matrix in the power grid field is inserted on this basis to obtain an initial AI model. The role of this lightweight adaptation matrix is to enable the initial large model to better adapt to the specific needs of the power grid field. Then, the initial AI model is trained using power grid field problem samples. During the training process, the initial AI model generates decisions for the power grid field problem samples. Then, expert scores of these generated decisions are obtained, which represent the evaluation of decision quality. The reward model is trained using these expert scores, and the reward model can distinguish the quality of decisions. Finally, the reward model is used to optimize the parameters of the lightweight adaptation matrix. By continuously adjusting the parameters of the lightweight adaptation matrix, the initial AI model can generate higher quality decisions when processing power grid field problems, and finally an AI model suitable for power grid operation decision management is obtained. Through the insertion of the lightweight adaptation matrix in the power grid field in the initial large model, and the training and optimization using the power grid field problem samples combined with expert scores and the reward model, the AI model can better adapt to the characteristics and needs of the power grid field, and improve the decision quality and accuracy of the model in the power grid operation decision problem.

[0037] In the embodiments of the present application, the model optimization layer can also include a model compression unit, a platform deployment unit and an operation and maintenance management unit. The model compression unit is used to reduce the size of the trained AI model through model pruning and model quantization technology. The platform deployment unit is used to deploy the compressed AI model to the application service layer, and regularly optimize and update the deployment of the AI model based on feedback to the decision answers. The operation and maintenance management unit is used to monitor the performance of the AI model, and when the model performance decreases by more than a preset performance threshold, the model training unit is automatically triggered to retrain the model.

[0038] In this embodiment, in addition to the model training unit, the model optimization layer also includes a model compression unit, a platform deployment unit, and an operation and maintenance management unit. The model compression unit uses model pruning and model quantization techniques to compress the size of the trained AI model, removing redundant neurons or connections in the model. Model quantization converts model parameters from high precision to low precision, reducing the storage space and computing resources occupied by the model. The platform deployment unit deploys the compressed AI model to the application service layer. At the same time, based on the feedback of experts on the decision answers, the AI model is optimized regularly, and the updated model is redeployed to ensure that the model always maintains good performance and adaptability. The operation and maintenance management unit monitors the performance indicators of the AI model in real time, such as accuracy and response time. When the model performance drops more than a preset performance threshold (e.g., 10%), the model training unit is automatically triggered to redevelop model training to ensure that the model can be restored in time and improve performance to meet the needs of power grid operation decision management.

[0039] In the embodiments of the present application, the application service layer can further include a report generation module. The report generation module is configured to receive financial statement monitoring information sent by a user through a user terminal, acquire a monitored financial statement corresponding to the financial statement monitoring information in the data resource layer in real time, acquire target power grid knowledge data segments related to the monitored financial statement in the data resource layer when financial data in the monitored financial statement exceeds preset financial data, generate a preliminary draft of a financial analysis and risk assessment report based on the pre-trained AI model according to a power industry standard template, the monitored financial statement, the preset financial data, and the target power grid knowledge data segments, dynamically associate equipment failure records and maintenance cost tables in the target power grid knowledge data segments to generate a root cause analysis chart, supplement the preliminary draft of the financial analysis and risk assessment report according to the root cause analysis chart, and obtain the financial analysis and risk assessment report.

[0040] In this embodiment, a report generation module is also included in the application service layer. First, financial report monitoring information sent by a user through a user terminal is received. Then, a monitored financial report corresponding to the monitoring information is acquired in real time from the data resource layer. Then, it is determined whether the financial data in the monitored financial report exceeds preset financial data. If it does, target power grid knowledge data segments related to the monitored financial report are acquired from the data resource layer. Subsequently, based on a pre-trained AI model, in combination with a power industry standard template, the monitored financial report, the preset financial data, and the target power grid knowledge data segments, a preliminary draft of a financial analysis and risk assessment report is generated. At the same time, the device fault records and maintenance cost tables in the target power grid knowledge data segments are dynamically associated, and a root cause analysis chart is generated through this association. Finally, the preliminary draft of the financial analysis and risk assessment report is supplemented and improved according to the generated root cause analysis chart, thereby obtaining a final financial analysis and risk assessment report. In the embodiment of the application, the report generation module can quickly and accurately generate a comprehensive and targeted financial analysis and risk assessment report by acquiring and analyzing financial report data in real time, combining related power grid knowledge data, using an AI model and a power industry standard template to generate a preliminary report, and supplementing the report with a root cause analysis chart, thereby providing strong support for power grid operation decision-making and improving the scientificity and timeliness of decision-making.

[0041] In the embodiment of the application, the report generation module is also used to automatically trigger an approval process and push it to a responsible person terminal when the financial data in the monitored financial report exceeds the preset financial data.

[0042] In this embodiment, the report generation module actively performs specific operations when it detects that the financial data in the monitored financial report exceeds the preset financial data. It has the function of automatically triggering an approval process, i.e., starting relevant approval links according to preset approval rules and processes. At the same time, it pushes the approval matter containing the abnormal financial data and related information to the responsible person terminal, ensuring that the responsible person can receive the notification in time and know the abnormal condition of the financial report, so as to quickly take corresponding measures for processing. Real-time monitoring and timely response to abnormal conditions of the financial report are achieved.

[0043] In the embodiment of the application, the application service layer also includes an information extraction module. The information extraction module is used for named entity recognition to extract entity information of device ID, fault type, and maintenance amount from the device fault records and maintenance cost tables in the target power grid knowledge data segments and construct a device-fault-cost triple. The report generation module is also used to generate a root cause analysis chart based on the device-fault-cost triple.

[0044] In this embodiment, the information extraction module of the application service layer can complete the named entity recognition task. Specifically, for the device fault record and maintenance cost table in the target power grid knowledge data segment, the device ID, fault type, maintenance amount and other key entity information are extracted. Then, using these extracted entity information, a device-fault-cost triple is constructed to associate and integrate the device, fault and corresponding maintenance cost. The report generation module generates a root cause analysis chart using the device-fault-cost triple. By generating a chart through such associated information, the relationship between device fault and maintenance cost can be more intuitively displayed, and the complex data relationship can be presented in the form of an intuitive chart, which helps users quickly understand the internal relationship between device fault and cost.

[0045] By applying the technical solutions of the present embodiment, the following beneficial effects can be achieved after the application and testing of the solutions:

[0046] (1) The present application realizes accurate decision-making through intelligent question answering and knowledge retrieval module. After the user asks a question in natural language, the system retrieves policy clauses and associated data from the power exclusive knowledge base based on vector similarity retrieval and dynamic threshold filtering, with a question answering accuracy of >92%. This significantly improves decision-making efficiency and reduces manual search time by more than 80%.

[0047] (2) The present application drives deep analysis through the data resource layer and the report generation module. It dynamically integrates financial statements, equipment maintenance records and real-time market data, uses file slicing technology and incremental cleaning to ensure data quality. The report generation module automatically associates equipment fault and cost data to output root cause analysis reports. The report generation time is reduced from 72 hours to 2 hours, helping power grid enterprises quickly identify operational risks.

[0048] (3) The present application realizes efficient and reliable operation and maintenance through the model optimization layer and the security mechanism. It uses low-rank adaptive fine-tuning technology to inject power grid professional knowledge, reducing training power consumption by 90%. It also uses model compression and automatic retraining mechanisms to ensure that the inference speed is improved and the performance is stable, with a decline of no more than 10%. In addition, through field-level encryption and private knowledge base, sensitive data such as device ID and maintenance amount are encrypted throughout the process, meeting the requirements of the "Electric Power Data Security Specification".

[0049] The technical features of the above embodiments can be combined in any way. To make the description concise, not all possible combinations of the technical features in the above embodiments are described, but as long as the combinations of the technical features do not exist, they should be considered as the scope of the present disclosure.

[0050] The above-described embodiments are merely illustrative of several embodiments of the present application, which are described in more detail and in a specific manner, but should not be construed as limiting the scope of the patent of the present application. It should be noted that, for those of ordinary skill in the art, several modifications and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the appended claims.

Claims

1. A power grid operation decision management system based on artificial intelligence, characterized in that, This includes an operational decision-making platform, which is communicatively connected to a data resource layer, a model optimization layer, and an application service layer. The application service layer is communicatively connected to a semantic parsing module, a knowledge retrieval module, and an intelligent question answering module. The application service layer is used to receive text messages about power grid operation and financial issues sent by users through user terminals, and to transmit the text messages about power grid operation and financial issues to the semantic parsing module. The semantic parsing module is used to perform natural language understanding on the text of the power grid operation financial issues to obtain a standardized issue text; The knowledge retrieval module is used to perform vector similarity retrieval in the data resource layer based on the standardized question text and recall candidate fragments; The intelligent question answering module is used to obtain a dynamic recall threshold, filter the candidate segments according to the dynamic recall threshold, filter out candidate segments whose topic relevance to the normalized question text is less than the dynamic recall threshold, generate the decision answer corresponding to the power grid operation financial question text through a pre-trained AI model based on the retained candidate segments, and return the decision answer to the user terminal. The data resource layer is used to dynamically build a power grid-specific knowledge base, supporting the uploading of various power grid knowledge data, including financial statements, equipment maintenance records, market transaction data, policy documents, and equipment fault records, while also supporting dynamic updates and cleaning of internal data. The model optimization layer is used to optimize the AI ​​model based on feedback from the decision answer.

2. The power grid operation decision management system based on artificial intelligence according to claim 1, characterized in that, The knowledge retrieval module is also used for: The standardized question text is converted into a question vector. All power grid knowledge data fragment vectors in the AIGC knowledge base are retrieved in the data resource layer. The similarity score between the question vector and each power grid knowledge data fragment vector is calculated. The first n candidate fragments are recalled after sorting them from high to low according to the similarity score, where n is a preset positive integer.

3. The power grid operation decision management system based on artificial intelligence according to claim 1, characterized in that, The semantic parsing module is also used for: The text of the power grid operation financial issues is subjected to issue type identification and preset keyword identification, wherein the preset keywords include numerical condition keywords and classified keywords; and a corresponding baseline recall threshold is determined according to the issue type corresponding to the text of the power grid operation financial issues. If only the numerical condition keywords are identified, then the recall threshold decrease value is determined based on the number of identified numerical condition keywords and the mapping relationship between the preset numerical condition keyword number range and the recall threshold decrease value. The baseline recall threshold is then adjusted based on the determined recall threshold decrease value to obtain the dynamic recall threshold. If only the classified keywords are identified, the recall threshold increase value is determined based on the number of identified classified keywords and the mapping relationship between the preset range of classified keyword numbers and the recall threshold increase value. The baseline recall threshold is then adjusted based on the determined recall threshold increase value to obtain the dynamic recall threshold. If both the numerical condition keywords and the classified keywords are identified simultaneously, then based on the number of identified numerical condition keywords and the mapping relationship between the number range of the number of numerical condition keywords and the recall threshold decrease value, a recall threshold decrease value is determined; and based on the number of identified classified keywords and the mapping relationship between the number range of the number of classified keywords and the recall threshold increase value, a recall threshold increase value is determined; based on the determined recall threshold decrease value and recall threshold increase value, a recall threshold adjustment value is determined, and based on the determined recall threshold adjustment value, the baseline recall threshold is adjusted to obtain the dynamic recall threshold; Otherwise, the baseline recall threshold will be determined as the dynamic recall threshold.

4. The power grid operation decision management system based on artificial intelligence according to claim 2, characterized in that, The data resource layer includes a structured data processing unit, an unstructured text processing unit, and an external data processing unit; the AIGC knowledge base includes a relational database and a vector database. The structured data processing unit is used to establish independent indexes on the tabular data in the power grid knowledge data and extract key fields to construct a relational database. The unstructured text processing unit is used to segment the text data in the power grid knowledge data into semantic paragraphs using file slicing technology, convert them into vectors through a word embedding model, and store them in a vector database. The external data processing unit is used to access meteorological early warning information in real time through an API interface, obtain relevant knowledge from the knowledge base based on the meteorological early warning information, and generate meteorological early warning decision information based on the relevant knowledge and send it to the management terminal.

5. The power grid operation decision management system based on artificial intelligence according to claim 4, characterized in that, The data resource layer is equipped with a security control mechanism, which includes: a permission hierarchical mechanism, a field-level encryption mechanism, and an operation audit mechanism. The permission leveling mechanism is used to mark permission tags on knowledge fragments in the AIGC knowledge base. The permission tags include public mode tags and private mode tags. The knowledge fragments marked with the private mode tags are only visible to the creator and the authorized user. The knowledge retrieval module searches for knowledge fragments corresponding to public mode tags and private mode tags that match the user. The field-level encryption mechanism is used to encrypt sensitive fields in knowledge fragments; The operation auditing mechanism is used to record various operational behaviors on the AIGC knowledge base.

6. The power grid operation decision management system based on artificial intelligence according to claim 1, characterized in that, The model optimization layer includes a model training unit, used for: An initial large model is obtained, and a lightweight adaptation matrix for the power grid domain is inserted into the initial large model to obtain an initial AI model. The initial AI model is trained based on power grid domain problem samples to optimize the parameters of the lightweight adaptation matrix, thus obtaining the AI ​​model. During the training process, the initial AI model is used to generate decisions for the power grid domain problem samples, and expert scores are obtained for the generated decisions. A reward model is trained using the expert scores to distinguish between high and low decision quality, and the parameters of the lightweight adaptation matrix are optimized using the decision quality evaluation information from the reward model.

7. The power grid operation decision management system based on artificial intelligence according to claim 6, characterized in that, The model optimization layer also includes a model compression unit, a platform deployment unit, and an operation and maintenance management unit; The model compression unit is used to reduce the size of the trained AI model through model pruning and model quantization techniques. The platform deployment unit is used to deploy the compressed AI model to the application service layer, and periodically optimize and update the AI ​​model based on feedback on the decision answer. The operation and maintenance management unit is used to monitor the performance of the AI ​​model, and when the model performance drops below a preset performance threshold, it automatically triggers the model training unit to retrain the model.

8. The power grid operation decision management system based on artificial intelligence according to claim 7, characterized in that, The application service layer also includes a report generation module; The report generation module is used to receive financial statement monitoring information sent by users through the user terminal, and to obtain the monitored financial statements corresponding to the monitoring information in real time in the data resource layer. When the financial data in the monitored financial statements exceeds the preset financial data, the module obtains the target power grid knowledge data fragments related to the monitored financial statements in the data resource layer. Based on a pre-trained AI model, the module generates a draft financial analysis and risk assessment report according to the power industry standard template, the monitored financial statements, the preset financial data, and the target power grid knowledge data fragments. At the same time, the module dynamically associates the equipment fault records and maintenance cost tables in the target power grid knowledge data fragments to generate root cause analysis charts. The module supplements the draft financial analysis and risk assessment report based on the root cause analysis charts to obtain the final financial analysis and risk assessment report.

9. The power grid operation decision management system based on artificial intelligence according to claim 8, characterized in that, The report generation module is also used to automatically trigger an approval process and push the data to the responsible person's terminal when the financial data in the monitored financial statements exceeds the preset financial data.

10. The power grid operation decision management system based on artificial intelligence according to claim 8, characterized in that, The application service layer also includes an information extraction module; The information extraction module is used for named entity recognition, extracting entity information such as equipment ID, fault type, and maintenance amount from equipment fault records and maintenance cost tables in the target power grid knowledge data fragment, and constructing equipment-fault-cost triplets; The report generation module is also used to generate root cause analysis charts based on the equipment-failure-cost triplet.

Images