Method for automatically generating orderly power utilization scheme based on bright large model

Abstract

The invention relates to the technical field of power load management, and discloses an orderly power utilization scheme automatic generation method based on a bright large model, and the method comprises the steps: 1, integrating user archive data and historical load data, collecting policy and regulation files and expert scheduling experience, forming a policy interpretation library and an expert experience library through semantic analysis, and carrying out the semantic analysis of the policy interpretation library and the expert experience library; finally, all knowledge base contents are processed into a vector form and stored in a vector database, and construction of a professional knowledge system is completed. 2, based on the constructed knowledge system, the system receives a natural language text instruction input by a user; according to the technical scheme, a full-process automatic scheme is adopted, the technical effects that scheme compilation is changed from a manual intensive type to a machine driving type, and minute-level response and generation are achieved are achieved, and compared with the technical scheme in the prior art that business personnel depends on manual integration and analysis of massive user archives and load data, the technical scheme has the advantages that the efficiency is high; the defects of complex process, long time consumption, easiness in making mistakes and slow response are overcome.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power load management, in particular to a method for automatically generating an orderly power utilization scheme based on a Guangming big model. BACKGROUND

[0002] Orderly power utilization is a key measure to ensure the safe and stable operation of the power grid and promote the optimal allocation of power resources. In particular, during peak power consumption periods such as the summer peak and winter peak, it is crucial to scientifically and accurately prepare and implement orderly power utilization schemes.

[0003] Currently, the preparation process of orderly power utilization schemes largely relies on the professional experience of power business personnel. The conventional approach is for business personnel to first manually integrate massive user profile data and historical load data from business systems such as marketing systems and load collection systems, and to consult and understand relevant policy and regulatory documents. Then, based on the obtained information and past dispatching experience, the users are manually analyzed and selected, and load reduction indicators are formulated.

[0004] However, the existing technology has the following shortcomings in the implementation process:

[0005] In the preparation of orderly power utilization schemes, the existing technology relies on business personnel to manually integrate and analyze massive user profile data and load data, resulting in complex processes, long time consumption, and susceptibility to errors. It is not easy to quickly respond to the dynamic changes in power supply and demand, which greatly limits the efficiency of power load management.

[0006] In the process of formulating orderly power utilization schemes, the existing technology is usually based on past experience and fixed rules, lacking the ability to deeply mine and correlate multi-dimensional power data. This results in low decision-making intelligence, making it difficult to predict load trends and understand user response potential, thereby affecting the scientificity and optimality of the scheme.

[0007] The existing technology can integrate heterogeneous data from marketing profiles and load collection, but it fails to achieve deep fusion and correlation analysis of data, and cannot extract high-value information to support load management decisions. This leads to the problem of not fully releasing the value of data.

[0008] The existing load management system is complex to operate, and the human-computer interaction experience is not good. Business personnel need to learn the system usage method through special learning, and it is not easy to obtain the data analysis results or generate schemes required for decision-making through intuitive and convenient methods, thereby reducing work efficiency and hindering the popularization and application of the system.

[0009] Therefore, the present application proposes a method for automatically generating an orderly power utilization scheme based on a Guangming big model to solve the above-mentioned problems. SUMMARY

[0010] In view of the deficiencies of the prior art, the application provides an ordered power utilization scheme automatic generation method based on a Guangming large model to solve the problems in the above background art.

[0011] To achieve the above object, the application is implemented by the following technical solutions: an ordered power utilization scheme automatic generation method based on a Guangming large model, comprising:

[0012] Step one: first integrate user profile data and historical load data, simultaneously collect policy and regulation files and expert dispatching experience, form a policy interpretation library and an expert experience library through semantic analysis, and finally process all knowledge base contents into vector form and store them in a vector database to complete the construction of a professional knowledge system;

[0013] Step two: based on the constructed knowledge system, the system receives a natural language text instruction or a graphical drag-and-drop operation instruction input by a user, calls the semantic understanding capability of the Guangming power large model, identifies the task intention and constraint conditions in the instruction, and forms a structured task description;

[0014] Step three: according to the generated structured task description, the system processes the task description into a task vector, calculates the semantic similarity of the task vector and a knowledge item vector in the constructed vector database, and selects knowledge with a similarity higher than a preset threshold to form a decision context;

[0015] Step four: taking the structured task description and the knowledge context as inputs, the system calls the Guangming power large model for reasoning, and finally generates an ordered power utilization scheme by the large model;

[0016] Step five: based on the ordered power utilization scheme, the system automatically calls an audit agent to verify the compliance of the scheme according to the policy interpretation library, and then calls an index intelligent decomposition agent to decompose the overall load regulation target of the scheme into an “individual strategy” execution index for the user;

[0017] In the step five, further comprising:

[0018] Sub-step , based on the generated structured ordered power utilization scheme , the system automatically calls an audit agent, the agent extracts all compliance constraint clauses from the policy interpretation library, and compares them with the fields in the ordered power utilization scheme one by one;

[0019] Set a Boolean compliance flag , the initial value is true, when all fields in the ordered power utilization scheme meet the corresponding constraint clauses, the compliance flag remains true, and the system enters the next sub-step;

[0020] Sub-step If compliance flag bit If true, the system is targeting the orderly power consumption plan. Each user in the user list The load characteristic analysis agent is invoked to obtain historical load data sequences for 96 points. The user's historical peak load contribution is calculated using the following formula. Compared with the load response potential assessment value :

[0021] ,

[0022] ,

[0023] in, This is a set of historical peak load periods for the system. Total number of users The user's historical maximum load, The user's historical average load;

[0024] Sub-step Based on the calculated contribution of historical peak load Compared with the load response potential assessment value The system calls the indicator intelligent decomposition intelligent agent, first calculating the intelligent agents of each user. Allocation weights ,That, and The preset weighting coefficients satisfy the following conditions: ;

[0025] The orderly power consumption plan will be implemented later. Overall load control target The weighted allocation algorithm is used to calculate the allocation to each user. Load reduction:

[0026] Ultimately, a list of implementation indicators for each household will be formed.

[0027] Preferably, the step of calling the Guangming Power large model specifically includes:

[0028] The prompt word engineering module combines the structured task description with the knowledge context to form enhanced prompt words; these enhanced prompt words are then input into the Guangming Power big data model to drive the generation of the orderly electricity consumption scheme.

[0029] Preferably, the step of the auditing agent verifying the scheme specifically includes:

[0030] The load reduction values, execution periods, and user industry category parameters in the plan are automatically compared with the constraint clauses stored in the policy interpretation database; when a parameter exceeds the range of the constraint clauses, the plan is marked as non-compliant.

[0031] Preferably, the intelligent decomposition agent of the index adopts a preset weighted allocation algorithm to calculate the load reduction allocated to the user based on the user's historical peak load contribution and load response potential assessment value.

[0032] Preferably, the historical peak load contribution and load response potential assessment value are obtained by the load characteristic analysis agent after performing time-series analysis on the user's historical 96-point load data.

[0033] Preferably, step one further includes:

[0034] Sub-step First, user profile data and historical user workload data are integrated. Simultaneously, policy and regulatory documents and expert scheduling experience are collected to form a raw text corpus. Then, data cleaning is performed based on the historical workload data, and if any time points exist... load value Missing information is filled in using a linear interpolation algorithm. The calculation formula is as follows:

[0035] ,

[0036] in, Time to be completed The load value, and The closest moments before and after the missing point and The known load value;

[0037] Sub-step Based on the acquired original text corpus, semantic parsing is performed using named entity recognition and relation extraction algorithms. When a text fragment simultaneously satisfies the condition that the main entity is a user category, the relation word is a scheduling behavior, and the object entity is a constraint value, the text fragment is extracted into a structured knowledge item and stored in the policy interpretation database and expert experience database according to the text source.

[0038] Sub-step For the generated structured knowledge entries, a pre-trained text embedding model is invoked. Knowledge entries in various text formats Convert to a fixed-dimensional floating-point vector The conversion process is represented as follows:

[0039] ,

[0040] wherein, is a knowledge item corresponding vector representation, is a text embedding model;

[0041] for the generated floating-point number vector perform validity check, when norm is greater than a preset non-zero threshold , the floating-point number vector is stored in the vector database, and finally a complete vectorized domain knowledge base is formed.

[0042] Preferably, in step two, further comprising:

[0043] sub-step , based on the constructed vectorized domain knowledge base, the system receives user input natural language text instructions or graphical drag operation instructions, and converts the received original instructions into a unified text string through a preset normalization function , the process is represented as:

[0044] ,

[0045] when the character length of the text string is greater than a preset minimum instruction length threshold , the subsequent processing flow is triggered.

[0046] sub-step , the system inputs the generated text string to the semantic understanding module of the Guangming Electric Power large model, and the module performs intent recognition and entity extraction in parallel; the intent recognition algorithm outputs candidate intent and the corresponding intent confidence , and the entity extraction algorithm extracts an entity set composed of entity type and entity value from the text:

[0047] ,

[0048] when the intent confidence satisfies the condition , the candidate intent is confirmed as the final task intent, wherein is a preset intent confirmation confidence threshold;

[0049] sub-step , according to the confirmed final task intent and the extracted entity set The system is for the task intent Match predefined task templates The template contains a set of required entity slots; the system uses the entity set... entity values ​​in Fill into task template The corresponding entity type To generate structured task descriptions ;

[0050] Before generation, the system performs an integrity check to ensure the task template is valid. All necessary entity slots can be found in the entity collection. Find the corresponding fill value, and after successful verification, finally generate the structured task description. .

[0051] Preferably, step three further includes:

[0052] Sub-step Based on the generated structured task description The system first serializes the data into a single text string. Then the same pre-trained text embedding model is called. , text string Convert to task vector The conversion process is represented as follows:

[0053] ,

[0054] After the transformation is complete, the task vector Perform validity verification when norm When the value is greater than zero, confirm the task vector. efficient;

[0055] Sub-step Use the confirmed valid task vector The system traverses the vectors of each knowledge entry in the vector database. Furthermore, the cosine similarity algorithm is used to calculate the task vector. With each knowledge item vector semantic similarity score between The calculation formula is:

[0056] ,

[0057] in, The dot product of vectors, Task vector of norm, is a knowledge entry vector of norm;

[0058] sub-step , the system selects the original knowledge entry text corresponding to the knowledge entry vector based on the calculated semantic similarity score ; when the corresponding semantic similarity score of the knowledge entry vector satisfies the condition , the system selects the original knowledge entry text corresponding to the knowledge entry vector , and combines all the selected knowledge entry texts to form a decision context for subsequent steps .

[0059] Preferably, in step four, further comprising:

[0060] sub-step , based on the formed decision context , the system calls the prompt word engineering module to inject the generated structured task description and the decision context into a pre-set prompt word template to generate an enhanced prompt word ; the prompt word engineering module performs integrity verification after generation, and confirms the enhanced prompt word valid when the task description and decision context contained in the enhanced prompt word are both non-empty.

[0061] Preferably, in step four, further comprising:

[0062] sub-step , taking the confirmed valid enhanced prompt word as input, calling the Guangming Power large model to perform inference generation operation, the process can be represented as:

[0063] ,

[0064] wherein, is the original text string output by the large model containing the ordered power consumption scheme content; the system performs validity determination on the original text string ; when the original text string does not contain a pre-set error identifier and the text length is greater than a minimum length threshold, it enters subsequent processing;

[0065] sub-step , based on the generated valid original text string , system call predefined scheme analysis function , the original text string Extract the relevant fields, and fill the extracted fields into the standard scheme data structure to finally form a structured ordered power consumption scheme ;

[0066] After forming, the system performs a final integrity check on the ordered power consumption scheme Ensure that all required key fields have been successfully filled, and output the scheme after passing the check.

[0067] The present application provides an ordered power consumption scheme automatic generation method based on a Guangming big model. It has the following beneficial effects:

[0068] 1、The present application adopts a full-process automatic scheme, which achieves the transformation of scheme compilation from labor-intensive to machine-driven, realizes the technical effect of minute-level response and generation, and solves the problems of complex process, long time consumption, easy to make mistakes and slow response compared with the technical scheme in the prior art which relies on business personnel to manually integrate and analyze massive user archives and load data.

[0069] 2、The present application adopts an intelligent decision-making scheme, which achieves the transformation of the scheme formulation process from post-experience dependence to pre-data insight, significantly improves the scientificity and optimality of the scheme, and solves the problem of low intelligent level of decision-making due to lack of deep mining ability of multi-dimensional power data compared with the technical scheme in the prior art which is based on past experience and fixed rules.

[0070] 3、The present application adopts a data deep fusion scheme, which achieves the technical effect of transforming heterogeneous and dispersed data sources into a high-value knowledge system that can be understood and called by a big model, solves the problem of failing to realize data deep fusion and correlation analysis, and leads to the problem of failing to fully release the value of data compared with the technical scheme in the prior art which can only perform surface integration of marketing archives and load collection data.

[0071] 4、The present application adopts a natural language and graphical interaction scheme, which achieves the technical effect that users can complete the compilation of complex schemes through intuitive methods such as dialogue or dragging, solves the problem of poor human-computer interaction experience and high operation threshold, and thus reduces work efficiency and hinders the popularization and application of the system compared with the complex operation system which needs to be learned and mastered in the prior art. BRIEF DESCRIPTION OF DRAWINGS

[0072] Figure 1 is a flowchart of the present application;

[0073] Figure 2 is a design scheme block diagram of the present application;

[0074] Figure 3 Software architecture diagram for example two. DETAILED DESCRIPTION

[0075] In order for those skilled in the art to understand the technical solutions of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, other embodiments obtained by those skilled in the art without creative labor should be within the scope of protection of the present application.

[0076] The present application will be described in detail below with reference to the drawings:

[0077] Example one:

[0078] Please refer to the accompanying Figure 1 and the accompanying Figure 2 The embodiments of the present application provide a method for automatically generating an orderly power consumption scheme based on a Guangming model, which comprises:

[0079] Step one: first integrate user profile data and historical load data, collect policy and regulation files and expert dispatching experience at the same time, form a policy interpretation library and an expert experience library through semantic analysis, finally process all the knowledge base contents into vector form and store them in a vector database to complete the construction of a professional knowledge system;

[0080] Step two: based on the constructed knowledge system, the system receives natural language text instructions or graphical drag-and-drop operation instructions input by the user, calls the semantic understanding capability of the Guangming power model, identifies the task intent and constraint conditions in the instructions, and forms a structured task description;

[0081] Step three: according to the generated structured task description, the system processes the task description into a task vector, calculates the semantic similarity between the task vector and the knowledge entry vector in the constructed vector database, and selects the knowledge with a similarity higher than a preset threshold to form a decision context;

[0082] Step four: taking the structured task description and the knowledge context as inputs together, calling the Guangming power model for reasoning, and finally generating an orderly power consumption scheme by the model;

[0083] Step five: based on the orderly power consumption scheme, the system automatically calls an audit agent to verify the compliance of the scheme according to the policy interpretation library, and then calls an index intelligent decomposition agent to decompose the overall load control target of the scheme into "one household one strategy" execution indexes for users;

[0084] In step one, further comprising:

[0085] Sub-step , first integrate user profile data and user historical load data, collect policy documents and expert dispatching experience to form an original text corpus, and based on historical load data, if there is a missing load value at a time point , a linear interpolation algorithm is used to complete the missing load value, the calculation formula is:

[0086]

[0087] wherein, is the load value to be completed at the time point , and and are the known load values at the nearest time points before and after the missing point and

[0088] Substep , based on the obtained original text corpus, semantic analysis is performed using named entity recognition and relationship extraction algorithm, when the text segment meets the conditions of user category as the main entity, dispatching behavior as the relationship word, and constraint value as the object entity, the text segment is extracted as a structured knowledge item, and according to the text source, it is stored in the policy interpretation library and the expert experience library;

[0089] Substep , for the generated structured knowledge item, a pre-trained text embedding model is called to convert each text-form knowledge item into a fixed-dimension floating-point number vector , and the conversion process is represented as:

[0090]

[0091] wherein, is the vector representation of the knowledge item , and is the text embedding model;

[0092] For the generated floating-point number vector , validity check is performed, when the norm of the floating-point number vector is greater than a preset non-zero threshold , the floating-point number vector is stored in the vector database, and finally a complete vectorized domain knowledge base is formed.

[0093] In step two, further comprising:

[0094] Substep ​​​​, based on the constructed vectorized domain knowledge base, the system receives a user input natural language text instruction or a graphical drag-and-drop operation instruction, and converts the received original instruction into a unified text string through a preset normalization function . , the process is represented as:

[0095] ,

[0096] When the character length of the text string is greater than the preset minimum instruction length threshold , the subsequent processing flow is triggered.

[0097] Sub-step , the system inputs the generated text string into the semantic understanding module of the Guangming Electric Power large model, and the module performs intent recognition and entity extraction in parallel; the intent recognition algorithm outputs candidate intent and the corresponding intent confidence , and the entity extraction algorithm extracts an entity set composed of entity types and entity values from the text:

[0098] ,

[0099] When the intent confidence satisfies the condition , the candidate intent is confirmed as the final task intent, wherein is a preset intent confirmation confidence threshold;

[0100] Sub-step , according to the confirmed final task intent and the extracted entity set , the system matches the task intent with a predefined task template , which contains a set of required entity slots; the system fills the entity values in the entity set into the corresponding entity types in the task template to generate a structured task description .

[0101] Before generation, the system performs integrity verification to ensure that all required entity slots in the task template can be filled in the entity set The corresponding fill value is found in the database, and after the check passes, the structured task description is finally generated .

[0102] Step three further comprises:

[0103] Sub-step , based on the generated structured task description , the system first serializes into a single text string , then calls the same pre-trained text embedding model , converts the text string into a task vector , and the conversion process is represented as:

[0104] ,

[0105] After the conversion is completed, the task vector is checked for validity, and when the norm is greater than zero, the task vector is confirmed to be valid;

[0106] Sub-step , using the confirmed valid task vector , the system traverses each knowledge entry vector in the vector database , and uses the cosine similarity algorithm to calculate the semantic similarity score between the task vector and each knowledge entry vector , the calculation formula is:

[0107] ,

[0108] where, is the dot product of the vector, is the norm of the task vector , is the norm of the knowledge entry vector ;

[0109] Sub-step , the system sets a similarity threshold based on the calculated semantic similarity score ; when the corresponding semantic similarity score of the knowledge entry vector satisfies the condition , the system selects the original knowledge entry text corresponding to the knowledge entry vector , and combines all the selected knowledge entry texts to form the decision context used in the subsequent steps ​.

[0110] In step four, further comprising:

[0111] Sub-step , based on the formed decision context , the system calls the prompt word engineering module to generate the structured task description and the decision context injects the preset prompt word template , and combines to generate an enhanced prompt word ; The prompt word engineering module performs integrity verification after generation, and when the task description contained in the enhanced prompt word and the decision context are both non-empty, it is confirmed that the enhanced prompt word is valid;

[0112] Sub-step , the confirmed valid enhanced prompt word is taken as input to call the Guangming Power large model to perform inference generation operation, the process can be represented as:

[0113] ,

[0114] Where, is the original text string containing the ordered power consumption scheme content output by the large model; the system performs validity determination on the original text string , when the original text string does not contain a preset error identifier and the text length is greater than the minimum length threshold, it enters subsequent processing;

[0115] Sub-step , based on the generated valid original text string , the system calls the pre-defined scheme parsing function to extract related fields from the original text string , and fills the extracted fields into the standard scheme data structure, finally forming a structured ordered power consumption scheme ;

[0116] After forming, the system performs the final integrity verification on the ordered power consumption scheme to ensure that all required key fields have been successfully filled, and outputs the scheme after verification.

[0117] In step five, further comprising:

[0118] Sub-step , based on the generated structured ordered power consumption scheme The system automatically invokes the review agent, which extracts all compliance constraints from the policy interpretation database and compares them one by one with the orderly electricity use plan. The fields in the comparison are compared;

[0119] Set Boolean compliance flag The initial value is true, and the power consumption scheme is in sequence. When all fields in the code satisfy the corresponding constraint clauses, the compliance flag is set. If the value remains true, the system will proceed to the next sub-step.

[0120] Sub-step If compliance flag bit If true, the system is targeting the orderly power consumption plan. Each user in the user list The load characteristic analysis agent is invoked to obtain historical load data sequences for 96 points. The user's historical peak load contribution is calculated using the following formula. Compared with the load response potential assessment value :

[0121] ,

[0122] ,

[0123] in, This is a set of historical peak load periods for the system. Total number of users The user's historical maximum load, The user's historical average load;

[0124] Sub-step Based on the calculated contribution of historical peak load Compared with the load response potential assessment value The system calls the indicator intelligent decomposition intelligent agent, first calculating the intelligent agents of each user. Allocation weights ,That, and The preset weighting coefficients satisfy the following conditions: ;

[0125] The orderly power consumption plan will be implemented later. Overall load control target The weighted allocation algorithm is used to calculate the allocation to each user. Load reduction:

[0126] Ultimately, a list of implementation indicators for each household will be formed.

[0127] Through the construction of a vectorized professional knowledge system, the deep integration and standardized processing of massive heterogeneous data in the power system are realized. This step will unify the user profile data and historical load data scattered in the marketing system and load acquisition system, as well as the implicit knowledge scattered in various policy documents and expert experience. Through semantic analysis technology, unstructured text is converted into structured knowledge items, and then through vectorization processing, a machine-understandable knowledge representation is formed. This solves the data silo problem under traditional methods and provides a rich, accurate, and computable knowledge base for subsequent intelligent decision-making, significantly improving the release degree of data value.

[0128] A major leap from traditional complex operation interface to natural language interaction is achieved, greatly reducing the system usage threshold. By calling the semantic understanding capability of the Guangming Power Large Model, the system can accurately analyze the user's requirements expressed in natural language, automatically identify task intent and extract constraints, and convert ambiguous human expression into structured task description. This innovation in human-computer interaction enables power business personnel to use the system without specialized training, while avoiding the cumbersome steps and potential errors in traditional interface operation, significantly improving work efficiency and user experience.

[0129] Through vector similarity calculation, accurate knowledge retrieval and intelligent matching are achieved, ensuring that the knowledge used in the decision-making process is both relevant and accurate. After vectorizing the task description, the system performs semantic similarity calculation in the constructed vector knowledge base to find the most relevant policy constraints, expert experience, and historical cases, forming a targeted decision context. This semantic understanding-based knowledge retrieval method is more intelligent and accurate than traditional keyword matching, avoiding decision bias caused by inaccurate knowledge retrieval and providing high-quality knowledge support for subsequent scheme generation.

[0130] Intelligent automatic generation of orderly power utilization schemes is achieved. By combining structured task description with selected decision context to form enhanced prompt words, the system can invoke the large model for deep reasoning, considering complex factors such as multi-dimensional constraints, user characteristics, load characteristics, etc., to automatically generate scientific and reasonable orderly power utilization schemes. This process breaks free from the limitations of traditional reliance on human experience and fixed rules, achieving a transition from experience-driven to data-driven and from static rules to dynamic reasoning, significantly improving the scientificity and optimality of generated schemes.

[0131] By leveraging multi-agent collaborative processes, the system automates the review and refinement of solutions, ensuring their compliance and feasibility. The review agent automatically verifies compliance with policies and regulations, avoiding omissions and errors that may occur during manual review. The intelligent decomposition agent scientifically allocates load based on users' historical load characteristics and response potential, precisely breaking down overall control objectives into tailored execution indicators for each user. This fully automated post-processing workflow guarantees solution quality and reduces execution plan development time from hours to minutes, significantly improving response speed and execution accuracy.

[0132] Example 2:

[0133] Please see the appendix Figure 3 This invention provides an automatic generation method for ordered power consumption schemes based on the Guangming large model, as follows:

[0134] I. Automatic generation of orderly power consumption plans

[0135] As attached Figure 3 As shown, the load management scheme auxiliary decision-making system achieves intelligent decision-making through a three-layer architecture of artificial intelligence platform, knowledge service, and Guangming big data model. In the orderly electricity consumption scheme development stage, the system embeds an orderly electricity consumption scheme intelligent agent invocation function, the specific implementation process of which is as follows:

[0136] Step 1: Local company staff input multi-dimensional data requirements, including region, industry, and peak electricity consumption periods, via text or voice through the system interface. The system, through its data collection module, calls basic service interfaces from the load management solution auxiliary decision-making database to obtain data on the user's historical load, adjustable load capacity, and industry-specific load reduction ratios.

[0137] Step Two: The system invokes the basic numerical calculation and analysis capabilities of the Guangming large model to perform load characteristic analysis on high-voltage users under the input conditions. The sub-agent calculates the user's load characteristic parameters using the following formula:

[0138] Maximum load calculation: ,

[0139] Average load calculation: ,

[0140] in, For a moment The load value, For a specified time period, This represents the number of sampling points within the time period.

[0141] Step three: The system calls the industry adjustable proportion standard library recommended by the State Grid Corporation through the knowledge service layer, and intelligently matches with the industry to which the power consumer belongs. The following algorithm is used to calculate the high-voltage user load reduction:

[0142] ,

[0143] Wherein, is the load reduction, is the industry load reduction coefficient, is the dynamic adjustment factor.

[0144] Step four: The system realizes user hierarchical control through the scheme evaluation module and the business logic reasoning ability of the Guangmingda model:

[0145] Exempt judgment is performed on six-protection five-limit users:

[0146] ,

[0147] Prioritized coding is performed on high-energy high-pollution enterprises:

[0148] ,

[0149] Comprehensive priority calculation:

[0150] ,

[0151] Wherein, and are weight coefficients, is the recommended priority, is the adjustable capacity evaluation value.

[0152] Finally, the system automatically generates a complete list of high-voltage user adjustable capacity, forming a complete and orderly power consumption scheme.

[0153] II. Intelligent decomposition of orderly power consumption execution indicators

[0154] Step one: The system obtains the current power reduction level and grade scheme approval load data through the feedback management interface, and the orderly power consumption scheme approval load identification sub-agent recommends the execution user and load reduction indicators according to the following logic:

[0155] ,

[0156] Step two: Based on the computing power of the Guangmingda model, the system calculates the correlation degree between the recent power consumption customer load before execution and the approved indicators:

[0157] ,

[0158] Wherein, The coefficient is dynamically adjusted, and is adaptively adjusted according to load fluctuation.

[0159] Step three: in combination with the analysis result of the user priority, the system automatically generates an ordered power consumption execution user list according to the enterprise recommendation priority and the size order of the adjustable ability by using the business logic reasoning ability of the large model.

[0160] Step four: the system realizes real-time monitoring of the ordered power consumption load execution by using the intelligent question and answer ability through the feedback result module.

[0161] Execution completion rate calculation:

[0162] ,

[0163] Early warning threshold determination:

[0164] When , the system automatically triggers the early warning mechanism, generates an early warning work order and pushes it to the relevant power officers.

[0165] Through the above implementation mode, the present application shortens the time of ordered power consumption scheme preparation , and the execution plan preparation time is shortened from the "hour level" to the "minute level", which effectively solves the technical problems of low scheme preparation efficiency and low execution plan decomposition precision in the prior art.

[0166] Although the embodiments of the present application have been shown and described, it can be understood by those skilled in the art that various changes, modifications, replacements and variations can be made to the embodiments without departing from the principles and spirits of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A method for automatically generating ordered electricity consumption schemes based on the Guangming large-scale model, characterized in that, include: Step 1: First, integrate user profile data and historical workload data, and at the same time collect policy and regulatory documents and expert scheduling experience. Through semantic parsing, form a policy interpretation library and an expert experience library. Finally, process all knowledge base content into vector form and store it in a vector database to complete the construction of the professional knowledge system. Step 2: Based on the constructed knowledge system, the system receives natural language text commands or graphical drag-and-drop operation commands input by the user, calls on the semantic understanding capabilities of the Guangming Power big data model, identifies the task intent and constraints from the commands, and forms a structured task description. Step 3: Based on the generated structured task description, the system processes the task description into task vectors, calculates the semantic similarity between the task vectors and knowledge entry vectors in the constructed vector database, and selects knowledge with similarity higher than a preset threshold to form a decision context. Step 4: Using the structured task description and knowledge context as input, the Guangming Power Big Model is invoked for reasoning, and finally the Big Model generates an ordered electricity consumption plan; Step 5: Based on the orderly electricity use plan, the system automatically calls the auditing agent to verify the compliance of the plan according to the policy interpretation library, and then calls the indicator decomposition agent to decompose the overall load control target of the plan into "one policy per household" execution indicators for users. Step five further includes: Sub-step Based on the generated structured and orderly power consumption scheme The system automatically invokes the review agent, which extracts all compliance constraints from the policy interpretation database and compares them one by one with the orderly electricity use plan. The fields in the comparison are compared; Set Boolean compliance flag The initial value is true, and the power consumption scheme is in sequence. When all fields in the code satisfy the corresponding constraint clauses, the compliance flag is set. If the value remains true, the system will proceed to the next sub-step. Sub-step If compliance flag bit If true, the system is targeting the orderly power consumption plan. Each user in the user list The load characteristic analysis agent is invoked to obtain historical load data sequences for 96 points. The user's historical peak load contribution is calculated using the following formula. Compared with the load response potential assessment value : ， ， in, This is a set of historical peak load periods for the system. Total number of users The user's historical maximum load, The user's historical average load; Sub-step Based on the calculated contribution of historical peak load Compared with the load response potential assessment value The system calls the indicator intelligent decomposition intelligent agent, first calculating the intelligent agents of each user. Allocation weights ,That, and The preset weighting coefficients satisfy the following conditions: ; The orderly power consumption plan will be implemented later. Overall load control target The weighted allocation algorithm is used to calculate the allocation to each user. Load reduction: Ultimately, a list of implementation indicators for each household will be formed.

2. The method for automatically generating an ordered electricity consumption scheme based on the Guangming large-scale model according to claim 1, characterized in that, The steps for calling the Guangming Power large model specifically include: The prompt word engineering module combines the structured task description with the knowledge context to form enhanced prompt words; these enhanced prompt words are then input into the Guangming Power big data model to drive the generation of the orderly electricity consumption scheme.

3. The method for automatically generating an ordered electricity consumption scheme based on the Guangming large-scale model according to claim 1, characterized in that, The steps for the auditing agent to verify the scheme specifically include: The load reduction values, execution periods, and user industry category parameters in the plan are automatically compared with the constraint clauses stored in the policy interpretation database; when a parameter exceeds the range of the constraint clauses, the plan is marked as non-compliant.

4. The method for automatically generating an ordered electricity consumption scheme based on the Guangming large-scale model according to claim 1, characterized in that, The intelligent decomposition agent of the index uses a preset weighted allocation algorithm to calculate the load reduction allocated to users based on their historical peak load contribution and load response potential assessment value.

5. The method for automatically generating an ordered electricity consumption scheme based on the Guangming large-scale model according to claim 4, characterized in that, The historical peak load contribution and load response potential assessment values ​​are obtained by the load characteristic analysis agent after performing time-series analysis on 96 historical load data points of the user.

6. The method for automatically generating an ordered electricity consumption scheme based on the Guangming large-scale model according to claim 1, characterized in that, Step one further includes: Sub-step First, user profile data and historical user workload data are integrated. Simultaneously, policy and regulatory documents and expert scheduling experience are collected to form a raw text corpus. Then, data cleaning is performed based on the historical workload data, and if any time points exist... load value Missing information is filled in using a linear interpolation algorithm. The calculation formula is as follows: ， in, Time to be completed The load value, and The closest moments before and after the missing point and The known load value; Sub-step Based on the acquired original text corpus, semantic parsing is performed using named entity recognition and relation extraction algorithms. When a text fragment simultaneously satisfies the condition that the main entity is a user category, the relation word is a scheduling behavior, and the object entity is a constraint value, the text fragment is extracted into a structured knowledge item and stored in the policy interpretation database and expert experience database according to the text source. Sub-step For the generated structured knowledge entries, a pre-trained text embedding model is invoked. Knowledge entries in various text formats Convert to a fixed-dimensional floating-point vector The conversion process is represented as follows: ， in, For knowledge entries The corresponding vector representation, For text embedding models; For the generated floating-point vector Perform validity verification when norm Greater than the preset non-zero threshold At that time, the floating-point vector The data is stored in a vector database, ultimately forming a complete vectorized domain knowledge base.

7. The method for automatically generating an ordered electricity consumption scheme based on the Guangming large-scale model according to claim 1, characterized in that, Step two further includes: Sub-step Based on the constructed vectorized domain knowledge base, the system receives natural language text commands or graphical drag-and-drop operation commands input by the user, and converts the received raw commands into... Through the preset normalization function Convert to a uniform text string This process is represented as: ， When the text string character length Greater than the preset minimum instruction length threshold When this happens, the subsequent processing flow is triggered. Sub-step The system will generate a text string. The input is fed into the semantic understanding module of the Guangming Power big data model, which performs intent recognition and entity extraction in parallel; the intent recognition algorithm outputs candidate intents. and the corresponding intent confidence level Entity extraction algorithms extract entity types from text. and entity value The entity set : ， When the confidence level of intent Meet the conditions At that time, confirm the candidate intention For the ultimate mission intent, among which, Set a preset confidence threshold for intent confirmation; Sub-step Based on the confirmed final mission intent and the extracted entity set The system is for the task intent Matching predefined task templates The template contains a set of required entity slots; the system uses the entity set... entity values ​​in Fill into task template The corresponding entity type To generate structured task descriptions ; Before generation, the system performs an integrity check to ensure the task template is valid. All necessary entity slots can be found in the entity collection. Find the corresponding fill value, and after successful verification, finally generate the structured task description. .

8. The method for automatically generating an ordered electricity consumption scheme based on the Guangming large-scale model according to claim 1, characterized in that, Step three further includes: Sub-step Based on the generated structured task description The system first serializes the data into a single text string. Then the same pre-trained text embedding model is called. , text string Convert to task vector The conversion process is represented as follows: ， After the transformation is complete, the task vector Perform validity verification when norm When the value is greater than zero, confirm the task vector. efficient; Sub-step Use the confirmed valid task vector The system traverses the vectors of each knowledge entry in the vector database. Furthermore, the cosine similarity algorithm is used to calculate the task vector. With each knowledge item vector semantic similarity score between The calculation formula is: ， in, The dot product of vectors, Task vector of Norm, Knowledge Entry Vector of Norm; Sub-step The system uses the calculated semantic similarity score as a basis. Set a similarity threshold When the knowledge entry vector The corresponding semantic similarity score meets the conditions At that time, the system will generate knowledge entry vectors. The corresponding original knowledge entry texts are selected, and all selected knowledge entry texts are combined to form a decision context for subsequent steps. .

9. The method for automatically generating an ordered electricity consumption scheme based on the Guangming large-scale model according to claim 1, characterized in that, Step four further includes: Sub-step Based on the formed decision context The system calls the prompt word engineering module to generate a structured task description. With the decision context Inject preset prompt template In the middle, combine to generate enhanced prompt words The prompt word engineering module performs an integrity check after generation. When the enhanced prompt word... When both the task description and decision context are not empty, confirm the enhanced prompt words. efficient.

10. The method for automatically generating an ordered electricity consumption scheme based on the Guangming large model according to claim 9, characterized in that, Step four further includes: Sub-step Confirmed valid enhancement prompts As input, the Guangming Power large model is called. The process of performing inference generation operations can be represented as follows: ， in, The system outputs the original text string containing the ordered electricity consumption scheme from the large model; the system processes the original text string. Perform a validity determination when the original text string If the text does not contain a preset error identifier and its length exceeds the minimum length threshold, proceed to the next step. Sub-step Based on the generated valid original text string The system calls a predefined scheme resolution function. The original text string Relevant fields are extracted and populated into the standard scheme data structure to form a structured and ordered electricity consumption scheme. ; After the system is formed, it will implement the ordered power consumption scheme. Perform a final integrity check to ensure that all required key fields have been successfully filled. Output the solution after the check passes.

Images