Method and related equipment for power grid task self-matching and dynamic scheduling of large language model and multi-agent cooperation

Abstract

The application relates to a power grid task self-matching and dynamic scheduling method and related equipment based on a large language model and multi-agent cooperation. The method comprises the following steps: using a large model to analyze and model task description information to obtain initial structured task information, calibrating the initial structured task information based on a power grid task semantic graph to obtain standard structured task information, and generating a power grid task based on the standard structured task information; performing multi-dimensional vector modeling to obtain resource objects; calculating the matching degree of the power grid task and the resource objects, determining the candidate resource objects of the power grid task based on the matching degree; making the task agent and the resource agent cooperate and schedule to determine the target resource object of the power grid task and generate task resource mapping information; and reviewing the task resource mapping information and assigning the task. The application can improve the task semantic analysis accuracy, improve the matching accuracy of resources and tasks, improve the adaptive and optimization capability of scheduling, and effectively avoid compliance risks.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power grid task scheduling, in particular to a power grid task self-matching and dynamic scheduling method based on large language model and multi-agent cooperation and related equipment. BACKGROUND

[0002] At present, part of the task management system has introduced a workflow engine based on BPM (Business Process Modeling) to realize the automatic circulation of task nodes, but still relies on artificial pre-set process templates, which is difficult to adapt to dynamic changes in business scenarios. For task scheduling, although there are task allocation technologies based on heuristic algorithms, linear programming, genetic algorithms and other algorithms, these methods are mostly static optimization, which cannot make dynamic optimization decisions according to task changes and resource state changes, and cannot adapt to task changes, resource state changes or emergencies in real time, making it difficult to improve the task response efficiency and resource utilization.

[0003] In terms of semantic understanding, traditional task management systems usually require structured input of tasks, and have weak analysis capabilities for task intentions or information appeals expressed in natural language, and cannot automatically identify task elements (such as task type, priority, time limit, dependency relationship, etc.) from unstructured input of users or systems, which limits the versatility and flexibility of the task system. SUMMARY

[0004] In view of the above deficiencies or shortcomings, the present application provides a power grid task self-matching and dynamic scheduling method based on large language model and multi-agent cooperation and related equipment, which can improve the task semantic analysis accuracy, improve the matching accuracy of resource objects and power grid tasks, and improve the adaptive and intelligent optimization ability of scheduling, effectively avoiding compliance risks.

[0005] According to the first aspect, the present application provides a power grid task self-matching and dynamic scheduling method based on large language model and multi-agent cooperation, in some embodiments, the method comprises:

[0006] The task description information is collected, the task description information is preprocessed, the preprocessed task description information is subjected to semantic analysis and structured modeling by using a power grid fine-tuning large model to obtain initial structured task information, the initial structured task information is subjected to semantic calibration based on a power grid task semantic graph to obtain standard structured task information, the standard structured task information is used to generate a power grid task, and the power grid task is added to a task pool; multi-dimensional resource object information is collected, multi-dimensional vector modeling is performed based on the multi-dimensional resource object information to obtain resource objects represented by multi-dimensional vectors, and the resource objects are added to a resource pool; based on a first matching degree scoring mechanism, the matching degree of each to-be-scheduled power grid task in the task pool and each resource object in the resource pool is calculated, and a plurality of first candidate resource objects corresponding to each to-be-scheduled power grid task are determined based on the matching degree calculation result; a task intelligent agent is allocated to each to-be-scheduled power grid task, a resource intelligent agent is allocated to each first candidate resource object, each task intelligent agent and each resource intelligent agent perform collaborative scheduling in a multi-agent reinforcement learning environment to determine a target resource object corresponding to each to-be-scheduled power grid task, and a plurality of task-resource mapping information are generated; each piece of task-resource mapping information is subjected to compliance review according to a compliance knowledge graph; and task dispatching is performed according to the task-resource mapping information that passes the compliance review.

[0007] In some embodiments, the power grid task semantic graph includes a plurality of standard task nodes; the semantic calibration of the initial structured task information based on the power grid task semantic graph to obtain the standard structured task information includes: determining a target task node from the plurality of standard task nodes according to the initial structured task information; calculating a matching degree between the initial structured task information and node information of the target task node by using a graph neural network; when the matching degree is greater than a preset threshold, obtaining the node information of the target task node as the standard structured task information; and when the matching degree is less than or equal to the preset threshold, returning to the step of performing semantic analysis and structured modeling on the preprocessed task description information by using the power grid fine-tuning large model.

[0008] In some embodiments, the calculation of the matching degree of each to-be-scheduled power grid task in the task pool and each resource object in the resource pool includes: embedding the power grid task into a vector space to obtain a task vector; calculating a skill matching score, a response ability score and a success rate score of each resource object based on the task vector; and calculating the matching degree of the task vector and each resource object based on the skill matching score, the response ability score and the success rate score of each resource object.

[0009] In some embodiments, the compliance review of each piece of task-resource mapping information according to the compliance knowledge graph comprises: taking each piece of task-resource mapping information as a review object, performing graph matching on the review object in the compliance knowledge graph to obtain a compliance path corresponding to the review object; verifying whether the review object satisfies all logical rules of the compliance path using a rule reasoning engine; if the review object satisfies all logical rules of the compliance path, determining that the review object passes the compliance review; and if there is no compliance path corresponding to the review object in the compliance knowledge graph, or the review object does not satisfy any logical rule of the compliance path, determining that the review object fails to pass the compliance review.

[0010] In some embodiments, after determining that the review object fails to pass the compliance review, the method further comprises: determining a conflict type between the review object and the compliance path; determining a disposal measure of a to-be-scheduled power grid task related to the review object according to the conflict type, and executing the disposal measure.

[0011] In some embodiments, the collaborative scheduling of the task agents and the resource agents in the multi-agent reinforcement learning environment is performed to determine the target resource object corresponding to each to-be-scheduled power grid task, comprising: constructing a scheduling graph; the scheduling nodes in the scheduling graph are each to-be-scheduled power grid task and each first candidate resource object, and the edges between any two scheduling nodes represent one or more relationships between the two scheduling nodes; the one or more relationships comprise an organizational constraint relationship; each task agent determines a target resource agent to be connected according to the scheduling graph, and sends a task matching request to all target resource agents in parallel; each resource agent, after receiving the task matching request, performs benefit evaluation based on the task information carried by the task matching request, and returns response information indicating acceptance of the task or rejection of the task to the task agent sending the task matching request based on the benefit evaluation result; the second candidate resource object corresponding to each to-be-scheduled power grid task is determined based on the response information of each resource agent; the matching degree of each to-be-scheduled power grid task and each second candidate resource object corresponding thereto is calculated based on a second matching degree scoring mechanism; and the target resource object corresponding to each to-be-scheduled power grid task is determined according to the matching degree of each to-be-scheduled power grid task and each second candidate resource object corresponding thereto.

[0012] In some embodiments, after the task dispatching according to the task-resource mapping information that passes the compliance review, the method further comprises: collecting multi-dimensional execution records of each power grid task after execution; dynamically adjusting the policy parameters in the second matching degree scoring mechanism through a reinforcement learning reward function based on the multi-dimensional execution records, and / or screening high-quality execution records meeting a preset quality requirement from the multi-dimensional execution records of the plurality of power grid tasks, and updating the power grid task semantic graph based on the high-quality execution records.

[0013] The application provides a power grid task self-matching and dynamic scheduling device based on a large language model and a multi-agent according to a second aspect. In some embodiments, the device comprises:

[0014] A power grid task construction module is configured to collect task description information, pre-process the task description information, perform semantic analysis and structured modeling on the pre-processed task description information using a power grid fine-tuning large model, obtain initial structured task information, perform semantic calibration on the initial structured task information based on a power grid task semantic graph, obtain standard structured task information, generate a power grid task based on the standard structured task information, and add the power grid task to a task pool.

[0015] A resource object construction module is configured to collect multi-dimensional resource object information, perform multi-dimensional vector modeling based on the multi-dimensional resource object information, obtain resource objects represented by multi-dimensional vectors, and add the resource objects to a resource pool.

[0016] A first resource object screening module is configured to calculate the matching degree of each to-be-scheduled power grid task in the task pool and each resource object in the resource pool based on a first matching degree scoring mechanism, and determine a plurality of first candidate resource objects corresponding to each to-be-scheduled power grid task based on the matching degree calculation results.

[0017] A second resource object screening module is configured to assign a task agent to each to-be-scheduled power grid task, assign a resource agent to each first candidate resource object, and enable the task agents and the resource agents to perform collaborative scheduling in a multi-agent reinforcement learning environment to determine a target resource object corresponding to each to-be-scheduled power grid task and generate a plurality of task-resource mapping information.

[0018] A compliance review module is configured to perform compliance review on each piece of task-resource mapping information according to a compliance knowledge graph.

[0019] A task assignment module is configured to perform task assignment based on the task-resource mapping information that passes the compliance review.

[0020] The application provides a computer device according to a third aspect, which comprises a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, the steps of the power grid task self-matching and dynamic scheduling method based on a large language model and a multi-agent provided in any of the embodiments are implemented.

[0021] The application provides a computer readable storage medium according to a fourth aspect, which stores a computer program. When the computer program is executed by a processor, the steps of the power grid task self-matching and dynamic scheduling method based on a large language model and a multi-agent provided in any of the embodiments are implemented.

[0022] In the above embodiments of the present application, by introducing the joint modeling technology of large language model (i.e., power grid fine-tuning large model) and knowledge graph (i.e., power grid task semantic graph), the semantic understanding and structured modeling capability for unstructured task input are greatly improved, the automatic extraction and standardization of task information are realized, the semantic drift phenomenon such as polysemy and non-standard description can be avoided, and the problem of low task semantic analysis accuracy is solved. In the resource modeling layer, the resource objects are modeled by multi-dimensional attributes, and the resource objects are represented by multi-dimensional vectors, which can improve the matching accuracy of resource objects and power grid tasks, and further make the subsequent task scheduling more personalized and more consistent with the business. On this basis, the task scheduling mechanism based on multi-agent reinforcement learning enables multiple task agents and resource agents to cooperate in scheduling, which can make optimal or suboptimal matching decisions in complex scenarios with dynamic task changes and resource state fluctuations, and has significant adaptive and intelligent optimization capabilities, thereby improving the scheduling response efficiency and resource utilization. After generating the task resource mapping information, compliance review is performed to ensure that each task dispatch meets the requirements of industry standards, personnel qualifications, safety regulations and other requirements, and effectively avoids the compliance risk of scheduling results. In addition, the present application also provides a continuous learning mechanism based on scheduling execution feedback, which enables the system to continuously accumulate experience and correct deviations from the task execution results, and realizes the self-evolution of the strategy. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 A flowchart of a large language model and multi-agent cooperative power grid task self-matching and dynamic scheduling method according to one or more embodiments of the present application is provided.

[0024] Figure 2 A structural block diagram of a large language model and multi-agent cooperative power grid task self-matching and dynamic scheduling device according to one or more embodiments of the present application is provided.

[0025] Figure 3 An internal structure diagram of a computer device according to one or more embodiments of the present application is provided. DETAILED DESCRIPTION

[0026] In order to make the purpose, technical scheme and advantages of the present application clearer, the embodiments of the present application will be described in detail below with reference to the drawings. It should be clear that the described embodiments are only some of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0027] The following description refers to the accompanying drawings. Unless otherwise noted, same or similar components in different drawings have same or similar reference numerals. The implementations described in the following exemplary embodiments are not meant to represent all implementations consistent with the present disclosure. Rather, they are simply examples of apparatuses and methods consistent with some aspects of the present disclosure as detailed in the appended claims.

[0028] In the description of the present application, it should be understood that the terms "first", "second", "third", etc. are only used to distinguish similar objects, and do not necessarily mean a specific order or sequence, nor can they be understood as indicating or implying relative importance. For those of ordinary skill in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances. In addition, in the description of the present application, "multiple" means two or more, unless otherwise specified. The "and / or" describes the association relationship between the associated objects, which means that there can be three relationships, for example, A and / or B can represent: A exists alone, A and B exist together, and B exists alone. The character " / " generally represents that the associated objects before and after are in an "or" relationship.

[0029] In view of the deficiencies or defects of the related art, the present application provides a power grid task self-matching and dynamic scheduling method based on large language model and multi-agent collaboration. The method is suitable for digital management scenarios with task scheduling, task matching, resource allocation, and collaborative execution as the core. The method can be widely applied to job task scheduling, inspection management, marketing inspection, fault repair, and other scenarios of power grid enterprises. Moreover, since the method provides a unified task governance and intelligent allocation mechanism, it is particularly suitable for environments with a large number of heterogeneous tasks, multi-source resource participation, and real-time dynamic adjustment requirements. In addition, the method is not only suitable for power grid enterprises, but also suitable for other industries such as government affairs processes, financial task allocation, and manufacturing operation and maintenance. The present application will take power grid enterprises as an example to illustrate the method.

[0030] The method can be applied to a server for implementing a series of operations such as task management, resource management, and task scheduling. The server can be a standalone server, and can also be a server cluster. Please refer to Figure 1The method includes steps S110-S160. The method includes at least the following improvements: (1) By introducing a large language model (i.e., a power grid fine-tuning large model) and knowledge graph (i.e., a power grid task semantic graph) joint modeling technology, the semantic understanding and structured modeling capability for unstructured task input is greatly improved, the automatic extraction and standardization of task information is realized, the semantic drift phenomenon such as polysemy and non-standard description can be avoided, and the problem of low task semantic analysis accuracy is solved. (2) At the resource modeling level, the resource object is modeled by multi-dimensional attributes, and the resource object is represented by a multi-dimensional vector, which can improve the matching accuracy of resource objects and power grid tasks, and further make the subsequent task scheduling more personalized and more closely related to the business. (3) The task scheduling mechanism based on multi-agent reinforcement learning allows multiple task agents and resource agents to collaborate in scheduling, which can achieve optimal or suboptimal matching decisions in complex scenarios with dynamic task changes and resource state fluctuations, and has significant adaptive and intelligent optimization capabilities, thereby improving the scheduling response efficiency and resource utilization. (4) After generating task resource mapping information, compliance review is performed to ensure that each task dispatch meets industry standards, personnel qualifications, safety regulations and other requirements, effectively avoiding compliance risks of scheduling results. (5) The continuous learning mechanism based on scheduling execution feedback enables the system to continuously accumulate experience and correct deviations from the task execution results, and realizes strategy self-evolution.

[0031] The various steps shown in the flowchart of FIG. 10 will be described in detail below with some embodiments. Figure 1

[0032] S110: Collect task description information, pre-process the task description information, use the power grid fine-tuning large model to perform semantic analysis and structured modeling on the pre-processed task description information, obtain initial structured task information, perform semantic calibration on the initial structured task information based on the power grid task semantic graph, obtain standard structured task information, generate a power grid task based on the standard structured task information, and add the power grid task to the task pool.

[0033] ​The task description information is used as the only original input source for subsequent semantic analysis, task type identification, resource matching, and order allocation decision. It is in the form of natural language or semi-structured text, and describes the information of the field operation tasks such as power grid operation and maintenance, repair, and dispatch. There can be multiple channels to collect task description information, which can be collected in real time through multiple channels and multiple protocols from people, machines, and systems. For example, an operation and maintenance personnel inputs "Please arrange to repair the main transformer of the substation in the south area this afternoon, and two people are required to work simultaneously" and "SF6 low pressure alarm of 303 switch on 35 kV side of #2 main transformer of 220 kV station, need to complete gas supplement and signal restoration within 2 hours" through a mobile terminal. For another example, a dispatch center operator fills in the text description in a defect single, a repair application, a test plan, and a patrol work order in an OMS system. Moreover, the task description information is not limited to text, and it can be natural language text, voice input, image screenshot, PDF (Portable Document Format) file, Word file, or multi-modal combined input.

[0034] The pre-processing of the task description information can convert the original data into structured or semi-structured text that can be directly understood by the power grid fine-tuning large model, including but not limited to time, location, task type, keyword, object, time, number of people required, qualification requirements, and other field information. The pre-processing operation can reduce the input noise of the power grid fine-tuning large model, improve the accuracy of subsequent semantic analysis, and retain traceable structured intermediate results.

[0035] The pre-processing can at least include the following operations: a. character cleaning, such as removing HTML tags, special symbols, and expression information; b. word segmentation and part-of-speech tagging, such as identifying professional vocabulary from the task description information through domain dictionary enhancement and other means, such as "SF6", "main transformer", etc.; c. named entity recognition, such as extracting key information such as site name, equipment number, voltage level, time limit, etc. from the task description information; d. dependency syntax analysis: used to locate the action-recipient relationship (such as "gas supplement → switch"); e. spelling correction, which can combine domain word library and language model to correct synonyms, such as unifying "low pressure" and "low pressure".

[0036] The power grid fine-tuning large model refers to a large language model obtained by fine-tuning a general large language model (such as ChatGLM, BLOOM, DeepSeek) using power grid domain corpus (such as defect report, maintenance procedure, dispatching instruction, safety regulation article). Compared with the general large language model, it has the ability to deeply understand the professional semantics of the power grid. The power grid fine-tuning large model is used for context semantic analysis and modeling of preprocessed task description information, and outputs initial structured task information, realizing accurate conversion from natural language to power grid knowledge structure. The initial structured task information refers to the output result of the power grid fine-tuning large model, which includes but is not limited to structured fields such as task type, professional attribute, processing object, key target, time requirement and constraint condition. Exemplarily, the field value of the task type can be device inspection, defect processing, fault repair, device replacement, maintenance test, safety inspection, marketing inspection, user visit, data archiving, work order audit, project collaboration, and other business self-defined types; the field value of the professional attribute can be power transmission, power transformation, power distribution, marketing, safety supervision, maintenance and operation, customer service, project construction, asset management, information communication, and comprehensive management; the field value of the processing object can be the south substation, 10kV switch cabinet No. 123, XX area user group, XX line number, certain data system interface, certain marketing inspection case, and specific device identification number (such as meter code and cable code), which is usually bound with the task type and supports automatic extraction of associated relationship combined with the knowledge graph; the field value of the key target can be inspection completion (patrol type task), defect elimination (processing type task), fault recovery (repair type task), data correction or archiving (information type task), case verification and case filing (inspection task), result review and feedback (process type task), and achievement reporting (plan type task); the field value of the time requirement can be start time, deadline, and response time limit.

[0037] For example, assuming that the task description information is “Please arrange to maintain the main transformer of the south substation this afternoon, and two people need to work simultaneously.”, the output result of the power grid fine-tuning large model can be as follows:

[0038] {“task type”: “power transformation maintenance”;

[0039] “processing object”: “main transformer of the south substation”;

[0040] “work time”: “this afternoon”;

[0041] “execution number requirement”: “2 people”;

[0042] “collaboration requirement”: “need two people to work simultaneously”;

[0043] “constraint condition”: “need to have high-voltage operation qualification”}.

[0044] For example, assume that the task description information is a PDF file containing "Work Order No.: 220331-BJ301; Task Content: Equipment replacement (110kV switch cabinet); Deadline: before 18:00 on April 15, 2024; Notes: need live operation qualification", the output result of the power grid fine-tuning large model can be as follows:

[0045] {“task type”: “equipment replacement”;

[0046] “processing object”: “110kV switch cabinet”;

[0047] “task ID”: “220331-BJ301”;

[0048] “deadline”: “2024-04-15 18:00”;

[0049] “qualification requirement”: “need live operation certificate”}.

[0050] The power grid task semantic graph is a knowledge graph that conforms to the pre-constructed power grid knowledge classification system (including task type ontology, relationship definition, upper and lower label, semantic similarity rule, etc.). It is used to perform semantic calibration and standardization on the initial structured task information generated by the power grid fine-tuning large model, which can eliminate the ambiguity that may exist in the initial structured task information, ensure that the task labels (related information of the power grid task, including but not limited to time, place, task type, keyword, object, time, number of people required, qualification requirement, etc. Field) are consistent within the entire network, and ensure the accuracy of subsequent resource matching and compliance review.

[0051] The power grid knowledge classification system is obtained by dividing the knowledge of the power grid enterprise using a multi-dimensional classification method. Common classification dimensions include knowledge type, business, specialty, equipment asset, and institution. Among them, knowledge type, business, and specialty can be used as main dimensions, for example, first divide multiple knowledge types, then divide multiple businesses under each knowledge type, and then divide multiple specialties under each business. Equipment asset and institution can be used as auxiliary dimensions.

[0052] The construction process of the power grid task semantic graph can be as follows:

[0053] (1) Extract the main dimension knowledge label as the task classification backbone. The task type ontology can be constructed in a three-dimensional combination of “knowledge type + business classification + specialty classification”; each combination is considered as a standardized task category, forming a semantic tree structure; for example: “operation guide > power transmission business > channel inspection”.

[0054] (2) Introduce non-standard expressions in historical task data. Non-standard expressions can be extracted from historical task data such as work order systems and maintenance platforms; build synonym / alias tables, such as "main transformer inspection" → "main transformer patrol".

[0055] (3) Establish semantic relationships between tasks and devices, organizational entities. You can use "has_asset", "has_department", "has_object" and other relationship edges to connect: "substation maintenance task" → has_asset → "10kV switch cabinet", "line patrol" → has_department → "power transmission and operation center".

[0056] (4) Set the scheduling meta attribute of each task category, including the professional of the task, the typical processing object, the default priority, the required skill certificate, etc.; facilitate subsequent structured output of tasks and resource capability adaptation.

[0057] (5) Build a graph service interface and connect it to the task scheduling system. You can provide a variety of graph API interfaces to the task scheduling system, such as querying standard task categories, mapping task labels to standard categories, and returning relevant power grid task scheduling rules and compliance requirements.

[0058] (6) Implement semantic standardization and label normalization services, such as inputting the initial structured task information output by the power grid fine-tuning large model into the mapping module of the power grid task semantic graph to achieve task label standardization, semantic clustering, ambiguity elimination, etc.

[0059] The power grid task semantic graph includes multiple standard task nodes, and the node information of the standard task node is also the task label, which includes but is not limited to time, place, task type, keyword, object, time, number of people required, qualification requirements, etc. The description of these fields can be referred to in the above, and will not be repeated here. Based on the power grid task semantic graph, the steps of performing semantic calibration on the initial structured task information to obtain standard structured task information can include: determining a target task node from the multiple standard task nodes according to the initial structured task information; using a graph neural network to calculate the matching degree between the initial structured task information and the node information of the target task node; when the matching degree is greater than a preset threshold, obtaining the node information of the target task node as the standard structured task information; when the matching degree is less than or equal to the preset threshold, returning to the step of using the power grid fine-tuning large model to perform semantic analysis and structured modeling on the preprocessed task description information.

[0060] The standard task nodes are usually more, and can be combined with hash dictionary, vector disambiguation and graph traversal to realize faster and more accurate positioning of the target task node (i.e. the standard task node most matched with the initial structured task information). For example, first build an inverted index of each standard task node of the power grid task semantic graph, and then build a hash dictionary. After obtaining the initial structured task information, the hash bucket is directly searched using the information, so that the node ID of the standard task node uniquely corresponding to the information can be located in seconds. If there are multiple standard task nodes corresponding to the initial structured task information, a graph neural network (such as a relational graph convolution network) can be used to calculate the similarity (such as the cosine similarity) between the initial structured task information and the node information of each standard task node. Finally, the standard task node with the highest similarity is taken as the target task node.

[0061] If the matching degree between the initial structured task information and the node information of the target task node is greater than a preset threshold, the node information of the target task node is taken as the standard structured task information. If the matching degree is less than or equal to the preset threshold, the step of using the power grid fine-tuning large model to perform semantic analysis and structured modeling on the preprocessed task description information is returned, so as to regenerate the initial structured task information from the power grid fine-tuning large model. If the initial structured task information generated by the power grid fine-tuning large model multiple times does not find a standard task node with a matching degree greater than the preset threshold, artificial correction of the initial structured task information can be requested to ensure that the task can be smoothly dispatched.

[0062] S120: Collect multi-dimensional resource object information, perform multi-dimensional vector modeling based on the multi-dimensional resource object information, obtain a resource object represented by a multi-dimensional vector, and add the resource object to a resource pool.

[0063] The resource object refers to the smallest entity unit that can be dispatched, allocated and bear specific work in the process of executing the power grid task, and usually exists in the form of personnel, team, tool, vehicle, spare parts, test equipment, unmanned aerial vehicle, robot, etc. For example, it can be "Mr. Li of the 220 kV professional maintenance team", "SF6 gas filling special tool kit", and "high-altitude working vehicle with license plate Beijing A12345". As the final executor of the dispatching system, the resource object can realize accurate matching and dynamic allocation of tasks and resources by modeling its multi-dimensional attributes such as capability, state and historical performance.

[0064] Multi-dimensional resource object information refers to all attribute data collected and integrated from multiple aspects such as static attributes, dynamic states, behavior records, and preference factors, for fully describing the capability, real-time state, and historical performance of each resource object. Static attributes may be, for example, qualifications, skill levels, professional labels, etc.; dynamic states may be, for example, current task load (e.g., class member A, task load 3, which are respectively the morning 9:00-10:00 inspection task, morning 11:00-12:00 defect processing, and afternoon 2:00-3:00 substation data verification. At this time, the current load level of the resource is: number of tasks: 3; current task density: 3 / 8 hours = 37.5%), current location, availability (e.g., remaining available working hours). Behavior records may be, for example, historical success rate (e.g., task success rate in the last 6 months), timeliness (e.g., average response time), preferred shift (e.g., preference for day shift or night shift), etc.

[0065] The multi-dimensional vector modeling method may be to convert attribute data of each aspect into a vector, and then splice vectors of each aspect, so that the final multi-dimensional vector is: static attribute vector + dynamic state vector + behavior record vector.

[0066] The preference factor represents the preference or selection habit of a certain resource under certain conditions, such as a resident area, a frequent collaborator, or a preference for processing a certain type of task. The preference factor does not participate in multi-dimensional vector modeling, and it can be used as a weighted correction item or a decision priority factor in the first matching degree scoring mechanism.

[0067] S130: Based on the first matching degree scoring mechanism, the matching degree of each to-be-scheduled power grid task in the task pool and each resource object in the resource pool is calculated, and a plurality of first candidate resource objects corresponding to each to-be-scheduled power grid task are determined based on the matching degree calculation result.

[0068] The to-be-scheduled power grid task refers to a power grid task to be scheduled. Each power grid task has corresponding indication information indicating the scheduling state (e.g., scheduled, to-be-scheduled, etc.), and the to-be-scheduled power grid task can be determined through the indication information.

[0069] The step of calculating the matching degree of each to-be-scheduled power grid task in the task pool and each resource object in the resource pool may include: embedding the power grid task into a vector space to obtain a task vector; calculating the skill matching score, response capability score, and success rate score of the task vector and each resource object; and calculating the matching degree of the task vector and each resource object according to the skill matching score, response capability score, and success rate score of the task vector and each resource object.

[0070] The first matching degree scoring mechanism is to calculate the matching degree of each to-be-scheduled power grid task and each resource object in the resource pool according to a preset matching degree scoring function. The matching degree scoring function may be:

[0071] Matching degree = a * skill matching score + b * response ability score + g * success rate score.

[0072] The sum of the three weights a, b and g is 1, and the system can dynamically adjust their values.

[0073] The skill matching score is the product of the value of the skill-related attribute (such as skill) in the task vector and the value of the skill-related attribute in the resource object; the response ability score is the product of the value of the response ability-related attribute (such as response time limit and response ability) in the task vector and the value of the response ability-related attribute in the resource object; and the success rate score is the product of the value of the success rate-related attribute (such as historical success rate) in the task vector and the value of the success rate-related attribute in the resource object. For example, the structure of the task vector T is [skill: 0.9, response time limit: 0.6, collaboration demand: 0.2], and the structure of the resource vector R is [skill: 0.85, availability: 0.7, response ability: 0.75, success rate: 0.92], then the skill matching score = 0.9 * 0.85 = 0.765, the response ability score = 0.6 * 0.75 = 0.45, and the success rate score = 0 * 0.92 = 0.

[0074] S140: assign a task agent to each to-be-scheduled power grid task, assign a resource agent to each first candidate resource object, and make each task agent and each resource agent perform collaborative scheduling in a multi-agent reinforcement learning environment to determine the target resource object corresponding to each to-be-scheduled power grid task and generate a plurality of task resource mapping information.

[0075] The task resource mapping information is the mapping information between the to-be-scheduled power grid task and the target resource object, which can be the task ID of the related power grid task and the resource ID of the resource object.

[0076] The embodiment can use a MARL (Multi-Agent Reinforcement Learning) framework, assign a task agent to each to-be-scheduled power grid task, assign a resource agent to each first candidate resource object, and then make each task agent and each resource agent perform collaborative scheduling.

[0077] The step of making each task agent and each resource agent in the multi-agent reinforcement learning environment to be cooperatively scheduled to determine the target resource object corresponding to each to-be-scheduled power grid task can include: constructing a scheduling graph; the scheduling nodes in the scheduling graph are each to-be-scheduled power grid task and each first candidate resource object, and the edges between any two scheduling nodes represent one or more relationships between the two scheduling nodes; the one or more relationships include an organizational constraint relationship; each task agent determines the target resource agent that needs to be connected according to the scheduling graph, and sends a task matching request to all target resource agents in parallel; after each resource agent receives the task matching request, it performs benefit evaluation based on the task information carried by the task matching request, and returns response information indicating acceptance or rejection of the task to the task agent that sent the task matching request based on the benefit evaluation result; determine the second candidate resource object corresponding to each to-be-scheduled power grid task based on the response information of each resource task agent; based on the second matching degree scoring mechanism, calculate the matching degree of each to-be-scheduled power grid task and each second candidate resource object corresponding thereto; determine the target resource object corresponding to each to-be-scheduled power grid task according to the matching degree of each to-be-scheduled power grid task and each second candidate resource object corresponding thereto.

[0078] The flow of cooperative scheduling can be as follows:

[0079] 1. Initialize the scheduling period: in the multi-agent reinforcement learning environment, load the initial state of the task pool and the resource pool, construct the scheduling graph at time t=0 in combination with the initial state and the preset organizational constraints. The scheduling nodes in the scheduling graph are two kinds, which are each to-be-scheduled power grid task and each first candidate resource object, and the edges between any two scheduling nodes are constructed on demand, which are used to represent one or more relationships between the two scheduling nodes, such as organizational constraint relationship, matching relationship, that is, there can be multiple edges between two scheduling nodes, such as edges representing organizational constraint relationship, edges representing matching relationship, regarding the construction of edges, taking the edges representing matching relationship as an example, they can be constructed according to the matching degree score calculated in step S130, such as if the matching degree score is greater than a preset threshold, the edge is constructed, and if it is less than or equal to the preset threshold, the edge is not constructed.

[0080] 2. The task agent initiates a task matching request: each to-be-scheduled power grid task is assigned a corresponding task agent, and each task agent determines the target resource agent that needs to be connected according to the scheduling graph, for example, according to whether there is an edge representing a matching relationship between the scheduling node corresponding to the to-be-scheduled power grid task (referring to the to-be-scheduled power grid task that the task agent is responsible for) and the scheduling node corresponding to other resource objects, the resource agent related to the other scheduling node connected to the to-be-scheduled power grid task through the edge representing the matching relationship is taken as the target resource agent. Then, a task matching request is sent to each target resource agent, and the request can carry the task information of the to-be-scheduled power grid task, such as the task vector.

[0081] 3. The resource agent evaluates the willingness to accept the task: each resource agent evaluates whether to accept the task according to the local state, which can include the following judgment logic: (1) whether there is an idle period; (2) whether the qualification and skill requirements are met; (3) whether the task matches the preference or historical experience; (4) whether there is a conflict in the multi-task scheduling sequence. When any of the judgment logics is true, there is a benefit (for example, one point can be obtained for each logic that is true, and 0 points can be obtained for each logic that is not true), and after all the judgment logics are completed, the sum of all the scores is taken as the total benefit. If the total benefit exceeds a preset threshold (such as 2 points), it is determined to accept the task, and if the total benefit does not exceed the preset threshold (such as 2 points), it is determined to reject the task.

[0082] 4. Matching score calculation and bidding game.

[0083] For any to-be-scheduled power grid task, if the resource agent feedbacks response information indicating acceptance of the task, the resource object responsible for the resource agent is the second candidate resource object corresponding to the to-be-scheduled power grid task. If multiple resource agents feedback response information indicating acceptance of the task to the task agent related to the to-be-scheduled power grid task, the scheduling system on the server can calculate the matching degree score of the to-be-scheduled power grid task and the resource object responsible for each resource agent based on the second matching degree scoring mechanism.

[0084] The second matching degree scoring mechanism calculates the matching degree between each to-be-scheduled power grid task and its corresponding second candidate resource object according to a preset matching degree scoring function. The matching degree scoring functions of the first and second matching degree scoring mechanisms are different.

[0085] The matching degree scoring function of the second matching degree scoring mechanism can be as follows:

[0086] Matching degree = a · skill matching score + β · load margin score + γ · geographic proximity score + δ · response history score.

[0087] In this matching degree scoring function, a, β, γ and δ are weight values, the sum of which is 1, and their respective values can be adjusted according to actual conditions.

[0088] After calculating the matching degree of each to-be-scheduled power grid task and each second candidate resource object corresponding thereto, the second candidate resource object with the highest matching degree is taken as a target resource object. Each task agent will send the to-be-scheduled power grid task to the target resource object, and then update the scheduling graph. For example, the scheduling state of the related to-be-scheduled power grid task in the scheduling graph is changed to scheduled, so that the power grid task will not be scheduled.

[0089] S150: According to the compliance knowledge graph, the compliance of each task resource mapping information is reviewed.

[0090] The compliance knowledge graph is a subgraph of the power grid task semantic graph, and represents the relationship network between the power grid task and the rules and regulations. Its structure is: task type-constraint rule-approval path.

[0091] The construction process of the compliance knowledge graph can be as follows:

[0092] (1) Task type standardization and classification modeling: abstract and standardize the modeling of power grid tasks, such as “equipment maintenance-live work” and “metering device replacement-business task”, and construct a task type tree.

[0093] (2) Rules and regulations knowledge extraction: extract compliance information from enterprise safety procedures, operation rules and internal management systems, and use rule extraction or manual correction to organize structured triples.

[0094] (3) Compliance rule mapping modeling: mapping system provisions and task types one by one to form a “task type-must meet rules” mapping, such as operation personnel must have a certificate, voltage level restriction, and approval responsibility person restriction.

[0095] (4) Approval process path modeling: modeling the task approval path, such as “submitter→team leader approval→operation and maintenance responsible person approval→safety review” as a path chain.

[0096] (5) Graph fusion and knowledge storage: use a graph database (such as Neo4j) to build the above elements into a knowledge graph, nodes include tasks, rules, and approval nodes, and edges include “requirement”, “constraint” and “approval sequence” and other relationships.

[0097] After the task resource mapping information is generated, the task resource mapping information is matched and rule-inferred based on the compliance knowledge graph to conduct compliance review on each task resource mapping information. The steps of reviewing the compliance of each task resource mapping information according to the compliance knowledge graph can include: taking each task resource mapping information as a review object, performing graph matching on the review object in the compliance knowledge graph to obtain a compliance path corresponding to the review object; verifying whether the review object satisfies all logical rules of the compliance path using a rule-inference engine; if the review object satisfies all logical rules of the compliance path, determining that the review object passes the compliance review; if there is no compliance path corresponding to the review object in the compliance knowledge graph, or the review object does not satisfy any logical rule of the compliance path, determining that the review object fails to pass the compliance review.

[0098] The generated task resource mapping information can be regarded as a whole as a review object; the task resource mapping information as a whole is order assignment suggestion information, which includes task identification information (such as task ID, task type, associated work order number), order assignment personnel information (such as a recommended personnel list, qualification information, and a belonging team), time information (such as recommended start and end time, operation window), resource information (such as required tools, vehicles, and auxiliary tools), approval path information (such as a recommended approval node sequence and a list of responsible persons), and execution risk prompts (such as special weather and cross-operation reminders).

[0099] For the review object, all possible compliance paths thereof are searched in the compliance knowledge graph using a graph matching technique. For example, a task instance subgraph is constructed based on the order assignment suggestion information, and the task instance subgraph is structurally matched with a standard subgraph defined in the compliance knowledge graph. The matching can check whether the task instance subgraph satisfies all defined paths, nodes, and relationships of the standard subgraph. If all the defined paths, nodes, and relationships are satisfied, it is determined that the standard subgraph is a compliance path. If one of the defined paths, nodes, or relationships is not satisfied, it indicates a conflict. For example, an execution personnel node in the order assignment suggestion information does not have a “high-voltage operation qualification”, which cannot satisfy the definition of the standard subgraph, and thus the two do not match, indicating a conflict.

[0100] Then, the rule-inference engine is used to verify whether the review object satisfies all logical rules in the compliance path, and finally, a determination is made on whether the review object passes or fails the compliance review according to the verification result. When the review object fails to pass the compliance review, conflict details can also be output. For example, the logical rules in the compliance path (such as “must”, “cannot”, and “sequence”) are used for inference verification. Assuming that the logical rule is “task type A and no X qualification of personnel, then no order assignment is allowed”, and the task type in the order assignment suggestion information is B, it is determined that the review object does not satisfy this logical rule.

[0101] The logical rules in the compliance path include, but are not limited to, task qualification requirements, job safety specifications, business approval conditions, job coordination rules, legal regulations, and the like. Exemplarily, the task qualification requirements are, for example, (1) the execution personnel need to have the "high-voltage work" qualification, and (2) the team needs to have the "non-stop work" special authorization; the job safety specifications are, for example, (1) high-altitude work needs to be coordinated by two people, and a full-time guardian is arranged, and (2) additional lighting approval must be applied for night work; the business approval conditions are, for example, (1) the operation involving the substation must obtain the dispatch approval, and (2) the same day cannot be disconnected for the same feeder multiple times; and the job coordination rules are, for example, (1) certain types of maintenance tasks need to be synchronized with the delivery of safety tools, and (2) marketing installation tasks need to be linked with the construction team.

[0102] In some embodiments, after determining that the review object does not pass the compliance review, the method further includes: determining a conflict type between the review object and the compliance path; determining a disposal measure of the review object related to the to-be-scheduled power grid task according to the conflict type, and executing the disposal measure.

[0103] The conflict type can be of various types, for example, qualification conflict (the designated personnel do not have the work qualification, certificate or authorization required by the task), safety specification conflict (the operation time, operation method, cooperation mechanism, etc. violate the safety operation specification), approval path conflict (the approval process is incomplete, missing key links, bypassing approval), resource allocation conflict (the tools, vehicles, auxiliary equipment, etc. are not matched with the task or are occupied), time sequence logic conflict (the dispatch task and the preceding task, dependent task have time or state conflicts), and organization coordination conflict (there are task overlaps, cross-operations, unclear responsibilities between multiple organizations / teams).

[0104] The disposal measure can include scheduling rollback, resource replacement, process change, and the like. Different conflict types require different disposal measures. For example, when the conflict type is a qualification conflict, resource replacement is performed; when the conflict type is a safety specification conflict or an approval path conflict, the process is changed; and when the conflict type is a time sequence logic conflict, scheduling rollback is performed. Resource replacement refers to replacing the relevant resource object with another resource object, and then re-executing the steps S130 and the following steps. Regarding the process change, taking the approval path conflict as an example, it can be to analyze the missing approval node, dynamically insert the missing approval node into the approval task to the workflow, and then notify the relevant approver, and after the approval is completed, the dispatch can continue. The operation of scheduling rollback can be to locate the preceding task or dependent task of the current to-be-scheduled power grid task, and then calculate the earliest start time, and based on moving or splitting the time window of the current to-be-scheduled power grid task.

[0105] S160: performing task dispatching according to the task resource mapping information passing the compliance review.

[0106] For the task resource mapping information passed through the compliance review, the server can generate a task execution path through a BPM (Business Process Management) platform or an internal process engine originally of the power grid enterprise. The task execution path includes a trigger node, task distribution, execution receipt, stage acceptance, exception report, achievement archiving, etc., each node being bound with a timestamp and an executor ID.

[0107] In some embodiments, after task assignment is performed according to the task resource mapping information passed through the compliance review, the method further includes: collecting multi-dimensional execution records of each power grid task after execution; dynamically adjusting the policy parameters in the second matching degree scoring mechanism through a reinforcement learning reward function based on the multi-dimensional execution records, and / or screening high-quality execution records meeting preset quality requirements from the multi-dimensional execution records of the plurality of power grid tasks, and updating the power grid task semantic graph based on the high-quality execution records.

[0108] The server can continuously collect multi-dimensional execution records associated with each power grid task after execution, which include but are not limited to task completion timeliness, feedback quality, user evaluation, whether to repair, resource response delay, etc. Based on the collected multi-dimensional execution records, the policy parameters (i.e., the values of a, b, g and d) in the second matching degree scoring mechanism can be dynamically adjusted through a reinforcement learning reward function. After execution of the power grid task, the scheduling effect can be dynamically evaluated according to the multi-dimensional execution records, and the above policy parameters can be fine-tuned through a reinforcement learning method, so as to realize optimization of the task assignment strategy, thereby realizing continuous evolution and intelligent improvement of the scheduling system in long-term operation.

[0109] In addition, high-quality execution records meeting preset quality requirements (which can include requirements such as data integrity and data accuracy) can be screened from the multi-dimensional execution records of the plurality of power grid tasks, and the high-quality task execution records can be embedded in the power grid task semantic graph to form case nodes and be bound with the task type relationship.

[0110] It should be noted that, as for each step included in the power grid task self-matching and dynamic scheduling method provided in any one of the above embodiments and the large language model and multi-agent cooperation, unless otherwise specified herein, the execution of these steps does not have strict order restrictions, and these steps can be executed in other orders. Moreover, at least part of these steps can include multiple sub-steps or multiple stages, which do not necessarily be executed at the same time, but can be executed at different times, and the execution order of these sub-steps or stages is not necessarily sequential, but can be executed in rotation or alternation with at least part of other steps or sub-steps or stages of other steps.

[0111] Based on the same inventive concept, the present application also provides a power grid task self-matching and dynamic scheduling device based on cooperation of a large language model and multiple agents. In some embodiments, as shown in Figure 2 The device includes the following modules:

[0112] The power grid task construction module 110 is configured to collect task description information, pre-process the task description information, perform semantic analysis and structured modeling on the pre-processed task description information using a power grid fine-tuning large model, obtain initial structured task information, perform semantic calibration on the initial structured task information based on a power grid task semantic graph, obtain standard structured task information, generate a power grid task based on the standard structured task information, and add the power grid task to a task pool.

[0113] The resource object construction module 120 is configured to collect multi-dimensional resource object information, perform multi-dimensional vector modeling based on the multi-dimensional resource object information, obtain resource objects represented by multi-dimensional vectors, and add the resource objects to a resource pool.

[0114] The first resource object screening module 130 is configured to calculate the matching degree between each to-be-scheduled power grid task in the task pool and each resource object in the resource pool based on a first matching degree scoring mechanism, and determine a plurality of first candidate resource objects corresponding to each to-be-scheduled power grid task based on the matching degree calculation results.

[0115] The second resource object screening module 140 is configured to assign a task agent to each to-be-scheduled power grid task, assign a resource agent to each first candidate resource object, and enable the task agents and the resource agents to perform collaborative scheduling in a multi-agent reinforcement learning environment to determine a target resource object corresponding to each to-be-scheduled power grid task and generate a plurality of task-resource mapping information.

[0116] The compliance review module 150 is configured to perform compliance review on each piece of task-resource mapping information according to a compliance knowledge graph.

[0117] The task assignment module 160 is configured to perform task assignment according to the task-resource mapping information that passes the compliance review.

[0118] In some embodiments, the power grid task semantic graph includes a plurality of standard task nodes; and the power grid task construction module 110 includes:

[0119] The task node determination sub-module is configured to determine a target task node from the plurality of standard task nodes according to the initial structured task information.

[0120] The task information matching sub-module is configured to calculate the matching degree between the initial structured task information and the node information of the target task node using a graph neural network.

[0121] The task information acquisition submodule is configured to acquire node information of the target task node as the standard structured task information when the matching degree is greater than the preset threshold.

[0122] The return submodule is configured to return to the step of performing semantic analysis and structured modeling on the preprocessed task description information by using the power grid fine-tuning large model when the matching degree is less than or equal to the preset threshold.

[0123] In some embodiments, the first resource object screening module 130 includes:

[0124] The embedding submodule is configured to embed the power grid task into a vector space to obtain a task vector.

[0125] The multi-dimensional score calculation submodule is configured to calculate a skill matching score, a response capability score, and a success rate score of the task vector and each resource object.

[0126] The first matching degree calculation submodule is configured to calculate a matching degree of the task vector and each resource object according to the skill matching score, the response capability score, and the success rate score of the task vector and each resource object.

[0127] In some embodiments, the compliance review module 150 includes:

[0128] The graph matching submodule is configured to map each piece of task resource mapping information as a review object, and perform graph matching on the review object in the compliance knowledge graph to obtain a compliance path corresponding to the review object.

[0129] The rule reasoning submodule is configured to verify, by using a rule reasoning engine, whether the review object satisfies all logical rules of the compliance path.

[0130] The review result determination submodule is configured to determine that the review object passes the compliance review if the review object satisfies all logical rules of the compliance path, and determine that the review object fails the compliance review if there is no compliance path corresponding to the review object in the compliance knowledge graph or the review object does not satisfy any logical rule of the compliance path.

[0131] In some embodiments, the device further includes a disposal module. The disposal module is configured to determine a conflict type between the review object and the compliance path after determining that the review object fails the compliance review, determine a disposal measure of a power grid task related to the review object according to the conflict type, and execute the disposal measure.

[0132] In some embodiments, the second resource object screening module 140 includes:

[0133] The scheduling graph construction submodule is configured to construct a scheduling graph; the scheduling nodes in the scheduling graph are each to-be-scheduled power grid task and each first candidate resource object, and the edges between any two scheduling nodes represent one or more relationships between the two scheduling nodes; the one or more relationships include an organizational constraint relationship;

[0134] The task agent control submodule is configured to control each task agent to determine a target resource agent that needs to be docked according to the scheduling graph, and to send a task matching request to all target resource agents in parallel;

[0135] The resource agent control submodule is configured to control each resource agent to, after receiving the task matching request, perform benefit evaluation based on the task information carried by the task matching request, and return response information indicating acceptance of the task or rejection of the task to the task agent that sent the task matching request based on the benefit evaluation result;

[0136] The resource object screening submodule is configured to determine, based on the response information of each resource task body, a second candidate resource object corresponding to each to-be-scheduled power grid task;

[0137] The second matching degree calculation submodule is configured to calculate, based on a second matching degree scoring mechanism, a matching degree of each to-be-scheduled power grid task and each second candidate resource object corresponding thereto;

[0138] The target resource object determination submodule is configured to determine, according to the matching degree of each to-be-scheduled power grid task and each second candidate resource object corresponding thereto, a target resource object corresponding to each to-be-scheduled power grid task.

[0139] In some embodiments, the apparatus further includes:

[0140] The task execution record collection module is configured to collect a multi-dimensional execution record of each power grid task after execution; and

[0141] The policy parameter adjustment module is configured to dynamically adjust a policy parameter in the second matching degree scoring mechanism based on the multi-dimensional execution record through a reinforcement learning reward function, and / or the graph updating module is configured to screen high-quality execution records meeting a preset quality requirement from the multi-dimensional execution records of the plurality of power grid tasks, and update the power grid task semantic graph based on the high-quality execution records.

[0142] The specific limitations of the power grid task self-matching and dynamic scheduling device based on the large language model and multi-agent collaboration can refer to the limitations of the power grid task self-matching and dynamic scheduling method based on the large language model and multi-agent collaboration described above, and will not be repeated here. Each module in the above power grid task self-matching and dynamic scheduling device based on the large language model and multi-agent collaboration can be realized by software, hardware, and combinations thereof, in whole or in part. The above modules can be embedded in or independent of the processor in the computer device in hardware form, or can be stored in the memory in the computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0143] The present application also provides a computer device, which in some embodiments includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, it can implement the steps of the power grid task self-matching and dynamic scheduling method based on the large language model and multi-agent collaboration provided in any of the above embodiments.

[0144] In some embodiments, the internal structure diagram of the computer device can be as shown in Figure 3 The computer device includes a processor, a memory, and a network interface connected by a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program, and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store power grid tasks, resource objects, and other data. The specific data stored can also refer to the limitations in the above method embodiments. The network interface of the computer device is used to communicate with external terminals through network connections. The computer program is executed by the processor to implement a power grid task self-matching and dynamic scheduling method based on the large language model and multi-agent collaboration.

[0145] Those skilled in the art can understand, Figure 3 The structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0146] The present application also provides a computer readable storage medium, which in some embodiments has a computer program stored thereon. When the computer program is executed by the processor, it implements the steps of the power grid task self-matching and dynamic scheduling method based on the large language model and multi-agent collaboration provided in any of the above embodiments.

[0147] In the above-described embodiments of the present application, the description of each embodiment has its own focus, and the parts not described in detail in a certain embodiment can be referred to the relevant description of other embodiments.

[0148] Those skilled in the art can understand that all or part of the processes in the above-mentioned method embodiments can be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer readable storage medium, and when the computer program is executed, the processes of the above-mentioned embodiments of the method can be included. Any reference to memory, storage, database or other medium used in each embodiment provided by the present application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink), DRAM (SLDRAM), memory bus (Rambus), direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0149] Any combination of the technical features of the above embodiments can be made. In order to make the description simple, all possible combinations of the technical features in the above-described embodiments have not been described, however, as long as the combination of the technical features does not exist, it should be considered as the scope of the present application.

[0150] The above-described embodiments only express several embodiments of the present application, and the description is more specific and detailed, but it should not be understood as a limitation on the scope of the patent. It should be noted that for ordinary skilled in the art, without departing from the concept of the present application, a number of modifications and improvements can be made, which are within the scope of the present application. Therefore, the scope of the patent of the present application should be subject to the appended claims.

Claims

1. A method for power grid task self-matching and dynamic scheduling of large language models and multi-agent collaboration, characterized in that, The method comprises: collecting task description information, preprocessing the task description information, using a power grid fine-tuning large model to perform semantic analysis and structured modeling on the preprocessed task description information to obtain initial structured task information, performing semantic calibration on the initial structured task information based on a power grid task semantic graph to obtain standard structured task information, generating a power grid task based on the standard structured task information, and adding the power grid task to a task pool; collecting multi-dimensional resource object information, performing multi-dimensional vector modeling based on the multi-dimensional resource object information to obtain resource objects represented by multi-dimensional vectors, and adding the resource objects to a resource pool; based on a first matching degree scoring mechanism, calculating the matching degree of each to-be-scheduled power grid task in the task pool and each resource object in the resource pool, and determining a plurality of first candidate resource objects corresponding to each to-be-scheduled power grid task based on the matching degree calculation result; allocating a task intelligent agent to each to-be-scheduled power grid task, allocating a resource intelligent agent to each first candidate resource object, and enabling each task intelligent agent and each resource intelligent agent to perform collaborative scheduling in a multi-agent reinforcement learning environment to determine a target resource object corresponding to each to-be-scheduled power grid task and generate a plurality of task resource mapping information; performing compliance review on each piece of task resource mapping information according to a compliance knowledge graph; performing task assignment according to the task resource mapping information that passes the compliance review; the enabling each task intelligent agent and each resource intelligent agent to perform collaborative scheduling in a multi-agent reinforcement learning environment to determine a target resource object corresponding to each to-be-scheduled power grid task comprises: constructing a scheduling graph; the scheduling nodes in the scheduling graph are each to-be-scheduled power grid task and each first candidate resource object, and the edges between any two scheduling nodes represent one or more relationships between the two scheduling nodes; the one or more relationships include an organizational constraint relationship; each task intelligent agent determines a target resource intelligent agent that needs to be interfaced according to the scheduling graph, and sends a task matching request to all target resource intelligent agents in parallel; after each resource intelligent agent receives a task matching request, it performs benefit evaluation based on the task information carried by the task matching request, and returns response information indicating acceptance of the task or rejection of the task to the task intelligent agent that sent the task matching request based on the benefit evaluation result; determining a second candidate resource object corresponding to each to-be-scheduled power grid task based on the response information of each resource intelligent agent; based on a second matching degree scoring mechanism, calculating the matching degree of each to-be-scheduled power grid task and each second candidate resource object corresponding thereto; determining a target resource object corresponding to each to-be-scheduled power grid task according to the matching degree of each to-be-scheduled power grid task and each second candidate resource object corresponding thereto.

2. The method of claim 1, wherein, The power grid task semantic graph comprises a plurality of standard task nodes; the performing semantic calibration on the initial structured task information based on the power grid task semantic graph to obtain standard structured task information comprises: determining a target task node from the plurality of standard task nodes according to the initial structured task information; calculating a matching degree between the initial structured task information and node information of the target task node using a graph neural network; when the matching degree is greater than a preset threshold, obtaining the node information of the target task node as standard structured task information; when the matching degree is less than or equal to the preset threshold, returning to the step of using the power grid fine-tuning large model to perform semantic analysis and structured modeling on the preprocessed task description information.

3. The method of claim 1, wherein, The calculating the matching degree of each to-be-scheduled power grid task in the task pool and each resource object in the resource pool includes: embedding the power grid task into a vector space to obtain a task vector; calculating a skill matching score, a response capability score, and a success rate score of the task vector and each resource object; calculating the matching degree of the task vector and each resource object according to the skill matching score, the response capability score, and the success rate score of the task vector and each resource object.

4. The method of claim 1, wherein, The compliance review of each task resource mapping information according to the compliance knowledge graph includes: taking each task resource mapping information as a review object, performing graph matching on the review object in the compliance knowledge graph to obtain a compliance path corresponding to the review object; verifying whether the review object satisfies all logical rules of the compliance path using a rule reasoning engine; if the review object satisfies all logical rules of the compliance path, determining that the review object passes the compliance review; if there is no compliance path corresponding to the review object in the compliance knowledge graph, or the review object does not satisfy any logical rule of the compliance path, determining that the review object fails the compliance review.

5. The method of claim 4, wherein, After determining that the review object fails the compliance review, the method further includes: determining a conflict type between the review object and the compliance path; determining a disposal measure of a to-be-scheduled power grid task related to the review object according to the conflict type, and executing the disposal measure.

6. The method of claim 1, wherein, After the task dispatching according to the task resource mapping information that passes the compliance review, the method further includes: collecting multi-dimensional execution records of each power grid task after execution; based on the multi-dimensional execution records, dynamically adjusting a policy parameter in the second matching degree scoring mechanism through a reinforcement learning reward function, and / or, screening high-quality execution records meeting a preset quality requirement from multi-dimensional execution records of a plurality of power grid tasks, and updating the power grid task semantic graph based on the high-quality execution records.

7. A power grid task self-matching and dynamic scheduling device for large language models and multi-agent collaboration, characterized in that, The device includes: The power grid task construction module is configured to collect task description information, pre-process the task description information, perform semantic analysis and structured modeling on the pre-processed task description information using a power grid fine-tuning large model to obtain initial structured task information, perform semantic calibration on the initial structured task information based on a power grid task semantic graph to obtain standard structured task information, generate a power grid task based on the standard structured task information, and add the power grid task to a task pool; The resource object construction module is configured to collect multi-dimensional resource object information, perform multi-dimensional vector modeling based on the multi-dimensional resource object information to obtain resource objects represented by multi-dimensional vectors, and add the resource objects to a resource pool; The first resource object screening module is configured to calculate, based on a first matching degree scoring mechanism, a matching degree of each to-be-scheduled power grid task in the task pool and each resource object in the resource pool, and determine, based on a matching degree calculation result, a plurality of first candidate resource objects corresponding to each to-be-scheduled power grid task; The second resource object screening module is configured to assign a task agent to each to-be-scheduled power grid task, assign a resource agent to each first candidate resource object, and enable each task agent and each resource agent to perform collaborative scheduling in a multi-agent reinforcement learning environment to determine a target resource object corresponding to each to-be-scheduled power grid task and generate a plurality of task-resource mapping information; The compliance review module is configured to perform compliance review on each piece of task-resource mapping information according to a compliance knowledge graph; The task dispatching module is configured to perform task dispatching according to the task-resource mapping information that passes the compliance review; The second resource object screening module includes: The scheduling graph construction submodule is configured to construct a scheduling graph; scheduling nodes in the scheduling graph are each to-be-scheduled power grid task and each first candidate resource object, and an edge between any two scheduling nodes represents one or more relationships between the two scheduling nodes; the one or more relationships include an organizational constraint relationship; The task agent control submodule is configured to control each task agent to determine, according to the scheduling graph, a target resource agent that needs to be interfaced, and to send, in parallel, a task matching request to all target resource agents; The resource agent control submodule is configured to control each resource agent, after receiving a task matching request, to perform benefit evaluation based on task information carried by the task matching request, and to return, to the task agent that sent the task matching request, response information indicating acceptance of the task or rejection of the task based on a benefit evaluation result; The resource object screening submodule is configured to determine, based on response information of each resource agent, a second candidate resource object corresponding to each to-be-scheduled power grid task; The second matching degree calculation submodule is configured to calculate, based on a second matching degree scoring mechanism, a matching degree of each to-be-scheduled power grid task and each second candidate resource object corresponding to the to-be-scheduled power grid task; The target resource object determination submodule is configured to determine, according to the matching degree of each to-be-scheduled power grid task and each second candidate resource object corresponding to the to-be-scheduled power grid task, a target resource object corresponding to each to-be-scheduled power grid task.

8. A computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The computer program is executed by the processor to implement the steps of the method of any one of claims 1 to 6.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the method of any one of claims 1 to 6.

Images