Large and small model collaborative multi-agent operation method and system

Abstract

The invention relates to the technical field of artificial intelligence, and provides a large and small model collaborative multi-agent operation method and system, and the method comprises the steps: analyzing a power task description through a large model agent, dividing the power task description into a plurality of activities, extracting the execution parameters of each activity, and determining a small model agent corresponding to each activity according to a large and small model collaborative relationship; based on the execution parameters, combining available resources and load states of the execution units, taking income and cost of task allocation as objective functions, and through a particle swarm optimization algorithm based on a non-dominated sorting mechanism, allocating the activities to different execution units, and processing the activities by a small model agent, the small model intelligent agent calls the small model through the interface to complete the corresponding activity; when the particle swarm optimization detects that resources are insufficient or tasks are distributed unevenly, the utilization rates are compared, and if deviation exceeds a threshold value, the tasks are reordered according to the priorities of the tasks. And the execution reliability of business processes such as fault diagnosis and the like is improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of artificial intelligence, and particularly relates to a multi-agent job method and system based on large and small model cooperation. BACKGROUND

[0002] The statements in this section merely provide background information related to the present application and do not necessarily constitute prior art.

[0003] With the continuous expansion of the scale of the power distribution network and the gradual improvement of the automation level, the power distribution network dispatching system plays an increasingly key role in daily operation and fault analysis. Especially in the complex scenario of sudden failure and cross-regional linkage, dispatchers need to quickly and accurately determine the fault point and take appropriate measures to ensure power supply safety and stability. However, the traditional fault analysis mainly relies on manual experience and manual dispatching operation, which is not only low in efficiency and high in cost, but also prone to cause the expansion of the fault range and processing errors due to information misjudgment or coordination delay.

[0004] In recent years, with the development of artificial intelligence technology, especially large language models (such as ChatGPT, LLaMA, etc.) and agent (Agent) frameworks, their application in power distribution dispatching has gradually expanded, and some units have tried to use pre-trained large models to directly understand and analyze multi-modal data such as telemetry and remote signaling, hoping to use the powerful understanding ability of large models to complete fault positioning and intelligent decision-making. However, this method has exposed many problems in actual deployment: Firstly, the current large models generally lack the ability to adapt to the professional knowledge and structured telemetry data of the power distribution system. Since the power distribution dispatching field accounts for a very small proportion in the pre-training data of large models, it is easy for large models to produce "fact illusion" or "loyalty illusion" when processing professional semantics, resulting in biased analysis conclusions; at the same time, large models still have significant bottlenecks in the unified processing of image, unstructured text, table and other multi-modal data, making it difficult to realize cross-modal information fusion and efficient reasoning.

[0005] Secondly, large models have problems such as unstable output and poor consistency when processing dynamic process tasks. The fault handling process in power distribution dispatching usually has high dynamicity and task heterogeneity, and it is difficult for a single large model to cover the intelligent needs of all process nodes. Some researches try to divide the task by forming a pipeline process with multiple large models, but due to the lack of unified process semantic modeling, the information between large models is fragmented, and errors are easily accumulated, ultimately leading to the failure of the overall analysis of the system.

[0006] In addition, in complex business scenarios, the distribution network scheduling process often spans multiple operation and maintenance units, management sections or stations. Due to the diversity of telemetry data sources and the complexity of the transmission chain, the authenticity of the data is difficult to verify, and there is currently a lack of effective mechanisms to ensure the reliable flow and consistent execution of multi-modal data among multiple agents. Once part of the model or node is tampered with, misjudged or fails, it is easy to cause global judgment errors, seriously threatening the safe and stable operation of the distribution network.

[0007] To solve the above technical problems, the traditional method mainly assists in research and judgment by deploying multiple small models or rule engines, but such methods have weak universality, high maintenance cost, and cannot realize intelligent collaboration in the whole process. SUMMARY

[0008] To solve the technical problems in the background art, the present application provides a multi-agent job method and system of large and small model cooperation, which realizes efficient fusion and unified processing of multi-modal data through the global semantic understanding, task planning and process scheduling capability of large models, combined with the accurate processing capability of small models in specific modalities or professional tasks, thereby effectively reducing the misjudgment that may be caused by a single model and improving the reliability of business process execution such as fault diagnosis.

[0009] To achieve the above purpose, the present application adopts the following technical solutions: The first aspect of the present application provides a multi-agent job method of large and small model cooperation, which comprises: obtaining a power task description; analyzing the power task description by a large model agent, dividing it into multiple activities, extracting the execution parameters of each activity, and determining the small model agent corresponding to each activity according to the cooperation relationship between large and small models; Based on the execution parameters, combining the available resources and load state of each execution unit, taking the benefit and cost of task allocation as the objective function, through the particle swarm algorithm based on the non-dominated sorting mechanism, each activity is allocated to different execution units for processing by the small model agent, and the small model agent calls the small model through the interface to complete the corresponding activity; wherein, when the particle swarm algorithm based on the non-dominated sorting mechanism detects resource shortage or uneven task allocation, the real-time utilization rate of each execution unit is calculated and compared with the average utilization rate, if the deviation exceeds the threshold, the low-priority tasks are migrated from the high-load execution unit to the low-load execution unit according to the priority.

[0010] Further, in the migration process, a greedy strategy is adopted to preferentially select the activity with the smallest migration cost and the largest reduction in imbalance for adjustment until the load is balanced.

[0011] Further, the migration cost includes activity waiting time, data transmission cost, restart warm-up time and rollback loss.

[0012] Further, the execution parameters include time limit and resource requirement.

[0013] Further, the large model agent includes a large model body, an MCP primitive and a standardized interface contained in the primitive.

[0014] Further, the particle swarm optimization algorithm based on the non-dominated sorting mechanism adopts a dynamic inertia weight decreasing with the number of iterations for particle speed updating.

[0015] Further, the large-small model cooperation relationship is represented as: ; wherein Agent i represents the i-th large model corresponding agent, lm i represents the i-th large model body, SM fa (lm i ) represents the i-th large model callable small model set, MCP pr (sm is ) and pr w () respectively represent the i-th large model callable s-th small model callable MCP primitive and corresponding standardized interface set.

[0016] The second aspect of the present application provides a large-small model collaborative multi-agent job system, which comprises: a task acquisition module configured to acquire a power task description; a task analysis module configured to analyze the power task description by a large model agent, divide it into multiple activities, extract the execution parameters of each activity, and determine the small model agent corresponding to each activity according to the large-small model cooperation relationship; a process execution module configured to, based on the execution parameters, combine the available resources and load states of each execution unit, take the benefits and costs of task allocation as the objective function, and through a particle swarm optimization algorithm based on the non-dominated sorting mechanism, allocate each activity to different execution units for processing by the small model agent, and the small model agent completes the corresponding activity by calling the small model through the interface; wherein when the particle swarm optimization algorithm based on the non-dominated sorting mechanism detects resource shortage or uneven task allocation, it calculates the real-time utilization rate of each execution unit and compares it with the average utilization rate, if the deviation exceeds the threshold, the low-priority tasks are migrated from the high-load execution unit to the low-load execution unit according to the priority.

[0017] The third aspect of the present application provides a computer readable storage medium having a computer program stored thereon, the program being executed by a processor to implement the steps of the above-mentioned large-small model collaborative multi-agent job method.

[0018] The fourth aspect of the present application provides a computer device comprising a computer readable storage medium, a processor and a computer program stored on the computer readable storage medium and executable on the processor, wherein the processor implements the steps of the multi-agent job method of the size model coordination when executing the program.

[0019] Compared with the prior art, the beneficial effects of the present application are: The present application innovatively provides a multi-agent job method of size model coordination, which realizes efficient fusion and unified processing of multi-modal data through the global semantic understanding, task planning and process scheduling capability of the large model, combined with the accurate processing capability of the small model in a specific modal or professional task, thereby effectively reducing the misjudgment that may be caused by a single model and improving the reliability of business process execution such as fault diagnosis. BRIEF DESCRIPTION OF DRAWINGS

[0020] The drawings accompanying the specification of the present application form a part thereof and serve to provide further understanding of the present application, the illustrative embodiments of the present application and its description serve to explain the present application, and do not constitute an improper limitation of the present application.

[0021] Figure 1 is a multi-modal agent cooperative distribution network scheduling fault research and judgment business process schematic diagram of embodiment one of the present application; Figure 2 is a size model cooperative distribution network scheduling fault research and judgment sub-business process schematic diagram of embodiment one of the present application; Figure 3 is a structure schematic diagram of a computer device of embodiment four of the present application. DETAILED DESCRIPTION

[0022] In order to make the purpose, technical scheme and advantages of the embodiments of the present application more clear, the technical scheme 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.

[0023] It should be pointed out that the following detailed description is exemplary and is intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as generally understood by those skilled in the art to which the present application belongs.

[0024] Embodiment one The present embodiment provides a multi-agent job method of size model coordination.

[0025] The multi-agent job method of size model coordination provided in the present embodiment can integrate the general inference capability of the large model and the professional precision advantage of the small model, and support dynamic, complex and multi-modal intelligent jobs.

[0026] Among them, the large model can be a general large language model (such as ChatGPT, GPT-4 / 4o, LLaMA, GLM) or a multi-modal large model (such as GPT-4V, Claude Vision, Kosmos-1), and in the power industry application, it can also be an industry-specific large model fine-tuned by power dispatching data. These large models mainly undertake tasks such as semantic analysis, global scheduling, and cross-modal reasoning.

[0027] Among them, the small model can be an image recognition model (such as YOLO, Mask R-CNN, used for fault equipment identification and smoke detection), a time series prediction model (such as LSTM, GRU, used for telemetry data trend analysis and anomaly detection), and a speech recognition model (such as Whisper, DeepSpeech, used for dispatch voice transcription). These small models have higher accuracy and efficiency in specific modal tasks.

[0028] The size model collaborative multi-agent job method provided in this embodiment combines the professional capabilities of expert small models and the task scheduling and semantic modeling advantages of large models. By deploying small model combinations driven by large models at each business critical node, a business agent (AI Agent) with dynamic reorganization, multi-modal input support, and process awareness is constructed. Further combined with agent trajectory analysis, intelligent collaboration and process reconstruction between multiple nodes are achieved, significantly improving the accuracy, stability, and transparency of fault analysis, and having good engineering implementability and industry application prospects.

[0029] Traditional distribution network dispatching systems mainly rely on manual experience and rule-based systems for analysis and judgment when facing sudden failures and complex environments, but these methods often cannot effectively integrate multi-modal data, and the output results lack consistency and stability, especially in a multi-source heterogeneous data environment, facing huge challenges in data authenticity, processing capacity, and intelligent decision-making.

[0030] The size model collaborative multi-agent job method provided in this embodiment introduces a collaborative mechanism between large models and small models, combining the powerful semantic understanding and task scheduling capabilities of large language models, and combining the accurate processing capabilities of expert small models in specific modalities, effectively addressing the shortcomings of existing technologies.

[0031] The embodiment provides a multi-agent job method of size model cooperation, which optimizes a business process, designs a multi-agent architecture based on size model cooperation, so that each agent can realize efficient cooperation in a multi-modal data input environment; the large model undertakes the planning, semantic analysis and global scheduling responsibilities of the overall task in the architecture, is responsible for processing high-level reasoning, decision-making and task allocation; at the same time, the small model is deployed at the key node of the business process and focuses on processing specific modalities or professional tasks, such as real-time analysis of telemetry data, identification of image data or processing of unstructured text; each small model independently completes its task according to the scheduling and requirements of the large model, and exchanges data and feeds back results with other agents through a standardized interface.

[0032] The multi-agent job method of size model cooperation provided by the embodiment has the core advantages that through accurate cooperation and data flow between agents, local intelligence and global control are organically combined, multiple agents construct a cooperation process through a standardized interface, and the flexibility and scalability of the business process are ensured, so that the execution strategy and task allocation of the agent can be dynamically adjusted according to real-time business requirements, so as to adapt to different business scenarios and fault handling requirements.

[0033] The multi-agent job method of size model cooperation provided by the embodiment also supports unified processing of multi-modal data, including but not limited to telemetry data, image data, unstructured text data and table data, greatly improving the efficiency and accuracy of data processing.

[0034] In the embodiment, a fault research and judgment task is regarded as a complex business process, and the process is composed of multiple links / nodes, which are called activities in this embodiment, and are commonly referred to as steps, for example, the process of the fault research and judgment task is: telemetry data acquisition→current / voltage curve analysis→image recognition equipment state→log information extraction→final comprehensive research and judgment, and each link here is an activity.

[0035] Different activities are executed by different agents (driven by large models or small models), and finally the whole task is completed through cooperation.

[0036] In the job, an activity is a processing node, which completes a specific subtask. An activity corresponds to an agent, and corresponds to a large model (core) + interface.

[0037] The small model is not an independent agent, but a “special tool” called by the large model, which interacts with the large model or the agent through a standardized interface to complete a specific modal task (such as telemetry analysis, image recognition), and the small model can be called by a single agent or reused in multiple sub-processes; One activity can contain multiple primitives / interface calls, which together constitute a "sub-business process". A sub-business process is a detailed execution step of an activity.

[0038] Specific examples are as follows: In the voltage sag alarm scenario, the task input is: the dispatching system generates an alarm, and the input content is "to analyze the voltage sag in the ×× area, and give the reason and treatment suggestion".

[0039] The activity is divided into: the master large model agent A analyzes the task and divides it into multiple activities: data integrity check, anomaly detection (voltage and current curve), image recognition (device status), log extraction (alarm text), result fusion and decision.

[0040] Activity execution is: Activity 1: Agent B calls the interface to access the database to check the telemetry data integrity, and outputs the data report; Activity 2: Agent C schedules the small model (LSTM / GRU) to analyze the current and voltage curve, and outputs the abnormal point and fluctuation characteristics.

[0041] Activity 3: Agent D calls the small model (YOLO) to identify the monitoring image, and outputs the device status (such as whether it smokes or sparks).

[0042] Activity 4: Agent E calls the small model (BERT / ERNIE) to extract the key information in the alarm log, and outputs the error code, timestamp, and operation record.

[0043] Activity 5: The master large model A summarizes all the outputs, performs cross-modal reasoning, and forms the final analysis conclusion and disposal suggestion.

[0044] A total fault analysis task is decomposed into multiple activities by a large model, each activity is processed by an agent, the agent calls a small model through an interface to complete a specific task, and finally a large model is used for global integration and decision. Each activity can be further refined into a series of interface calls, which are called sub-business processes.

[0045] The multi-agent job method provided by the embodiment of the large and small model cooperation comprises the following steps: Step 1, define the multi-modal agent business process execution (job) element.

[0046] In order to realize the multi-modal intelligent processing of distribution network dispatching fault analysis, a unified job method is proposed, which models the core elements involved in the business process in detail. This method combines multiple large models and small models to optimize the processing efficiency and accuracy of the overall fault analysis task through multi-agent collaboration. For example Figure 1The definitions of the key elements and their specific functions are as follows: (1) Large model set. In the process of distribution network dispatching fault research and judgment, multiple large models of different modalities are involved to process information obtained from multiple data sources (such as telemetry data, image data, and text data). The complete large model set in the fault research and judgment process is represented as: ; where i represents the large model number corresponding to the i-th activity in the fault research and judgment business process, and n represents the total number of activities in the fault research and judgment task process, LM fa represents the large model set corresponding to all activities in the fault research and judgment business, lm i represents the i-th large model; the large model is responsible for semantic analysis, global scheduling, and task planning in the entire fault research and judgment task process, and provides task allocation and instruction guidance for small models.

[0047] (2) MCP primitive set. In the fault research and judgment process, each activity corresponds to an agent, which interacts with the outside through MCP primitives to complete the corresponding activity. This interaction process is considered as a sub-business process associated with the activity. Each sub-business process (i.e., activity) contains a set of MCP primitives, which define the basic operations and communication protocols of task execution. The primitive set is represented as: ; where pr w represents a specific MCP primitive, w represents the primitive type, including data sampling (Sampling), resource calling (Resource), tool calling (Tool), and prompt generation (Prompt); each primitive corresponds to different tasks and data processing methods.

[0048] (3) Interface set. Each MCP primitive contains multiple standardized interfaces for interaction with external systems and other agents. The interface set is represented as: ; where API j represents a specific standardized interface associated with the primitive, which directly supports the interaction between the large model and the outside world, such as data acquisition, database query, algorithm calling, image sampling, and specified prompt acquisition, p represents the number of standardized interfaces included in the primitive; the calling of multiple interfaces can be mapped to multiple activities in the sub-business process, and the calling of multiple primitives can be mapped to multiple parallel activities in the sub-business process, thereby realizing the unified processing and coordination of multi-modal data.

[0049] (3) Agent. The large model with external interaction capability is regarded as an independent agent, and each agent is composed of a large model entity, an MCP primitive, and a standardized interface contained in the primitive, which is specifically represented as: ; In practical applications, the MCP primitive and the standardized interface constitute the MCP Server in the MCP protocol, and the large model constitutes the MCP Host in the actual deployment container or physical host. The task of each agent is to perform fault diagnosis tasks according to the MCP protocol, while cooperating with other agents to complete cross-modal data processing and task reasoning.

[0050] Through the definition of such an agent and the standardized design of the MCP interface, seamless collaboration of multi-modal agents is achieved, making the entire distribution network scheduling fault diagnosis system flexible to adapt to different types of tasks and scenarios, further improving the accuracy and efficiency of fault diagnosis.

[0051] Step 2, build a multi-modal agent collaborative fault diagnosis business process.

[0052] In order to solve the problem of multi-modal data processing in the fault diagnosis process of distribution network scheduling, a fault diagnosis business process architecture based on multi-modal agent collaboration is proposed. This architecture, through the cooperation between multiple agents, combines the global scheduling of large models and the specialized task processing capabilities of small models, aims to achieve efficient and accurate fault diagnosis and effectively improve the intelligent level of the entire business process. The specific business process includes the following main parts: Main business process. The main business process is completed by the cooperation of multiple large models. Each large model interacts with other models and system modules through MCP when processing business activities. The task allocation of large models is coordinated by MCP Server nodes. Large models parse input task instructions and data through MCP interfaces. When large models receive instructions or new business tasks from external users, they will first parse the task content, extract the execution parameters such as completion time limit, complexity, and resource requirements, and submit the sorted task information to the MCP-Server node (Model Context Protocol Server). MCP-Server nodes will combine the available resources and load status of each execution unit to build a benefit and cost model for activity allocation. In order to reasonably evaluate different task allocation schemes, MCP-Server nodes introduce a benefit function and a cost function to reflect the task completion efficiency and resource utilization balance, respectively, and weigh them through a unified multi-objective optimization model.

[0053] For example, the master large model (Agent A) splits the task and allocates activities: Activity 1: Data inspection -> to Agent B (large model); Activity 2: Anomaly detection -> to Agent C (large model); After receiving the "anomaly detection" activity, Agent C further splits: Subtask 1: Analyze voltage curve with LSTM small model; Subtask 2: Analyze switch status with XGBoost small model; Small models execute tasks respectively, output results -> aggregate back to Agent C's large model -> pass to master Agent A.

[0054] Among them, the reward function reflects the task completion efficiency, calculated as follows: ; Where N represents the number of activities, w i is the priority weight of activity i, T i is the activity completion time; the greater the reward value, the higher the overall completion efficiency.

[0055] The cost function reflects the resource utilization balance, calculated as follows: ; Where, M represents the number of execution units, U j is the actual utilization rate of the jth execution unit, is the average utilization rate of all execution units; the smaller the cost value, the more balanced the resource allocation.

[0056] The comprehensive objective function is as follows: ; Where, , is the adjustment parameter, used to balance the completion efficiency and resource balance. Through this function, the system can improve efficiency while maintaining reasonable resource allocation.

[0057] And use the improved particle swarm algorithm (I-PSO) with non-dominated sorting mechanism to solve the multi-objective optimization of candidate allocation schemes, to balance task completion time and resource utilization balance. The steps of the improved I-PSO algorithm are as follows: Initialize particle swarm: randomly generate several particles, each particle corresponds to an activity allocation scheme, and initialize its speed and position, and calculate the reward and cost fitness value of each particle.

[0058] Non-dominated sorting and Pareto layering: Based on the two objectives of benefit R and cost C, the comprehensive objective F is obtained, and the particle population is sorted based on F, and the solution set is divided into different Pareto front layers. The low-order front layer is preferentially retained and used to guide the optimization direction of the group.

[0059] Elite reservation and diversity maintenance: In the updating process, the elite solution in the preamble layer is always reserved, and the congestion measure is used to ensure the balanced distribution of solutions in the target space, preventing the solution set from being too concentrated.

[0060] Update rule optimization: When updating the speed and position of the particle, not only the historical optimal solution and the global optimal solution of the group are referred to, but also a representative solution is selected from the current Pareto front layer as an additional reference to improve the global search. The representative solution is a small number of feasible and non-identical solutions selected from the current non-dominated front, such as the method that meets the highest benefit / lowest cost.

[0061] Dynamic inertia weight adjustment: A dynamic inertia weight ω(t) that decreases with the number of iterations is introduced, which gives priority to global search at the beginning and gradually enhances local development at the end to balance exploration and convergence. The dynamic inertia weight is gradually reduced with the number of iterations, which is used in the particle speed updating step of the improved I-PSO optimization module, that is, the parameter used by the algorithm to generate new solutions every generation. Its function is "more exploration at the beginning and fast convergence at the end", the formula is: ; where, is the maximum inertia weight value preset by the system, is the minimum inertia weight value preset by the system, is the actual number of iterations, is the maximum number of iterations preset by the system. The "more exploration at the beginning and fast convergence at the end" means that the initial larger weight value helps global exploration, and the gradually reduced weight value focuses on local development.

[0062] Adaptive correction mechanism: According to the estimated time / capacity / bandwidth requirement of each task and the available capacity of each execution unit, the utilization rate of each unit and the imbalance degree of utilization rate between units are calculated. If the utilization rate of a certain unit exceeds the upper limit or the overall allocation is uneven, it is determined that there is resource shortage or task allocation imbalance. When resource shortage or task allocation imbalance is detected, the real-time utilization rate of each execution unit is first calculated and compared with the average utilization rate. If the deviation exceeds the preset threshold, the activities are reordered according to their priority, and low-priority activities are migrated from high-load units to low-load units. During the migration process, a greedy strategy is used to preferentially select activities with the smallest migration cost and the most significant reduction in imbalance degree for adjustment until the overall load of the system tends to be balanced, thereby ensuring the feasibility of task allocation. The migration cost is calculated by taking the activity waiting time, data transmission cost (e.g. cost or event), restart warm-up time, and possible rollback loss as calculation parameters. Since the units of the calculation parameters are all seconds or cost unit prices, the parameters can be converted to the same unit and added to obtain the migration cost. If the saved cost (time or cost) is greater than the migration cost, the migration is performed, otherwise it is not migrated.

[0063] Termination and solution output: When the upper limit of the number of iterations or the Pareto solution set is stable, the final Pareto optimal solution set is output, the MCP-Server selects the allocation scheme with the highest comprehensive score, generates a task allocation matrix, and issues an appropriate standard MCP interface for each sub-task to execute.

[0064] Compared with traditional methods, the improved I-PSO algorithm can balance between task completion efficiency and resource balance, ensuring multi-objective optimization capability while improving solution set quality and diversity, effectively avoiding local optimum through non-dominated sorting and crowding degree mechanism. In addition, the algorithm introduces dynamic weight and adaptive correction mechanism, which can maintain strong adaptability and stability under different business load scenarios, and achieve a good balance between complexity and convergence speed, with high engineering implementability and application value.

[0065] After each interface program is executed, the output data is returned, and the large model determines whether to continue executing subsequent activities according to the validity of these data. If the current output data is valid and meets the preset conditions, the large model will continue to perform the next operation; otherwise, the system will make appropriate adjustments to ensure that the task proceeds according to the best process.

[0066] Sub-service process interaction. When the execution of the large model reaches a certain condition, the system automatically triggers the next sub-service process. Through the prompt primitive of the MCP Server, the large model can activate the next agent and pass on the necessary instructions and data. The call of the prompt primitive not only ensures the sequential execution of the process, but also dynamically adjusts the task allocation and execution strategy of each agent to adapt to different fault scenarios. At this time, the large model not only takes charge of global coordination, but also obtains the data required by the current sub-service process through the MCP interface and transmits it to the subsequent agent in real time for processing. Through this mechanism, the system realizes the orderly advancement of tasks and seamless collaboration between agents.

[0067] Multi-agent collaborative execution. During collaborative process execution, the large model controls the start time of each agent according to task dependency, so that related models are triggered to run after meeting the preconditions, and are pushed forward in parallel when necessary to ensure continuous advancement of the overall process without interruption. After each large model processes the sub-service process in sequence or in parallel according to task allocation, the system integrates the output results of each agent through a unified MCP protocol to form a complete collaboration link. Ultimately, through multi-level and multi-agent collaboration, the system can efficiently complete the fault diagnosis task and output reliable results. In this process, the task execution of each agent is coordinated and monitored by global scheduling, ensuring accurate processing of multi-modal data and reliability of fault location. Through coordination and information sharing between agents, the fault diagnosis system can achieve efficient execution and ensure controllable advancement of the business process, ultimately achieving rapid and accurate positioning and processing of faults.

[0068] Step 3, multi-modal small model collaborative job.

[0069] To achieve efficient processing of multi-modal data in distribution network scheduling fault diagnosis, a multi-modal small model collaborative job method is proposed, which defines small model collaboration structure and sub-service process to realize the organic combination of local accurate reasoning and global process control. As shown in Figure 2 , it specifically includes the following contents: Small model set definition. Machine learning models that can independently reason and output effective results based on specified data structures are called expert small models, with parameter sizes at least an order of magnitude lower than large models, focusing on processing specific modal data (such as images, telemetry data, etc.).

[0070] For a large model lm i in the fault diagnosis business process, its callable small model set is defined as: ; Where sm is represents the large model lm iThe system uses callable small model identifiers, where m represents the total number of small models involved in the entire business process. Different large models can reuse the same set of small models. This design enables the system to efficiently reuse small models across multiple business nodes, reducing computational costs and improving processing efficiency.

[0071] The set of MCP interfaces for small models. Each small model interacts with other intelligent agents through standardized interfaces and supports task processing within sub-business processes. The set of MCP primitives and corresponding standardized interfaces that small models can call are as follows: ; ; These interfaces support multiple operations of the small model in sub-business processes, such as data acquisition, database query, algorithm invocation, and prompt word generation, ensuring that the small model can perform its specialized tasks throughout the business process and achieve efficient data interaction with other intelligent agents.

[0072] Definition of Collaboration Relationship between Large and Small Models. The collaboration relationship between the large and small models can be defined using an agent quadruple, where each small model can either act as an independent expert model for inference or as an activity node in a sub-business process of the large model for collaboration. Specifically defined as follows: ; This agent collaboration model enables seamless integration between large and small models, fully leveraging the advantages of large models in global scheduling and the precise processing capabilities of small models in specific tasks. Each agent interacts with the outside world through the MCP Server, dynamically adjusting the order and content of task execution to ensure the flexibility and adjustability of the entire fault assessment process.

[0073] Collaborative Business Process Between Large and Small Models. Multiple agents and small models jointly constitute the collaborative fault assessment business process for power distribution network scheduling. After decomposing the overall task into multiple sub-processes, the large model will invoke the corresponding small models one by one according to predetermined dependencies, allowing them to process their respective data content. Input data can come directly from the output of the large model or from external data forwarded by the MCP Server. Furthermore, small models can obtain other relevant data from the MCP Server through MCP interaction when necessary. Through this collaborative process, the collaborative relationship between the large and small models can efficiently and dynamically handle different types of business tasks.

[0074] The execution of the entire collaborative process not only improves the accuracy and efficiency of fault diagnosis but also achieves an organic combination of local intelligence and global control. Distribution network dispatchers can use this method to construct flexible and reusable small-model processing sequences, dynamically scheduling intelligent agents at each key node according to specific tasks to achieve more accurate and efficient fault diagnosis. Simultaneously, the system ensures process transparency and stability, effectively reducing risks arising from erroneous judgments and data inconsistencies. Through the multimodal large-scale model collaborative operation method of this invention, the distribution network dispatch fault diagnosis system can achieve efficient processing of cross-modal data while ensuring accuracy and stability under different business scenarios, providing a more intelligent and efficient operational guarantee for the power system.

[0075] This embodiment provides a multi-agent operation method that combines large and small models, which can significantly improve the accuracy and operational stability of the distribution network scheduling fault assessment system. By combining the global semantic understanding, task planning and process scheduling capabilities of the large model with the precise processing capabilities of the small model in specific modalities or professional tasks, it achieves efficient fusion and unified processing of multimodal data (including telemetry, teleindication, images and text, etc.), thereby effectively reducing the misjudgment that may be caused by a single model and improving the reliability of fault diagnosis and assessment results.

[0076] This embodiment provides a multi-agent operation method that combines large and small models. It can achieve dynamic scheduling and optimization of business processes through the agent collaboration mechanism. The large model can flexibly adjust the execution order and calling logic of the small model according to the analysis scenario, and quickly adapt to different types of fault handling needs. The multi-modal agent collaboration mechanism greatly reduces the frequency of manual intervention and shortens the fault location and handling time, thereby significantly improving the processing efficiency of distribution network scheduling fault analysis and ensuring the safe and stable operation of the power supply system.

[0077] This embodiment provides a multi-agent operation method that combines large and small models, exhibiting excellent scalability and long-term adaptability. Users can add new small models, update fault judgment logic, or design new agents at any time according to business needs to meet the diverse and ever-changing power grid dispatching environment. Simultaneously, it maintains efficient and accurate fault judgment capabilities even in complex cross-regional and multi-source heterogeneous data scenarios, and ensures high reliability and stability during long-term operation through the traceability and controllability of the agent collaboration process.

[0078] Example 2 This embodiment provides a multi-agent operation system for collaborative large and small model operations, including: The task acquisition module is configured to: acquire the power task description; The task parsing module is configured to: parse the power task description through a large model agent, divide it into multiple activities, extract the execution parameters of each activity, and determine the small model agent corresponding to each activity based on the cooperation relationship between the large and small models; The process execution module is configured to: based on execution parameters, combined with the available resources and load status of each execution unit, and with the benefit and cost of task allocation as the objective function, use a particle swarm optimization algorithm based on a non-dominated sorting mechanism to allocate each activity to different execution units, which are then processed by small model agents. The small model agents complete the corresponding activities by calling small models through an interface. Specifically, when the particle swarm optimization algorithm based on the non-dominated sorting mechanism detects resource shortages or uneven task allocation, it calculates the real-time utilization rate of each execution unit and compares it with the average utilization rate. If the deviation exceeds a threshold, the tasks are reordered according to their priority, and low-priority tasks are migrated from high-load execution units to low-load execution units.

[0079] It should be noted that each module in this embodiment corresponds one-to-one with each step in Embodiment 1, and their specific implementation processes are the same, so they will not be repeated here.

[0080] Example 3 This embodiment provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of a multi-agent operation method for large-scale model collaboration as described in Embodiment 1 above.

[0081] Example 4 This embodiment provides a computer device, such as... Figure 3 As shown, the system includes a computer-readable storage medium 1003, a processor 1001, a communication interface 1002, and a computer program stored on the computer-readable storage medium 1003 and executable on the processor 1001. The processor 1001, communication interface 1002, and computer-readable storage medium 1003 can be connected via a bus or other means. The communication interface 1002 is used to receive and send data. When the processor 1001 executes the program, it implements the steps of the multi-agent operation method for large and small model collaboration as described in Embodiment 1 above.

[0082] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A multi-agent job method of size model coordination, characterized by, The method comprises: obtaining a power task description; analyzing the power task description by a large model agent, dividing it into multiple activities, extracting the execution parameters of each activity, and determining the small model agent corresponding to each activity according to the cooperation relationship between the large model and the small model; based on the execution parameters, combining the available resources and load state of each execution unit, taking the benefits and costs of task allocation as the objective function, and through the particle swarm algorithm based on the non-dominated sorting mechanism, each activity is allocated to different execution units for processing by the small model agent, and the small model agent completes the corresponding activity by calling the small model through the interface; wherein, when detecting resource shortage or uneven task allocation, the particle swarm algorithm based on the non-dominated sorting mechanism calculates the real-time utilization rate of each execution unit and compares it with the average utilization rate, and if the deviation exceeds the threshold, the low-priority tasks are migrated from the execution unit with high load to the execution unit with low load according to the priority of the tasks.

2. The size model collaborative multi-agent job method of claim 1, wherein, In the migration process, a greedy strategy is adopted to preferentially select the activity with the smallest migration cost and the largest reduction in imbalance degree for adjustment until the load is balanced.

3. The size model collaborative multi-agent job method of claim 2, wherein, The migration cost includes activity waiting time, data transmission cost, restart warm-up time and rollback loss.

4. The size model collaborative multi-agent job method of claim 1, wherein, The execution parameters include time limit for completion and resource requirements.

5. The size model collaborative multi-agent job method of claim 1, wherein, The large model agent includes a large model body, an MCP primitive, and a standardized interface contained in the primitive.

6. The size model collaborative multi-agent job method of claim 1, wherein, The particle swarm algorithm based on the non-dominated sorting mechanism adopts a dynamic inertia weight that decreases with the number of iterations for particle velocity update.

7. The size model collaborative multi-agent job method of claim 1, wherein, The size model cooperation relationship is represented as: ; wherein, Agent i represents the intelligent agent corresponding to the i-th large model, lm i represents the i-th large model ontology, SM fa (lm i ) represents the set of small models callable by the i-th large model, MCP pr (sm is ) and pr w () respectively represent the MCP primitives and the corresponding standardized interface set callable by the s-th small model callable by the i-th large model.

8. A size model cooperative multi-agent job system, characterized by, The method comprises: a task acquisition module configured to obtain a power task description; a task analysis module configured to analyze the power task description by a large model agent, divide it into multiple activities, extract the execution parameters of each activity, and determine the small model agent corresponding to each activity according to the cooperation relationship between the large model and the small model; a process execution module configured to, based on the execution parameters, combine the available resources and load state of each execution unit, take the benefits and costs of task allocation as the objective function, and through the particle swarm algorithm based on the non-dominated sorting mechanism, allocate each activity to different execution units for processing by the small model agent, and the small model agent completes the corresponding activity by calling the small model through the interface; wherein, when detecting resource shortage or uneven task allocation, the particle swarm algorithm based on the non-dominated sorting mechanism calculates the real-time utilization rate of each execution unit and compares it with the average utilization rate, and if the deviation exceeds the threshold, the low-priority tasks are migrated from the execution unit with high load to the execution unit with low load according to the priority of the tasks.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to realize the steps in the large and small model cooperative multi-agent job method of any one of claims 1-7.

10. A computer device, comprising a computer readable storage medium, a processor, and a computer program stored on the computer readable storage medium and executable on the processor, wherein, The processor executes the program to realize the steps in the large and small model cooperative multi-agent job method of any one of claims 1-7.

Images