A power supply station-oriented multi-agent collaborative scheduling system and a scheduling method thereof

Abstract

This application discloses a multi-agent collaborative scheduling system and method for power supply stations, relating to the field of multi-agent collaborative scheduling. The system includes: a master control agent subsystem, used to perform intent recognition, task planning, and task decomposition on service requests sent by the power supply station's front-end equipment, and to call service agents to execute decomposition tasks based on the A2A protocol; an MCP tool standard subsystem, used to encapsulate and manage data in the power supply station's database based on the MCP protocol to generate MCP tools; and a service agent subsystem, used to create service agents to call the MCP tool to read data stored in the power supply station's database, execute corresponding decomposition tasks based on the read data, and feed back the task execution results to the master control agent. This application enables multi-agent collaborative scheduling in real-world power supply station scenarios involving cross-systems, cross-role operations, long tasks, and strong compliance.

Description

Technical Field

[0001] This application relates to the field of multi-agent cooperative scheduling, and in particular to a multi-agent cooperative scheduling system and scheduling method for power supply stations. Background Technology

[0002] With the development of enterprise digitalization and the application of large-scale models, the technical evolution of power supply station operations (work order generation, task scheduling, personnel dispatching, on-site guidance, evaluation, summarization, service scripts, etc.) has mainly undergone a transformation from "single-point informatization → process automation → intelligent assistantization." Current mainstream practices can be summarized into the following categories, which often coexist in combination: Centered on BPM (Business Process Management) or RPA (Robotic Process Automation), it integrates with existing business systems (such as work orders, GIS, archives, HR, etc.) to automate fixed processes and form filling through process nodes and rule bases. Based on personnel skills, geographical location, duration, workload, etc., it uses heuristic or simple optimization algorithms (such as greedy / rule scoring) for automatic or recommended order dispatch. With the help of data platform / indicator library and knowledge base (FAQ / work instructions), it provides reference information and script templates for agents / on-site through retrieval or RAG. Around front-end scenarios such as customer service, order entry, and repair consultation, it deploys conversational assistants, with LLM completing intent recognition, question answering, and template generation.

[0003] In summary, existing solutions either focus on fixed processes and rule execution (BPM / RPA / scheduling optimization) or on point-like intelligence enhanced by dialogue / retrieval (single Copilot / RAG). In real-world power supply substation scenarios involving cross-systems, cross-role, long-term tasks, and strong compliance, their overall collaborative capabilities remain limited. Summary of the Invention

[0004] The purpose of this application is to provide a multi-agent collaborative scheduling system and scheduling method for power supply stations, which can improve the efficiency of task scheduling through multi-agent collaboration in the practical scenario of power supply stations with "multi-stage, strong constraints, and high concurrency".

[0005] To achieve the above objectives, this application provides the following solution: In a first aspect, this application provides a multi-agent collaborative scheduling system for power supply stations, including: a master control agent subsystem, a business agent subsystem, an MCP tool standard subsystem, and an A2A agent communication subsystem; The main control intelligent agent subsystem is used to perform intent recognition on the service requests sent by the power supply station front-end equipment, plan the whole process tasks that meet the intent recognition results and decompose the tasks, arrange the dependencies between the decomposed tasks, and call the corresponding business intelligent agents to execute the corresponding decomposed tasks according to the dependencies between the decomposed tasks. The MCP tool standard subsystem is used to encapsulate and manage the data in the power supply station's database based on the MCP protocol, and generate MCP tools so that various business intelligence agents can call the MCP tools to read data from various external systems of the power supply station; The Business Intelligence Subsystem is used to create a corresponding Business Intelligence Agent for each dismantling task in the Main Control Intelligence Agent Subsystem. Each Business Intelligence Agent calls the MCP tool to read the data stored in the power supply station's database, executes the corresponding dismantling task based on the read data, and feeds back the task execution results to the Main Control Intelligence Agent. The A2A intelligent agent communication subsystem is used to implement task delegation, context sharing, state synchronization and event feedback between the master intelligent agent and various business intelligent agents using the A2A protocol.

[0006] Secondly, this application provides a scheduling method for a multi-agent cooperative scheduling system for power supply stations, including: The main control intelligent agent is used to identify the intent of the service requests sent by the front-end equipment of the power supply station, plan the whole process tasks that meet the intent identification results, decompose the tasks, and arrange the dependencies between the decomposed tasks. Based on the MCP protocol, the data in the power supply station's database is encapsulated and managed to generate MCP tools; For each decomposition task in the main control intelligent agent subsystem, create a corresponding business intelligent agent; The master control agent uses the A2A protocol to call the corresponding business agent to execute the corresponding dismantling task based on the dependencies between the dismantling tasks. When each business intelligence agent executes its corresponding dismantling task, it calls the MCP tool to read data from the power supply station's database; Each business intelligence agent executes the corresponding decomposition task based on the data it reads, and feeds back the task execution results to the master intelligence agent based on the A2A protocol; The master control agent feeds back the task execution results of each business agent to the power supply station's front-end equipment.

[0007] According to the specific embodiments provided in this application, this application has the following technical effects: This application provides a multi-agent collaborative scheduling system and method for power supply stations. The system includes: a master control agent subsystem, used to perform intent recognition, task planning, task decomposition, and A2A protocol-based invocation of corresponding business agents to execute decomposition tasks for service requests sent by the power supply station's front-end equipment; an MCP tool standard subsystem, used to encapsulate and manage data in the power supply station's database based on the MCP protocol to generate an MCP tool; and a business agent subsystem, used to create corresponding business agents, each of which calls the MCP tool to read data stored in the power supply station's database, executes corresponding decomposition tasks based on the read data, and feeds back the task execution results to the master control agent. This application provides a multi-agent solution of "MCP standard interface + A2A collaborative orchestration," connecting the master control agent ("one leader and three staff") with various business agents via the A2A protocol to achieve cross-stage task planning, resource orchestration, and process control. This application can achieve multi-agent collaborative scheduling in real-world power supply station scenarios involving cross-systems, cross-role tasks, long tasks, and strong compliance, improving the efficiency of task scheduling and execution. Attached Figure Description

[0008] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 A functional module diagram of a multi-agent collaborative scheduling system for power supply stations provided in an embodiment of this application; Figure 2 A schematic diagram of the overall architecture of a multi-agent cooperative scheduling system for power supply stations provided in an embodiment of this application; Figure 3 A schematic diagram illustrating the interface relationship between a main control intelligent agent, a business intelligent agent, and the MCP protocol layer, provided in an embodiment of this application; Figure 4 This is a schematic diagram of the lifecycle of a tool invocation process based on the MCP protocol provided in an embodiment of this application; Figure 5 A schematic diagram of the lifecycle of agent discovery and task assignment based on the A2A protocol provided in an embodiment of this application; Figure 6 This is a flowchart illustrating a scheduling method for a multi-agent collaborative scheduling system for power supply stations, provided as an embodiment of this application. Detailed Implementation

[0010] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0011] While existing solutions have played a role in improving efficiency and reducing workload, they still have the following common shortcomings in the practical scenarios of power supply stations characterized by "multiple links, strong constraints, and high concurrency," making it difficult to support a closed-loop process of "order acceptance → execution → evaluation → continuous optimization": (1) Existing systems mostly connect directly via their own APIs / SDKs, resulting in inconsistent plugin / function call standards and a lack of unified, discoverable, authorizable, and auditable descriptions of tool capabilities. New systems / tools require repeated integration and adaptation, leading to slow expansion, weak governance, and difficulty in large-scale replication. This application addresses this issue by introducing the MCP protocol.

[0012] (2) Single agents often remain at the level of "single-stage automation," making it difficult to perform unified task decomposition and dependency orchestration for work order generation, scheduling, dispatching, guidance, evaluation, and summary. Once concurrent conflicts (people / vehicles / materials), cross-domain data, time limits, and compliance constraints are involved, it is easy to experience chain breaks or mutual waiting; complex business often requires multiple steps of retrieval, verification, matching, and write-back. This application introduces a collaborative framework of master control agent and multi-business agent, through which the master control agent performs unified task decomposition and dependency locking for the entire lifecycle from work order generation to work summary, thereby solving the problems of cross-domain data chain breaks and time-series coordination.

[0013] (3) In reality, there are common anomalies such as timeouts, rate limiting, interface changes, and partial success. Existing solutions generally lack standardized retry / degradation / alternative paths and manual fallback mechanisms. In response to common anomalies such as timeouts, rate limiting, interface changes, and partial success, this application establishes a unified anomaly handling link through a master control agent. When a business sub-agent fails to call the business system or tool via MCP, the master control agent dynamically triggers retries, degradation, alternative agent switching, compensation execution, or manual fallback based on the anomaly type, task stage, and candidate processing link, thereby avoiding task interruption, long-term blocking, or overall failure.

[0014] (4) For long-process tasks (spanning hours / shifts), the lack of heartbeat and breakpoint resume functionality can easily lead to "interruption and untraceability". This application addresses long-process tasks that span hours and shifts by maintaining a unified task state machine through a master control agent and relying on the A2A protocol to achieve heartbeat reporting, stage state synchronization, and abnormal resume control, supporting breakpoint recovery and full-process traceability after task interruption.

[0015] (5) Compliance of communication scripts, operational qualifications, on-site photo verification, and anonymization of sensitive fields are often treated as "external supplementary checks." When retrieval / generation bypasses the standards and compliance constraints, it is easy to generate illusionary strategies or illegal communication scripts. This application eliminates illusionary strategies and illegal communication scripts by using gateway-level unified authentication and full-link CoT (CoT) traceability, combined with a built-in compliance quality inspection component. This compliance quality inspection component is built into the unified governance link of the main control agent and can be used for secondary verification after the output results of key business sub-agents are returned.

[0016] (6) Most solutions only return "results / suggestions," lacking end-to-end SLA and KPI data (processing time, percentage of orders handled by multiple personnel, first-time completion rate, second-visit rate, complaint rate, etc.), making it difficult to drive strategy evolution, route optimization, and profile updates. Benefits are not quantifiable, and operations are unsustainable. This application uses this master control agent to uniformly aggregate task events, processing time, status transitions, and result data from various business sub-agents, and combines this with business system feedback data to form an end-to-end SLA / KPI indicator system to support strategy optimization, process governance, and operational evaluation.

[0017] (7) In heterogeneous systems, network fluctuations, peak concurrency, and temporary rule changes, single agents or loosely distributed multiple agents are prone to inconsistencies (including inconsistencies in task status, rule versions, result judgments, and cross-system data). This application improves consistency and stability in multi-agent collaborative scenarios by unifying the orchestration and arbitration of the master intelligent agent, combined with the state synchronization capability of the A2A protocol and the standardized tool access mechanism of MCP.

[0018] (8) The lack of unified authentication at the channel level (identity, role, minimum privilege) and protocol-level state tracking leads to unpredictable collaborative behavior and a lack of full-link traceability (who, when, which tool, which parameters, and what results were produced), making auditing and review difficult. This application utilizes a master control agent to perform identity verification, role mapping, and minimum privilege control on user requests and calls from various business agents through a unified access gateway. When accessing MCP tools and business systems, business agents inherit or map controlled credentials to avoid unauthorized calls. This application implements identity authentication, role mapping, and minimum privilege control at the unified access layer, and performs full-process tracking of task creation, decomposition, transfer, execution, exceptions, and completion at the A2A protocol layer, ensuring that collaborative behavior is controllable, auditable, and reviewable.

[0019] In addition, this application uses a feedback loop based on log data from the task execution process to drive the adaptive evolution of multi-agent collaborative strategies, ensuring the determinism, compliance, and quantifiable benefits of system behavior in heterogeneous environments.

[0020] In response, this application proposes a multi-agent collaborative scheduling system and scheduling method for power supply stations. It is based on a three-element intelligent agent (master intelligent agent) collaboratively controlling multiple service intelligent agents, and combines the MCP protocol and A2A protocol to execute relevant tasks to meet service requirements.

[0021] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] In one exemplary embodiment, such as Figure 1 and Figure 2 As shown, a multi-agent collaborative scheduling system for power supply stations is provided, including: a master control agent subsystem, a business agent subsystem, an MCP tool standard subsystem, and an A2A agent communication subsystem.

[0023] Figure 2 The diagram illustrates the overall architecture of the power supply station's multi-agent collaborative scheduling system. From bottom to top, the diagram includes the data layer, basic capability layer, agent service layer, and business application layer (power supply station integrated business digital platform, work summary, strategy assistance, etc.). The basic capability layer includes an artificial intelligence platform (containing the Guangming Power large model, RAG module, etc.) and a data platform (data storage, data processing, data analysis, etc.); the agent service layer includes a leader-three-person agent, A2A protocol, various business agents, and MCP service access. Figure 2 The document illustrates the two-layer standardized architecture (MCP protocol + A2A protocol) of this application and the relationships between the layers.

[0024] The master control intelligent agent subsystem, business intelligent agent subsystem, MCP tool standard subsystem, and A2A intelligent agent communication subsystem described in this application correspond to the intelligent agent service layer in the overall architecture.

[0025] The main control intelligent agent subsystem is used to perform intent recognition on the service requests sent by the power supply station front-end equipment, plan the entire process of tasks that meet the intent recognition results and decompose the tasks, arrange the dependencies between the decomposed tasks, and call the corresponding business intelligent agents to execute the corresponding decomposed tasks according to the dependencies between the decomposed tasks.

[0026] In another exemplary embodiment of this application, the master control agent subsystem is further configured to distribute cross-agent communication routes, manage the global state of task execution, and perform self-healing processing for abnormal situations during task execution.

[0027] The master control agent subsystem serves as the central decision-making brain and collaborative orchestration foundation of the scheduling system. Independent of specific business logic, it focuses on task management and computing power scheduling throughout the entire lifecycle. Its core functions include intent recognition, global task decomposition, dependency orchestration, cross-agent communication routing and distribution, global state management, and anomaly self-healing.

[0028] The system can simultaneously receive task requests from user natural language input, external work order pushes, or system event triggers. The main control agent identifies task types (such as repair reports, inspections, power restoration, and evaluations) and extracts key information (user number, transformer area, time window, priority, SLA constraints, etc.) through a task parsing model to generate structured "task objects".

[0029] The controlling agent invokes internal templates based on the task type (the templates contain corresponding processing flows for different task requirements), breaks down the task into multiple executable subtasks, and constructs a directed acyclic graph (DAG) to describe their dependencies: Parallel execution of subtasks that can be parallelized (such as multi-work order generation and resource verification); For tasks with sequential constraints (such as scheduling → assignment → guidance → evaluation), perform serial scheduling; For subtasks competing for the same resource, a resource contention detection and dynamic programming algorithm is configured to automatically identify conflict points and reset the task execution priority, thereby achieving orderly allocation and conflict-free access to resources.

[0030] Each business intelligence agent uses the MCP tool to obtain the real-time status and geographical location of personnel, vehicles, and materials. The master control intelligence agent dynamically selects suitable executors based on ETA (estimated arrival time from the start of the order to the destination), personnel skill tags, workload level (such as task difficulty), and historical performance.

[0031] When multiple tasks compete for the same resource (such as a vehicle or work team), the master control agent performs secondary planning based on task priority, timeliness indicators, and fairness weights. If the constraints still cannot be met, the scheduling process is rolled back to reallocate the resource, and the reasons for failure and the blacklist are recorded in the log.

[0032] In response to network fluctuations, interface timeouts, data loss, and compliance verification failures, the system supports retries, degradation, manual takeover, and breakpoint resumption. This ensures the continuity and traceability of long-cycle tasks.

[0033] The main control intelligent agent subsystem adopts a closed-loop control architecture of "perception-planning-scheduling-monitoring", which specifically includes the following core functional modules: The global planning and intent parsing module is used to receive business requests (such as upper-layer requests in natural language form) sent by the power supply station's front-end equipment using the main control agent, perform semantic parsing and intent alignment using the large model's reasoning capabilities, and obtain intent recognition results. Based on the intent recognition results, it plans the full-process tasks that satisfy the intent recognition results. Based on the preset atomic capability library, it decomposes the full-process tasks (complex long-process tasks) that satisfy the intent recognition results into parts. Based on the decomposed tasks, it constructs a directed acyclic graph (DAG) containing temporal and data dependencies, and clarifies the input and output standards of each decomposed task.

[0034] The dynamic scheduling and routing engine module is used to intelligently distribute communication routes based on task attributes and the real-time load status of downstream business agents by the master control agent, and to call the corresponding business agents to execute the corresponding dismantling tasks according to the dependencies between the dismantling tasks.

[0035] In another exemplary embodiment of this application, the dynamic scheduling and routing engine module is further configured to, when multiple tasks compete for the same scheduling resource, utilize the master control agent to perform secondary planning based on task priority, timeliness indicators and fairness weights. If the secondary planning still results in concurrent resource conflicts, the scheduling resources corresponding to each task are re-determined.

[0036] In another exemplary embodiment of this application, the dynamic scheduling and routing engine module is further configured to utilize the master control agent to call multiple business agents in parallel via the A2A protocol to form a distributed task orchestration network.

[0037] The dynamic scheduling and routing engine module, as the system's execution hub, is responsible for intelligent routing and distribution based on task attributes (such as real-time performance and complexity) and the real-time load status of downstream business agents. It maintains a "business agent registry," enabling it to dynamically select the most suitable business agent to execute subtasks based on capability matching. When concurrent tasks are involved, it performs resource locking and priority arbitration based on a global cost function to prevent deadlocks or resource contention. "Selecting the most suitable business agent" means that when multiple candidate business agents exist to handle the same task, the master agent is not fixed to a single agent but dynamically selects the most appropriate execution target based on task characteristics and the current state of the candidate agents. The "most suitable" is determined primarily by considering the following factors: 1) Capability matching degree: Whether the candidate business intelligence agent has the professional capabilities, rule base and tool chain to handle this type of task; 2) Task adaptability: Whether it is suitable for the current task type, task complexity, and processing time requirements; 3) Real-time load status: Is the candidate business agent currently busy? Is the queue length too high? 4) MCP tool availability: Whether the business system, knowledge base, or tool it is connected to is currently available; 5) Historical performance: Success rate, latency, and processing quality on similar tasks; 6) Permissions and compliance constraints: Whether the user has the permission to access the corresponding business data and perform the corresponding operations.

[0038] The main control agent determines the execution target by dynamically arbitrating multiple candidate business agents based on preset scheduling rules or cost evaluation mechanisms.

[0039] The context bus and state management module is used to seamlessly transfer the execution results between different business agents using the master agent, and to resume execution and restore state when the task is interrupted.

[0040] In another exemplary embodiment of this application, the context bus and state management module is further configured to use the master control agent to feed back the task execution results of each business agent to the power supply station front-end equipment.

[0041] The context bus and state management module provides unified short-term and long-term memory management. It is responsible for the lossless transmission of execution results between different business agents, ensuring that downstream agents can obtain key parameters generated by upstream agents; at the same time, it maintains the global session state in real time (such as task progress and intermediate variables), supports task resumption and state recovery after interruption, and achieves data consistency throughout the entire link.

[0042] This application utilizes a master control agent to uniformly orchestrate, manage the state of, govern anomalies, control access, and audit and trace multiple business sub-agents, thereby supporting the stable collaborative execution of complex and long-process businesses in heterogeneous system environments.

[0043] In this application, the master control agent adopts a "global arbitration mechanism based on multi-factor cost function" for task routing and scheduling decisions, candidate selection of multiple business sub-agents, priority selection and failure switching of MCP tool call links. When multiple business agents, multiple tool links or multiple execution paths can handle the same task, the master control agent comprehensively considers multiple factors (the matching degree between the current task type and business domain, the current load and availability of the sub-agent, the success rate of tool / MCP call, processing latency and timeout risk, historical execution effect, data permissions and compliance constraints, execution cost and resource consumption) to score and select candidate execution schemes, thereby determining the optimal scheduling result. This enables optimal resource scheduling under concurrent conflicts and millisecond-level adaptation to rule changes.

[0044] The Business Intelligence Subsystem is used to create a corresponding Business Intelligence Agent for each dismantling task in the Main Control Intelligence Agent Subsystem. Each Business Intelligence Agent calls the MCP tool to read the data stored in the power supply station's database, executes the corresponding dismantling task based on the read data, and feeds back the task execution results to the Main Control Intelligence Agent.

[0045] like Figure 2 and Figure 3 As shown, corresponding scenario-based intelligent agents (business intelligent agents) are constructed around the seven core aspects of power supply station operations: work order generation, work order scheduling, work order dispatch, strategy generation, work order evaluation, work summary, and service scripts. Figure 3 This diagram illustrates the interface between a main control intelligent agent with three members, a business intelligent agent, and the MCP protocol layer. Figure 3 In this process, the front-end interactive page calls the main control intelligent agent (comprising a leader and three staff members) via HTTP. Based on pre-configured rules, the main control intelligent agent sequentially or conditionally calls various business intelligent agents, including the work order generation intelligent agent, work order scheduling intelligent agent, work order dispatch intelligent agent, strategy generation intelligent agent, work order evaluation intelligent agent, work summary intelligent agent, and service script intelligent agent, via the A2A protocol. Each business intelligent agent then accesses the power supply station's business database (including work order information, user profiles, equipment profiles, human resources information, operational data, and historical records) and its self-built strategy knowledge base (which stores task processing strategies) through the MCP (Model Context Protocol) layer, achieving unified data and knowledge acquisition.

[0046] The business intelligence agents include: work order generation intelligence agent, work order scheduling intelligence agent, work order dispatch intelligence agent, strategy generation intelligence agent, work order evaluation intelligence agent, work summary intelligence agent, and service script intelligence agent.

[0047] The work order generation intelligent agent is used to call the work order information stored in the power supply station's database based on the MCP tool, generate work order information for the task corresponding to the current business request, and send the work order information to the main control intelligent agent.

[0048] The work order scheduling intelligent agent is used to call the resources stored in the power supply station's database based on the MCP tool, determine the resources that can be called to meet the current business request, and send the callable resources to the main control intelligent agent.

[0049] The work order dispatching AI agent is used to call the resource attribute tags stored in the power supply station's database based on the MCP tool, determine the resources to be dispatched for the task from the currently available resources, and send the resources to be dispatched for the task to the main control AI agent.

[0050] The strategy generation agent is used to call the task response strategies stored in the power supply station's database based on the MCP tool, determine the execution strategy to be used for the current task according to the task response strategy, and send the execution strategy to the master control agent.

[0051] The work order evaluation agent is used to call the task execution status recorded in the power supply station's database based on the MCP tool, evaluate the task execution status, and send the evaluation results to the main control agent.

[0052] The work summary agent is used to call the task execution status recorded in the power supply station's database based on the MCP tool, summarize the work based on the task execution status, and send the work summary results to the main control agent.

[0053] The service script intelligent agent is used to generate service scripts based on the service object tags recorded in the power supply station's database using the MCP tool, and then send the service scripts to the main control intelligent agent.

[0054] Each business agent has a clear input / output interface, independent MCP tool invocation capability and local policy logic. It can execute sub-tasks independently or achieve multi-agent collaboration under the master control scheduling.

[0055] In another exemplary embodiment of this application, the business intelligence agent subsystem is further configured to optimize the business intelligence agent (such as the work order scheduling intelligence agent, the work order dispatch intelligence agent, the work order strategy intelligence agent, and the service script intelligence agent) based on the task execution log.

[0056] The MCP tool standard subsystem is used to encapsulate and manage data in the power supply station's database based on the MCP protocol, generating MCP tools so that various business agents can call these tools to read data from various external systems of the power supply station. For example... Figure 4 The diagram shows a lifecycle illustration of a tool invocation process based on the MCP protocol. Figure 4 The example demonstrates the interaction process between the user, LLM / agent, MCP client, MCP server, and data source / service: First, the user sends a command to the LLM / agent; the LLM / agent analyzes the command and determines the tool to be invoked; then, it sends a request to the MCP client to use the specific tool, the MCP client invokes the MCP server, the MCP server retrieves data from the corresponding data source / service and returns it; finally, the LLM / agent generates the tool execution result based on the returned data and feeds it back to the user.

[0057] The MCP tool standard subsystem is responsible for abstracting, encapsulating, and managing various external systems and data sources of the power supply station, including: work order system, GIS positioning system, human resources files, knowledge base, on-site photo verification, work qualification inspection, alarm and customer service system, etc.

[0058] All external tools (external databases) are registered through MCP description files, which include interface functions, input / output structures, parameter templates, access control, authentication policies, and logging requirements, achieving a resource access method of "unified discovery, unified invocation, and unified auditing".

[0059] Each business intelligence agent injects rich and structured power sector knowledge and labeled assets into the standardized query through Retrieval Enhancement Generation (RAG) technology.

[0060] The A2A intelligent agent communication subsystem is used to implement task delegation, context sharing, state synchronization and event feedback between the master intelligent agent and various business intelligent agents using the A2A protocol.

[0061] The A2A protocol provides a standardized communication mechanism for multi-agent systems, responsible for task delegation, context sharing, state synchronization, and event feedback. This protocol supports: JSON-RPC 2.0 formatted messages, full-link tracing via trace_id, streaming responses, and long-running tasks. Communication uses the JSON-RPC format, and trace_id is used throughout the entire call process. The A2A protocol supports streaming messages and heartbeat mechanisms to ensure uninterrupted long-running tasks. If a business agent exits abnormally, the master agent can restore the context and continue running under the same trace_id.

[0062] The master control agent can use the A2A protocol to call multiple business agents in parallel to form a "distributed task orchestration network". For example... Figure 5 As shown, a schematic diagram of the lifecycle of agent discovery and task issuance based on the A2A protocol is illustrated. Figure 5 The system comprises three roles: client, A2A server, and authentication server. First, the client obtains an Agent Card by accessing paths such as / .well-known / agent-card, enabling agent discovery. Then, based on the security scheme in the Agent Card, it interacts with the authentication server to complete identity verification and obtain tokens such as JWT. Subsequently, the client sends tasks to the A2A server via the / sendMessage interface based on the URL provided in the Agent Card. During task execution, the client can also obtain task status and stage results in a streaming manner through the / sendMessageStream interface until the task is completed.

[0063] All message interactions between the master control agent and various business agents are verified through authentication, role authorization, and parameter validation. Sensitive fields are anonymized before transmission, and the interaction log retains only the minimum necessary information for auditing. When the master control agent calls a business agent, it generates a unified signature request. The system records the call parameters, time, response status, and output results, and forms a complete call chain through the audit module.

[0064] In another exemplary embodiment of this application, the multi-agent cooperative scheduling system for power supply stations further includes: The monitoring and traceability subsystem is used to establish a unique identifier (such as trace_id) for each task execution, record task input, task execution path, tool calls, execution logs (abnormal logs), and task execution results, and generate a visualized execution chain for subsequent auditing and policy backtracking. Here, "task input" refers to the initial task data received and processed by the master control agent when initiating task scheduling. Its sources typically include user-submitted business requests and relevant contextual information already available on the system side. "Task input" includes not only the user's business request itself but also related work order information, user identity information, device object information, historical processing records, rule parameters, and subtask parameters generated by the master control agent after decomposition.

[0065] Based on the above-mentioned calling system, a scheduling process of "execution-feedback-optimization" can be implemented: Execution phase: Each business intelligence agent executes tasks according to the master control intelligence agent's plan. Each business intelligence agent calls the MCP tool to interact with external systems (external databases) and obtain the execution results of each decomposed task; all links are automatically timed.

[0066] Feedback phase: The system collects execution logs, task results, exception records and user feedback to form a set of quantifiable indicators.

[0067] Optimization Phase: The work order scheduling agent and strategy generation agent learn strategies based on log data and update the work order dispatch agent and internal knowledge templates based on log data. Simultaneously, compliance clauses, task templates, and prompts are version-managed, forming a self-evolving system. This application not only monitors and provides feedback on the task execution process but also continuously optimizes and manages the rules and execution configurations used in collaborative scheduling based on execution results, anomaly records, and user feedback, forming a sustainable iterative closed-loop optimization mechanism. The version management of "compliance clauses, task templates, and prompts" can apply to both the main control agent and various business sub-agents: for the main control agent, it mainly involves version management of task decomposition rules, routing strategies, scheduling strategies, and unified compliance constraints; for business agents, it mainly involves version management of their respective processing templates, domain prompts, execution strategies, and call parameter configurations.

[0068] This application enables the formation of a traceable closed loop of "order acceptance → execution → evaluation → continuous optimization," significantly shortening processing time and improving dispatch accuracy and service consistency. Through continuous feedback and evolution, the system gradually achieves intelligent governance and self-optimization of power supply station business processes.

[0069] The scheduling system in this application unifies the capabilities of external systems and tools through MCP, achieves cross-agent task orchestration through the A2A protocol, and combines a "one leader, three staff" master control agent with seven types of business agents to achieve intelligent collaboration and closed-loop scheduling of power supply station operations. Specifically: it adopts a two-layer standardized architecture: pioneering the combined application of the MCP and A2A protocols, unifying the "tool call interface" and "agent communication protocol" respectively, forming a standardized system of capabilities across systems and tasks; it adopts a master control-business multi-agent collaborative framework: through the "one leader, three staff" master control architecture and the layered collaboration of seven types of business agents, it achieves parallel decomposition and dynamic scheduling of complex tasks; it adopts a DAG-based task dependency modeling and conflict rearrangement mechanism: using explicit task graphs to describe task dependencies, resource constraints, and timing relationships, supporting conflict detection, path rearrangement, and automatic rollback; and it adopts a compliance embedding and full-link traceability mechanism: embedding compliance steps such as qualification verification, script standardization, and sensitive field desensitization into task nodes to achieve audit traceability and risk controllability.

[0070] Compared with existing systems that primarily rely on BPM, RPA, or standalone LLM, this invention has the following significant technical advantages: (1) Significantly improved efficiency: Multi-agent collaboration reduces manual coordination steps, shortening the average work order response time by 40%–60%.

[0071] (2) Improved work assignment accuracy: The scheduling strategy with integrated dimension constraints improves dispatch accuracy by approximately 25% and first-time delivery rate by 15%.

[0072] (3) High scalability and reusability: For any new business scenario, you only need to register the corresponding MCP tool and agent description file, without modifying the core logic, and the expansion cycle is shortened to one-third of the original.

[0073] (4) Improved compliance and auditability: All tasks are bound to a trace_id, ensuring traceability and compliance, and meeting the regulatory requirements of the power industry.

[0074] (5) Intelligent evolution and continuous optimization: The system has self-learning and strategy feedback capabilities, and can automatically adjust the task assignment and script models based on task performance, supporting cross-regional experience transfer and knowledge sharing.

[0075] (6) Promote the digital transformation of power supply stations: This will enable a leap from "rule-driven automation" to "intelligent decision-driven autonomy," providing a replicable intelligent governance paradigm for the energy and power industry.

[0076] In another exemplary embodiment of this application, such as Figure 6 As shown, a scheduling method for a multi-agent cooperative scheduling system for power supply stations is provided, including: S1: Utilize the main control intelligent agent to perform intent recognition on the service requests sent by the power supply station's front-end equipment, plan the entire process of tasks that satisfy the intent recognition results, decompose the tasks, and orchestrate the dependencies between the decomposed tasks.

[0077] S2: Based on the MCP protocol, encapsulate and manage the data in the power supply station's database to generate MCP tools.

[0078] S3: For each decomposition task in the master control intelligent agent subsystem, create a corresponding business intelligent agent.

[0079] S4: The master control agent calls the corresponding business agent to execute the corresponding dismantling task based on the dependency relationship between the dismantling tasks using the A2A protocol.

[0080] S5: When each business intelligence agent executes its corresponding dismantling task, it calls the MCP tool to read data from the power supply station's database.

[0081] S6: Each business intelligence agent executes the corresponding decomposition task based on the read data, and feeds back the task execution results to the master intelligence agent based on the A2A protocol.

[0082] S7: The main control agent feeds back the task execution results of each business agent to the power supply station's front-end equipment.

[0083] This application also provides an application scenario for the aforementioned multi-agent collaborative scheduling system for power supply stations. Specifically, the multi-agent collaborative scheduling system for power supply stations provided in this embodiment can be applied to a scenario of power outage repair for a specific household within a power supply station. Upon receiving a power outage repair request from the power supply station's front-end equipment, the main control agent performs intent recognition on the request, plans a complete workflow task to satisfy the intent recognition result (work order generation (generating a work order number for the task), work order scheduling (identifying available scheduling resources such as personnel, vehicles, and materials), work order assignment (assigning a repairman), service script (generating corresponding service scripts based on the service recipient's personality traits), strategy generation (specifically, how to investigate and repair the power outage), work order evaluation (evaluating processing satisfaction, time, etc.), and work summary), and decomposes the tasks and arranges the dependencies between the decomposed tasks. For each dismantling task in the master control intelligent agent subsystem, a corresponding business intelligent agent is created. Based on the dependencies between dismantling tasks, the master control intelligent agent uses the A2A protocol to call the corresponding business intelligent agent to execute the corresponding dismantling task. When executing their respective dismantling tasks, each business intelligent agent calls the MCP tool to read data from the power supply station's database. Each business intelligent agent executes its corresponding dismantling task based on the read data and feeds back the task execution results to the master control intelligent agent based on the A2A protocol. The master control intelligent agent then feeds back the task execution results of each business intelligent agent to the power supply station's front-end equipment.

[0084] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Moreover, the collection, use and processing of the relevant data are carried out in compliance with the relevant data protection laws and policies of the country where the location is located, and with the authorization granted by the owner of the corresponding device.

[0085] The databases involved in the various embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases.

[0086] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0087] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A multi-agent collaborative scheduling system for power supply stations, characterized in that, include: The system comprises a master control intelligent agent subsystem, a business intelligent agent subsystem, an MCP tool standard subsystem, and an A2A intelligent agent communication subsystem. The main control intelligent agent subsystem is used to perform intent recognition on the service requests sent by the power supply station front-end equipment, plan the whole process tasks that meet the intent recognition results and decompose the tasks, arrange the dependencies between the decomposed tasks, and call the corresponding business intelligent agents to execute the corresponding decomposed tasks according to the dependencies between the decomposed tasks. The MCP tool standard subsystem is used to encapsulate and manage the data in the power supply station's database based on the MCP protocol, and generate MCP tools so that various business intelligence agents can call the MCP tools to read data from various external systems of the power supply station; The Business Intelligence Subsystem is used to create a corresponding Business Intelligence Agent for each dismantling task in the Main Control Intelligence Agent Subsystem. Each Business Intelligence Agent calls the MCP tool to read the data stored in the power supply station's database, executes the corresponding dismantling task based on the read data, and feeds back the task execution results to the Main Control Intelligence Agent. The A2A intelligent agent communication subsystem is used to implement task delegation, context sharing, state synchronization and event feedback between the master intelligent agent and various business intelligent agents using the A2A protocol.

2. The multi-agent cooperative scheduling system for power supply stations according to claim 1, characterized in that, The master control agent subsystem is also used to distribute communication routes across agents, manage the global state of task execution, and handle abnormal situations during task execution.

3. The power supply oriented multi-agent collaborative scheduling system according to claim 2, wherein, The main control intelligent agent subsystem includes: The global planning and intent parsing module is used to receive service requests sent by the power supply station front-end equipment using the main control agent, perform semantic parsing and intent alignment using the large model reasoning capability, and obtain intent recognition results. Based on the intent recognition results, the module plans the full-process tasks that satisfy the intent recognition results. Based on the preset atomic capability library, the module decomposes the full-process tasks that satisfy the intent recognition results, constructs a directed acyclic graph containing temporal and data dependencies based on the decomposed tasks, and clarifies the input and output standards of each decomposed task. The dynamic scheduling and routing engine module is used to intelligently distribute communication routes based on task attributes and the real-time load status of downstream business agents by the main control agent, and to call the corresponding business agents to execute the corresponding decomposition tasks according to the dependencies between the decomposition tasks. The context bus and state management module is used to seamlessly transfer the execution results between different business agents using the master agent, and to resume execution and restore state when the task is interrupted.

4. The power supply oriented multi-agent collaborative scheduling system according to claim 3, wherein, The dynamic scheduling and routing engine module is also used to perform secondary planning based on task priority, timeliness indicators and fairness weights when multiple tasks compete for the same scheduling resource. If the secondary planning still results in concurrent resource conflicts, the scheduling resources corresponding to each task are re-determined.

5. The power supply oriented multi-agent collaborative scheduling system according to claim 3, wherein, The dynamic scheduling and routing engine module is also used to utilize the master control agent to call multiple business agents in parallel via the A2A protocol to form a distributed task orchestration network.

6. The power supply oriented multi-agent collaborative scheduling system according to claim 3, wherein, The context bus and state management module is also used to feed back the task execution results of each business agent to the power supply station front-end equipment using the master control agent.

7. The power supply oriented multi-agent collaborative scheduling system according to claim 1, wherein, The multi-agent cooperative scheduling system for power supply stations also includes: The monitoring and logging subsystem is used to establish a unique identifier for each task execution, record task input, task execution path, tool calls, execution logs and task execution results, and generate a visual execution chain.

8. The multi-agent cooperative scheduling system for power supply stations according to claim 7, characterized in that, Business intelligence agents include: work order generation intelligence agent, work order scheduling intelligence agent, work order dispatching intelligence agent, strategy generation intelligence agent, work order evaluation intelligence agent, work summary intelligence agent, and service script intelligence agent. The work order generation intelligent agent is used to call the work order information stored in the power supply station's database based on the MCP tool, generate work order information for the task corresponding to the current business request, and send the work order information to the main control intelligent agent. The work order scheduling intelligent agent is used to call the resources stored in the power supply station's database based on the MCP tool, determine the resources that can be called to meet the current business request, and send the callable resources to the main control intelligent agent; The work order dispatching intelligent agent is used to call the resource attribute tags stored in the power supply station's database based on the MCP tool, determine the resources to be dispatched for the task from the currently available resources, and send the resources to be dispatched for the task to the main control intelligent agent. The strategy generation agent is used to call the task response strategies stored in the power supply station's database based on the MCP tool, determine the execution strategy to be used for the current task according to the task response strategy, and send the execution strategy to the master control agent. The work order evaluation agent is used to call the task execution status recorded in the power supply station's database based on the MCP tool, evaluate the task execution status, and send the evaluation results to the main control agent. The work summary agent is used to call the task execution status recorded in the power supply station's database based on the MCP tool, summarize the work based on the task execution status, and send the work summary results to the main control agent; The service script intelligent agent is used to generate service scripts based on the service object tags recorded in the power supply station's database using the MCP tool, and then send the service scripts to the main control intelligent agent.

9. The multi-agent cooperative scheduling system for power supply stations according to claim 8, characterized in that, The business intelligence agent subsystem is also used to optimize the work order scheduling intelligence agent, work order dispatch intelligence agent, and work order strategy intelligence agent based on task execution logs.

10. A scheduling method for a power supply oriented multi-agent collaborative scheduling system, characterized in that, include: The main control intelligent agent is used to identify the intent of the service requests sent by the front-end equipment of the power supply station, plan the whole process tasks that meet the intent identification results, decompose the tasks, and arrange the dependencies between the decomposed tasks. Based on the MCP protocol, the data in the power supply station's database is encapsulated and managed to generate MCP tools; For each decomposition task in the main control intelligent agent subsystem, create a corresponding business intelligent agent; The master control agent uses the A2A protocol to call the corresponding business agent to execute the corresponding dismantling task based on the dependencies between the dismantling tasks. When each business intelligence agent executes its corresponding dismantling task, it calls the MCP tool to read data from the power supply station's database; Each business intelligence agent executes the corresponding decomposition task based on the read data, and feeds back the task execution results to the master intelligence agent based on the A2A protocol; The master control agent feeds back the task execution results of each business agent to the power supply station's front-end equipment.

Images