A multi-agent collaborative analysis and auxiliary treatment method, system, device and storage medium

Abstract

The application discloses a kind of multi-agent collaborative analysis and auxiliary disposal method, system, equipment and storage medium, it is related to artificial intelligence and control operation system technical field.The method includes that electric power real-time business task is parsed, task context is obtained and structured task description is generated;According to structured task description, query business intelligent agent ability directory, tool call contract, preposed constraint template and historical execution record, generate candidate intelligent agent set and corresponding execution link;According to preposed constraint template, execution link is executed before constraint verification: if there is execution link that passes through execution before constraint verification, execution link is issued to corresponding business intelligent agent and tool gateway execution, and execution result is obtained, and execution result is trusted verification;If not, according to task context, determine the risk level of task and trigger hierarchical rollback disposal.The application improves the reliability, reviewability and safety boundary of multi-agent collaborative analysis and auxiliary disposal.

Description

Technical Field

[0001] This invention relates to a method, system, device, and storage medium for multi-agent collaborative analysis and assisted handling, belonging to the technical field of artificial intelligence and control operation systems. Background Technology

[0002] With the advancement of new power system construction, power grid control and operation systems are gradually shifting from a human-computer interaction model primarily based on fixed menus, dedicated interfaces, and manual queries to an intelligent interaction model oriented towards natural language requests from operators. During shifts, monitoring, alarm handling, and auxiliary analysis, operators frequently need to submit tasks such as "inquire about the cause of consecutive over-limit alarms in a certain area this morning," "determine if there is abnormal flickering on the current monitoring screen," and "analyze the possible causes of protection actions on a certain line." These tasks are not simple question-and-answer sessions but typically require the collaborative completion of multiple business intelligence agents and tools, including intelligent query analysis, intelligent alarm processing, screen diagnostics, knowledge retrieval, rule engines, and report generation.

[0003] Real-time power operations are characterized by strong constraints, high reliability, and traceability. A single intelligent interaction is often subject to multiple constraints, including device objects, geographical scope, time windows, operating status, data permissions, professional procedures, tool parameters, return structures, and security boundaries. While general-purpose large language models possess strong natural language understanding capabilities, the lack of path constraints and execution verification specific to power control scenarios may result in unexecutable, unverifiable, or even misleading call chains or handling suggestions.

[0004] Existing multi-agent collaborative technologies typically focus on role segmentation, dialogue collaboration, and task decomposition, which can improve the ability to handle complex tasks in open-domain scenarios. However, in constrained industrial scenarios such as power grid control and operation systems, business agents cannot arbitrarily choose tools and action sequences. An agent that appears semantically matched may still fail to execute due to a lack of equipment objects, time windows, access permissions, or necessary tool parameters; a tool that appears callable may still fail to support the final conclusion because the return structure does not meet the acceptance criteria. Therefore, simply selecting agents or tools based on task semantics cannot guarantee the feasibility of the execution chain.

[0005] Existing rule-based routing solutions typically rely on keywords, business dictionaries, or static dispatch rules. When user expressions contain multiple intents, synonyms, abbreviations, non-standard terminology, or omitting context, rule-based routing is prone to mis-dispatch. For example, alarm cause analysis might be misrouted to a general status query agent, or screen anomaly diagnosis might be misrouted to a general query agent. While existing solutions based on direct routing using large models can recognize some complex semantics, they are prone to continuing tool calls when device objects are missing, time windows are ambiguous, permissions are insufficient, or tool contracts are not met, leading to call failures, empty results, or incomprehensible output.

[0006] Furthermore, the analysis and auxiliary handling in real-time power operations differ in risk levels. Read-only data query tasks can be executed automatically when permissions and parameter conditions are met; however, tasks involving control recommendations, handling points, equipment operation, parameter modification, or operation mode adjustment require more stringent risk identification, manual confirmation, or only auxiliary recommendations. If a multi-agent collaborative system lacks structured clarification, degradation, and manual confirmation mechanisms, it is prone to oscillating between "silence and no response" and "forced conclusions" when information is incomplete, making it difficult to meet the security and verifiability requirements of control operation scenarios.

[0007] Therefore, a multi-agent collaborative analysis and auxiliary handling method for real-time power business is needed, so that the system can not only determine "which agent will execute" and "which tools will be called", but also determine whether the execution link meets the capability constraints, context constraints, permission constraints, tool call contract constraints, return structure constraints and risk control constraints before the action, and provide a controllable fallback path when there is no execution link. Summary of the Invention

[0008] The purpose of this invention is to propose a multi-agent collaborative analysis and auxiliary handling method, system, device and storage medium. By setting a constraint hierarchical routing mechanism between the business intelligence agent and the tool pool, the system completes executability verification before entering the tool call, and triggers rollback processing according to the risk level when the link is unexecutable or the result is unreliable, thereby improving the reliability, interpretability and verifiability of the real-time business intelligence interaction results of the power industry.

[0009] To achieve the above objectives, the present invention is implemented using the following technical solution:

[0010] In a first aspect, the present invention proposes a multi-agent collaborative analysis and auxiliary processing method, comprising:

[0011] Obtain real-time power service tasks, parse the real-time power service tasks, and obtain the task context;

[0012] Generate a structured task description based on the task context;

[0013] Based on the structured task description, query the business intelligence agent capability catalog, tool call contract, pre-constraint template, and historical execution record;

[0014] A candidate agent set is generated based on the structured task description, business agent capability catalog, tool invocation contract, and historical execution records;

[0015] Based on the candidate agent set, the business agent capability catalog, and the tool invocation contract, determine the binding relationship between the candidate agent and one or more invokable tools, and generate the corresponding execution chain;

[0016] Perform pre-execution constraint verification on the execution link based on the aforementioned pre-constraint template:

[0017] If an execution link passes the pre-execution constraint verification, the execution link is sent to the corresponding business intelligence agent and tool gateway for execution, and the execution result is obtained; the execution result is subjected to field integrity verification, consistency verification, credibility verification and interpretation organization to determine whether the execution result is credible, and the credible execution result is recorded;

[0018] If there is no execution link that passes the pre-execution constraint verification or the execution result is unreliable, the risk level of the task is determined according to the task context and a graded rollback is triggered.

[0019] Furthermore, the real-time power business tasks include natural language query tasks, real-time measurement status query tasks, continuous over-limit alarm analysis tasks, protection action cause analysis tasks, control screen anomaly diagnosis tasks, trend analysis tasks, knowledge procedure retrieval tasks, auxiliary handling suggestion generation tasks, or sub-tasks dispatched by other intelligent agents.

[0020] Furthermore, the task context includes natural language input, business events, alarm information, screen diagnostic requests, device objects, time windows, current operating status, task objectives, expected output, access permissions, risk levels, and manual confirmation strategies.

[0021] Furthermore, the business intelligence agent capability catalog includes a description of each business intelligence agent's capabilities, risk costs, applicable business objects, set of callable tools, set of preconditions, output type, historical success rate, and disabling conditions. Among them, each business intelligence agent includes an intelligent data query agent, an intelligent alarm agent, an image diagnosis agent, a knowledge retrieval agent, a recommendation and handling agent, a behavior analysis agent, and a report generation agent.

[0022] Furthermore, a candidate agent set is generated based on the structured task description, the business agent capability catalog, the tool invocation contract, and historical execution records, including:

[0023] Calculate the semantic similarity between the structured task description and the capability descriptions of each business agent; obtain historical success priors based on the historical execution records; generate a comprehensive matching score based on the task objectives and device entity types in the structured task description, the risk costs in the business agent capability catalog, semantic similarity, historical success priors, and context satisfaction; sort each business agent according to the comprehensive matching score to obtain a ranking result; select a preset number of candidate agents from the ranking result to form a candidate agent set.

[0024] Furthermore, the pre-execution constraint verification includes verifying the context integrity, parameter value range, permission conditions, resource reachability, tool call contract, return structure, and risk boundary for each execution link.

[0025] Furthermore, the step of distributing the execution link to the corresponding business intelligence agent and tool gateway for execution, and obtaining the execution result, includes:

[0026] Through the tool gateway, the tools to be executed in the execution chain are called sequentially according to the order of their arrangement. Each tool to be executed satisfies the tool type, input parameters, permission conditions and return structure requirements specified in the corresponding tool call contract.

[0027] After each invocation of the currently executed tool and receipt of the intermediate result returned by the currently executed tool, the intermediate result is written to the task context to update the task context;

[0028] Based on the updated task context, determine the input parameters or calling conditions of the next tool to be executed in the execution chain, and continue to call the next tool to be executed until all tools to be executed in the execution chain have been called;

[0029] The execution result is generated based on the intermediate results returned by all pending tools and the updated task context.

[0030] Further, the step of determining the risk level of the task based on the task context and triggering a tiered rollback process includes:

[0031] When a time window or device object is detected to be missing in the task context, a follow-up question is sent to the user for clarification.

[0032] When the current running state is a normal read-only query, the query is executed using a conservative default window, and the assumptions used are marked in the output results.

[0033] When there is no access permission, the automatic execution of control actions is denied;

[0034] When the risk level exceeds the preset threshold, a manual confirmation prompt, a two-person confirmation prompt, or an auxiliary handling suggestion will be generated.

[0035] Furthermore, the method also includes iteratively optimizing the business intelligence agent capability catalog, tool invocation contract, and historical execution record based on the reliable execution results and feedback from the tiered rollback handling, until a preset iteration termination condition is met.

[0036] Secondly, this invention proposes a multi-agent collaborative analysis and auxiliary processing system, comprising:

[0037] The task access module is used to receive real-time power business tasks;

[0038] The task parsing module parses the real-time power business task to obtain the task context and generates a structured task description based on the task context.

[0039] The data query module is used to query the business intelligence agent capability catalog, tool call contract, pre-constraint template and historical execution record based on the structured task description;

[0040] The candidate generation module is used to generate a set of candidate intelligent agents based on the structured task description, the business intelligent agent capability catalog, the tool invocation contract, and the historical execution record;

[0041] The execution link generation module is used to determine the binding relationship between the candidate intelligent agent and one or more callable tools based on the candidate intelligent agent set, the business intelligent agent capability catalog and the tool invocation contract, and generate the corresponding execution link;

[0042] The pre-execution constraint verification module is used to perform pre-execution constraint verification on the execution link according to the pre-constraint template.

[0043] The collaborative execution module is used to distribute the verified execution links to the corresponding business intelligence agents and tool gateways for execution, and obtain the execution results; perform field integrity verification, consistency verification, credibility verification and interpretation organization on the execution results, determine whether the execution results are credible, and record credible execution results;

[0044] The graded rollback module is used to determine the risk level of an execution link that fails the pre-execution constraint verification or whose execution result is unreliable, and to trigger graded rollback processing based on the task context.

[0045] Thirdly, the present invention proposes a computer device, including a processor, a memory, and a communication interface. The memory stores a computer program. When the computer program is executed by the processor, the computer device executes the computer program to realize the multi-agent collaborative analysis and auxiliary processing method as described above.

[0046] Fourthly, the present invention proposes a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the aforementioned multi-agent collaborative analysis and auxiliary processing method.

[0047] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0048] 1. This invention transforms the role selection and tool invocation in multi-agent collaboration from empirical assignment to calculable and verifiable execution link selection by using a business intelligence agent capability catalog, tool invocation contract, and pre-constraint template, thereby improving the stability of routing decisions.

[0049] 2. This invention moves the verification of context integrity, parameter value range, access permissions, resource reachability, tool contract, return structure, and risk boundary to before tool invocation, which can intercept non-executable paths before entering the execution phase, reducing invalid calls and uninterpretable results.

[0050] 3. This invention forms a three-stage control link through capability coarse screening, pre-execution constraint verification, and hierarchical rollback, which can handle failure modes such as no executable link, missing context, unsatisfied tool contract, and insufficient confidence of boundary samples.

[0051] 4. This invention addresses real-time power business scenarios such as natural language querying, alarm analysis, image diagnosis, and auxiliary handling. It sets differentiated rollback strategies based on task context, business intelligent agent capability catalog, tool call contract, pre-constraint template, historical success priors, and risk level. When information is incomplete or high-risk actions are involved, it prioritizes triggering clarification and follow-up questions, manual confirmation, or only generating auxiliary suggestions, thereby improving the controllability of system behavior boundaries.

[0052] 5. This invention records task input, routing decisions, constraint verification, tool calls, rollback processing, and output results through end-to-end auditing and tracing, enabling the intelligent interaction process of real-time power business to be traceable, verifiable, and continuously optimized. Attached Figure Description

[0053] Figure 1 This is a schematic diagram of the overall structure of the multi-agent collaborative analysis and assisted processing system provided in an embodiment of the present invention;

[0054] Figure 2This is a flowchart of a multi-agent collaborative analysis and auxiliary processing method provided in an embodiment of the present invention;

[0055] Figure 3 A flowchart of the constraint hierarchy routing mechanism provided in this embodiment of the invention;

[0056] Figure 4 A flowchart of tiered rollback and risk control provided for embodiments of the present invention;

[0057] Figure 5 This is a schematic diagram of the alarm analysis and screen diagnosis application link of the power grid operation monitoring system provided in the embodiments of the present invention. Detailed Implementation

[0058] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the present invention or its application or use.

[0059] Example 1:

[0060] This embodiment proposes a multi-agent collaborative analysis and auxiliary processing method, such as... Figure 2 As shown, it includes the following steps:

[0061] S1: Receive real-time power service tasks.

[0062] The system receives one or more of the following tasks from users: natural language questioning tasks, real-time measurement status query tasks, continuous over-limit alarm analysis tasks, protection action cause analysis tasks, control screen anomaly diagnosis tasks, trend analysis tasks, knowledge procedure retrieval tasks, auxiliary handling suggestion generation tasks, and sub-tasks dispatched by other intelligent agents.

[0063] S2: Parse the task context.

[0064] The system performs natural language understanding, entity extraction, time window identification, task target identification, risk level identification, and permission context reading on real-time power business tasks to form a task context. The task context includes natural language input, business events, alarm information, screen diagnostic requests, equipment objects, time windows, current operating status, task targets, expected outputs, access permissions, risk levels, and manual confirmation strategies.

[0065] This data is used to transform users' natural language requests into verifiable, routeable, and executable structured tasks. The general function of each field is as follows:

[0066] Natural language input serves as the source of the original task, such as querying the cause of consecutive out-of-limit alarms in a certain area this morning. Business events are system-triggered event objects, such as alarm events, protection action events, and inspection task events. Alarm information is used to determine whether to enter the alarm analysis chain, including alarm type, level, object, and time. Screen diagnostic requests are used to determine whether to enter the screen diagnostic agent, such as screen flickering, abnormal colors, or abnormal graphic elements. Device objects are used to identify the target equipment for querying, diagnosing, or analyzing, such as lines, buses, switches, and main transformers. Time windows are used to limit the data query range, such as the last 30 minutes, early this morning, or before and after a certain accident. Current operating status is used to determine whether the analysis conclusions meet real-time operating conditions, such as whether the equipment is energized, under maintenance, or locked. Task objectives are used to determine the task type, such as querying data, alarm analysis, screen diagnostics, knowledge retrieval, and auxiliary handling. Expected output is used to determine the result format, such as evidence lists, cause ranking, handling suggestions, and reports. Access permissions are used to determine whether users, agents, or tools have the right to access target data or perform corresponding operations. The risk level determines whether automation is allowed, or whether clarification, manual confirmation, or simply generating suggestions is required. A manual confirmation strategy is used for high-risk tasks, such as determining whether single-person confirmation, dual-person confirmation, or manual takeover is necessary.

[0067] The fields in the task context are not isolated data, but are used for subsequent task structuring, candidate agent matching, toolchain generation, pre-execution constraint verification, and risk rollback control. For example, the device object and time window are used to generate tool call parameters, the task objective and expected output are used to select the business agent, and access permissions and risk level are used for pre-execution verification and rollback strategy determination.

[0068] S3: Generate a structured task description.

[0069] Organize the task context into a structured task description that includes input text, entity set, task objective, current context state, candidate tool types, acceptance criteria, and risk control requirements.

[0070] S4: Query capability catalog and tool call contract.

[0071] Based on the structured task description, query the business intelligence agent capability catalog, tool invocation contract, precondition template, and historical execution record to obtain information on intelligence agents and tools that can participate in this task.

[0072] The various business intelligent agents include intelligent data query agent, intelligent alarm agent, image diagnosis agent, knowledge retrieval agent, recommendation and handling agent, behavior analysis agent, and report generation agent.

[0073] The Business Intelligence Agent Capability Directory includes a description of each agent's capabilities, risk cost, applicable business objects, set of callable tools, set of preconditions, output type, historical success rate, and disabling conditions. The capability description is used to match the agent with the task semantics, determining if the agent is suitable for the current task. The risk cost is used to rank candidate agents, with higher-risk agents or processes requiring more stringent validation. The applicable business objects determine if the agent is suitable for the current device or business object, such as lines, transformers, alarms, and displays. The set of callable tools generates the execution path, determining which tools the agent can subsequently call. The set of preconditions is used for pre-execution constraint validation, determining if the current task meets the calling conditions. The output type determines if the agent's output meets user expectations, such as evidence lists, conclusions, and suggestions. The historical success rate is used for candidate ranking, prioritizing agents with higher success rates in similar tasks. Disabling conditions exclude unsuitable agents, such as those with insufficient permissions, unavailable resources, or those disabled in high-risk scenarios. Each field in the Business Intelligence Agent Capability Directory plays a specific role in subsequent processes. Among them, capability description, applicable business objects and output type are used for matching and screening candidate intelligent agents; risk cost and historical success rate are used for prioritizing candidate intelligent agents; callable tool set is used for generating candidate toolchains; and precondition set and disallowed conditions are used for constraint verification before execution and filtering of non-executable links, respectively.

[0074] A tool invocation contract is a machine-readable constraint description of each invokable tool within the system, equivalent to a tool invocation specification / interface contract. It includes one or more of the following: tool name, tool type, input parameter definitions, required parameters, parameter types, parameter value ranges, business preconditions, access permission conditions, resource reachability conditions, invocation timeout, return field structure, result acceptance conditions, exception codes, and rollback policies. Tool types include data query services, real-time measurement services, historical alarm query tools, rule engines, knowledge retrieval tools, image diagnostic tools, model inference services, file parsing tools, report generation tools, database query interfaces, web service interfaces, or specialized algorithm interfaces. Taking the historical alarm query tool invocation contract as an example, invoking the historical alarm query tool requires providing the device object, time window, and alarm type; the returned result must include fields such as alarm time, device ID, alarm level, and alarm description.

[0075] The pre-constraint template is a template for pre-execution constraint verification rules. It is used to determine whether a candidate execution chain has the necessary conditions for execution before calling the tool, such as whether there is a missing device object, a missing time window, whether the user has the necessary permissions, whether the tool resources are accessible, whether it involves high-risk operations, and whether the returned results have acceptance requirements.

[0076] S5: Generate a set of candidate agents.

[0077] The system statistically analyzes and summarizes historical execution records, extracts historical success priors, and uses them as empirical parameters or scoring factors to participate in the ranking calculation of candidate agents. Simultaneously, it calculates the semantic similarity between the structured task description and the capability descriptions of each business agent; obtains the task objective and device entity type from the structured task description; generates a comprehensive matching score based on semantic similarity, risk cost, historical success priors, task objective, device entity type, and contextual satisfaction; ranks each business agent according to the comprehensive matching score, obtaining the ranking result; and selects a preset number of candidate agents from the ranking result to form a candidate agent set.

[0078] S6: Bind candidate toolchain.

[0079] The system determines the binding relationship between the candidate agent and one or more callable tools based on the candidate agent's capability description, the set of tools that the agent can call, the tool calling contracts of each tool, the current task context, and historical execution records, and generates the corresponding execution chain.

[0080] S7: Constraint verification before execution.

[0081] The execution chain is subjected to pre-execution constraint verification based on the pre-constraint template. The pre-execution constraint verification includes context integrity, parameter value range, permission conditions, resource reachability, tool invocation contract, return structure, and risk boundary verification for each execution chain. Among them, context integrity verification is used to determine whether the device object, region object, time window, and task target are complete; parameter value range verification is used to determine whether the tool parameters meet the requirements of type, range, enumerated value, and required fields; permission condition verification is used to determine whether the user, intelligent agent, and tool have the permission to access the target business data; resource reachability verification is used to determine whether the tool service, knowledge base, real-time data source, and image diagnostic resources are reachable; return structure verification is used to determine whether the tool return structure contains the fields required for result acceptance conditions; and risk boundary verification is used to determine whether the execution chain involves remote control, remote adjustment, setting modification, parameter download, configuration change, or other high-risk actions.

[0082] S8: Execute the verified link.

[0083] When a valid execution chain exists, the system uses the tool gateway to sequentially call the tools to be executed in the execution chain according to their order. Each time the current tool to be executed is called, the system performs analysis, diagnosis, retrieval, reasoning, or report generation operations and writes the generated intermediate results into the task context to update the task context.

[0084] Based on the updated task context, determine the input parameters or calling conditions of the next tool to be executed in the execution chain, and continue to call the next tool to be executed until all tools to be executed in the execution chain have been called;

[0085] The execution result is generated based on the intermediate results returned by all pending tools and the updated task context.

[0086] S9: Verify the execution results.

[0087] The system performs field integrity checks, consistency checks, credibility checks, and interpretation organization on the execution results based on acceptance criteria to determine whether the execution results are credible. When the execution results meet the acceptance criteria, the system generates final analysis results or auxiliary handling suggestions. Credibility checks determine whether the results returned by the tool or the analysis results of the agent are reliable. If the credibility of the results is insufficient, the system will not directly output a definite conclusion, but will trigger a rollback strategy, such as clarification and follow-up questions, manual confirmation, or only generating auxiliary suggestions. Interpretation organization organizes the tool results, evidence, rule hits, and analysis conclusions into an output that can be understood by the operators.

[0088] S10: Perform a tiered rollback.

[0089] When no execution link passes verification or the execution result fails to meet acceptance criteria, the system performs tiered rollback based on risk level and type of missing information:

[0090] When a time window or device object is detected to be missing in the task context, a follow-up question is sent to the user for clarification.

[0091] When the current running state is a normal read-only query, the query is executed using a conservative default window, and the assumptions used are marked in the output results.

[0092] When there is no access permission or the risk level is higher than the preset threshold, the automatic execution of control actions will be refused, and at least one of the following will be generated: manual confirmation prompt, dual confirmation prompt, and auxiliary handling suggestions.

[0093] Step S11: Output the results and audit the results.

[0094] The system outputs natural language interpretation, structured analysis results, auxiliary handling suggestions, evidence lists or reports, and records end-to-end audit data.

[0095] Step S12: Update historical priors and configurations.

[0096] The system updates historical execution records, capability catalogs, tool call contracts, and rollback rules based on task success indications, user feedback, manual confirmation results, tool call counts, execution latency, and risk interception results.

[0097] Example 2:

[0098] This embodiment proposes a multi-agent collaborative analysis and auxiliary processing system, such as... Figure 1 As shown, the system includes a task access module, a task parsing module, a data query module, a candidate generation module, an execution link generation module, a pre-execution constraint verification module, a collaborative execution module, a result verification module, a tiered rollback module, and an audit trail module. These modules can be deployed on the intelligent interactive service side of the power grid control and operation system, or incrementally deployed as a collaborative control layer between the decision center and the tool pool.

[0099] The task access module is used to receive real-time power business tasks. Tasks can originate from operators' natural language text, speech-to-text transcription, alarm events, control screen diagnostic requests, shift workflow nodes, scheduled inspection tasks, or sub-tasks dispatched by other intelligent agents. Task formats can include natural language, structured event objects, alarm objects, screen screenshot identifiers, device entity sets, or workflow context.

[0100] The task parsing module is used to parse real-time power business tasks to obtain the task context. The task context includes at least the task objective, business scenario, equipment object, area scope, time window, current operating status, context status, data permissions, expected output, task priority, risk level, whether manual confirmation is required, and whether automatic execution is allowed. For example, for a request to analyze the causes of consecutive out-of-limit alarms in a certain area this morning and provide key handling points, the task parsing module can extract the area object, time window, alarm analysis objective, handling suggestion output requirements, and a risk level higher than that of a normal question.

[0101] The data query module is used to query the business intelligence agent capability catalog, tool invocation contracts, precondition templates, and historical execution records. The business intelligence agent capability catalog can include intelligent data query agents, intelligent alarm agents, image diagnostic agents, knowledge retrieval agents, recommendation and handling agents, behavior analysis agents, and report generation agents; each business intelligence agent corresponds to a capability description, applicable task type, set of callable tools, set of preconditions, output type, risk cost, historical success rate, and disabling conditions. The tool invocation contract describes the executable boundaries of each tool. Taking the historical alarm query tool as an example, its tool invocation contract can include the tool name, input parameters, required parameters, time window format, device object type, access permission conditions, return field structure, timeout policy, exception code, and acceptance conditions. Taking the image diagnostic tool as an example, its tool invocation contract can include image identifier, region coordinates, diagnostic type, time window, image frame sampling rules, return result fields, confidence threshold, and exception prompt format.

[0102] The candidate generation module generates a set of candidate agents based on the structured task description. It comprehensively calculates a matching score by considering the similarity between the task semantics and capability description, the matching degree between the device entity type and the applicable objects of the agent, historical success priors, contextual satisfaction, and risk cost. The module then selects the Top-k candidate agents based on this score. The purpose of the candidate generation module is to narrow down the execution scope, reducing the search space required for subsequent verification and tool binding.

[0103] The execution link generation module is used to determine the binding relationship between the candidate intelligent agent and one or more callable tools based on the candidate intelligent agent set, the business intelligent agent capability catalog and the tool invocation contract, and generate the corresponding execution link;

[0104] The pre-execution constraint verification module is used to verify candidate agents and their toolchains item by item. The verification content includes: whether the device object, area range, time window and task target are complete; whether the tool input parameters meet the requirements of type, range and enumeration value; whether the user, agent and tool have the right to access the target business data; whether the real-time measurement service, historical alarm service, rule base, knowledge base and image diagnostic resources are accessible; whether the tool return structure contains the fields required for acceptance conditions; whether the execution link involves high-risk actions such as remote control, remote adjustment, setting modification, parameter download or configuration change.

[0105] The collaborative execution module distributes the execution chain, which has passed pre-execution constraint verification, to the business agent and the tool gateway. The tool gateway encapsulates data query services, historical alarm query tools, real-time status retrieval tools, rule engines, knowledge retrieval tools, image diagnostic tools, file parsing tools, and report generation tools. The return result of each tool call is written into the task context and used as input conditions for subsequent agents or tools.

[0106] The results verification module is used to verify the tool's returned results and final output. Verification includes field completeness, business consistency, result reliability, coverage of key evidence, and whether the output structure meets user expectations. For alarm analysis tasks, the results verification module checks whether the output includes an alarm summary, a ranking of possible causes, related measurement evidence, rule matching criteria, and key handling points. For screen diagnosis tasks, the results verification module checks whether the output includes the screen area, observation period, diagnostic conclusion, confidence level, and suggested verification items.

[0107] The hierarchical rollback module is used to determine the handling strategy for execution links that fail the pre-execution constraint verification or whose execution results do not meet the verification requirements, based on the task context; and to generate corresponding handling outputs based on the determined handling strategy.

[0108] The audit trail module records data throughout the entire process. Recorded content may include natural language input, structured task descriptions, candidate agent scores, reasons for failed validations, tool call parameters, tool return structures, result acceptance status, rollback processing details, manual confirmation records, output conclusions, key screenshots, log identifiers, and audit serial numbers. These records can be used for post-event review, accountability tracking, parameter tuning, and capability catalog updates.

[0109] Example 3:

[0110] This embodiment, based on Embodiment 2, provides a specific constraint-level routing mechanism. For example... Figure 3 As shown.

[0111] Let the real-time power service task be represented as ,in, For natural language input or business event content, This is a collection of entities such as equipment entities, region entities, and time windows. For the current context state, For the mission objective, Risk level. The set of business intelligence agents is... The toolset is Each business intelligence agent Corresponding ability description Precondition set Callable tool subset and risk costs .

[0112] The comprehensive matching score of candidate agents can be expressed as:

[0113]

[0114] In the formula, This indicates the semantic similarity between task and capability descriptions. This indicates that historical success was a priori. This indicates the degree of matching between the entity and the applicable objects of the capability. This indicates whether the context satisfies the preconditions. - The scores are set with preset weights. The system uses these scores to form a set of candidate agents.

[0115] For each candidate agent, the system uses a subset of callable tools. Generate candidate execution links with task objective o Candidate execution chains Capability constraints, context constraints, permission constraints, and tool invocation contract constraints must be satisfied. The system can represent the executability of the execution chain as follows:

[0116]

[0117] In the formula, This indicates the result of the context integrity check. This indicates the result of the parameter value range verification. This indicates the permission verification result. This indicates the resource reachability verification result. This indicates that the tool calls the contract verification result. This indicates that the structure verification result is being returned. This indicates the result of the risk boundary verification. If any verification result is 0, the execution link must not proceed to the tool execution phase.

[0118] The key to the above mechanism lies in moving the tool executability check forward to the routing decision stage. Even if a business intelligence agent is highly semantically matched to the task, if its toolchain lacks necessary parameters, permission conditions are not met, resources are unreachable, the return structure does not meet acceptance criteria, or the risk boundary does not allow automatic execution, the system will not directly call the tool, but will continue to check the next candidate link or enter a tiered rollback process.

[0119] In engineering implementation, the candidate generation complexity can be approximated as: The constraint verification complexity of Top-k candidates is O(n). , where q is the average number of constraints in the candidate execution link. Since k in real-time power business can usually be a small constant, the overall complexity is approximately linearly related to the number of business agents and the number of constraint templates, making it suitable for deployment between the intelligent interactive decision-making center and tool pool of the power grid operation monitoring system.

[0120] Example 4:

[0121] This embodiment, based on Embodiment 2, uses the power grid OCS alarm analysis and image diagnosis service as an example to illustrate the application process of the multi-agent collaborative analysis and auxiliary handling method in real-time power services. Figure 5 As shown.

[0122] After the operators issue an analysis request, the system extracts entities and time windows through NLU (Natural Language Understanding) parsing. Then, through SOP-constrained hierarchical routing (SCHR), it simultaneously activates two intelligent agent links: alarm analysis and image diagnosis. The alarm link retrieves historical alarms, reads real-time measurements, and matches them with the rule base, prioritizing "single-phase grounding of adjacent lines" as the most likely cause and providing key handling points. The image link synchronously retrieves monitoring images, identifies voltage curve inflection points and high-temperature areas in the heat map through image diagnosis, and generates abnormal evidence and verification suggestions. Finally, the evidence lists output by the two links, including protection action records, over-limit data, and annotated screenshots, are integrated into a unified analysis conclusion, along with auxiliary handling suggestions such as "outage fault interval, switch to backup power supply." The entire process is audited, enabling manual fault assessment and significantly improving the efficiency and reliability of dispatching and handling.

[0123] Example 5:

[0124] This embodiment, based on Embodiment 2, uses an alarm analysis scenario from a power grid operation monitoring system as an example to illustrate the application process of the multi-agent collaborative analysis and auxiliary handling method in real-time power services.

[0125] In this embodiment, the operator inputs "Analyze the possible causes of continuous over-limit alarms in a certain area this morning and provide key points for handling".

[0126] After receiving the request, the task parsing module extracts the object entity "a certain area", the time window "this morning", the task objective "alarm analysis + auxiliary handling suggestions" and the risk level "analysis and suggestion class".

[0127] The candidate generation module calculates candidate agent scores based on task semantics and the capability catalog. The intelligent alarm agent receives a high score due to the dual semantic matching of "alarm" and "analysis," while the intelligent query agent receives the second-highest score due to matching "region" and "query." The screen diagnosis agent and the report generation agent are excluded or ranked lower due to their large semantic distance. The system thus narrows down the candidate set to at least the number of business agents most likely to perform the task.

[0128] The pre-execution constraint verification module first checks the preconditions of the intelligent alarm agent, including whether the historical alarm chain is reachable in recent times, whether the real-time measurement data of the corresponding area is available, whether the relevant rule base entries exist, and whether the user has permission to read the operational data of the area. After the verification is passed, the system binds the intelligent alarm agent to a toolchain of "historical alarm query + real-time status retrieval + knowledge rule retrieval" and performs contract verification on each tool.

[0129] The tool gateway sequentially calls the historical alarm query tool, real-time status retrieval tool, and knowledge rule retrieval tool, returning out-of-limit records, real-time measurement data, and rule matching results. The result verification module checks whether the returned fields include the alarm occurrence time, device object, associated measurement, rule hit item, and handling points. After the consistency check passes, it organizes the data into a structured output of "alarm summary, possible cause ranking, evidence list, and auxiliary handling points".

[0130] like Figure 4 As shown, if the user does not provide a specific time window, the system will not directly assume a default time range and give a definitive conclusion. Instead, it will trigger clarification and follow-up questions based on the risk level, such as prompting the user to confirm the period of concern. For read-only tasks that allow conservative queries, the system can use a preset default window to execute the query, but will clearly indicate the assumptions in the output. For tasks with medium-risk analysis recommendations, the system requests confirmation to limit the scope of the conclusion. For tasks involving high-risk handling, the system refuses to automatically execute control actions and generates prompts for manual confirmation and double confirmation. For structured output with insufficient credibility, the system will first attempt to retry parsing; if the retry fails, the task will be transferred to manual management.

[0131] Example 6:

[0132] This embodiment, based on Embodiment 2, uses a power grid operation monitoring system control screen diagnosis scenario as an example to illustrate the application process of the multi-agent collaborative analysis and auxiliary handling method in real-time power services.

[0133] In a scenario involving image diagnostics, a user inputs "determine if there is abnormal flickering in a certain area of ​​the currently monitored image." This task description is vague and may lack precise area numbers, observation time windows, and image identifiers. If rule-based routing is used, the task may be incorrectly assigned to a general-purpose query agent because "abnormal flickering" is not a standard keyword. If a large model is used for direct routing, the image diagnostic tool may be called directly, resulting in insufficient parameters or returning empty results.

[0134] The constraint-hierarchical routing mechanism in this invention first identifies the task as a high-uncertainty image diagnosis task and includes the image diagnosis agent in the candidate set. The pre-execution constraint verification module then checks the image diagnosis tool call contract. If it finds a missing region number and observation time window, it determines the execution link is unexecutable and triggers clarification follow-up questions.

[0135] After the user provides the region ID and observation period, the system regenerates a structured task description and binds it to the image diagnostic agent and image diagnostic tool. The tool gateway invokes image diagnostic capabilities based on the image identifier, region coordinates, and observation period, returning suspected flickering events, confidence levels, and keyframe information. The result verification module checks whether the returned structure includes event time, region location, anomaly type, confidence level, and suggested verification items, and organizes it into interpretable output.

[0136] This embodiment demonstrates that in scenarios with incomplete parameters or high task uncertainty, "clarification before execution" is more reliable than directly invoking tools. Through a tiered rollback mechanism, the system can maintain controllable behavioral boundaries when information is incomplete, avoiding silence or unresponsiveness or forcibly providing unverifiable conclusions.

[0137] Example 7:

[0138] This embodiment, building upon Embodiment 2, uses anonymized real-world processes and template-enhanced samples to construct an evaluation set for verifying the technical effectiveness of the constraint hierarchy routing mechanism. The evaluation samples can cover task types such as state query count, alarm analysis, and screen diagnostics. Each sample is labeled with the target agent, the toolchain to be invoked, key pre-parameters, and result acceptance conditions. Evaluation metrics may include task success rate, routing accuracy, tool invocation success rate, average end-to-end latency, and average number of tool invocations.

[0139] In an exemplary test, the full-constraint hierarchical routing mechanism achieved a 93.3% task success rate, 96.7% routing accuracy, 98.3% tool invocation success rate, and an average end-to-end latency of approximately 319.4 ms on 100 anonymized test samples. Compared to random routing, rule-based routing, and direct routing for large models, its performance improvement primarily stems from improved preferred route quality and interception of non-executable paths, rather than relying on more tool retries.

[0140] In one ablation test, the task success rate was 80.0% when using only the large model for direct routing; after adding capability screening, the success rate increased to 86.7%; after further adding pre-execution constraint verification, the success rate increased to 90.0%; and after adding tiered rollback, the success rate increased to 93.3%. These results indicate that capability screening, pre-execution constraint verification, and tiered rollback make independent contributions to role selection, tool executability, and boundary sample recovery, respectively.

[0141] The above values ​​are merely an example of desensitization to illustrate the technical effects of this embodiment and do not constitute a limitation on the scope of protection of the claims. Those skilled in the art can adjust the number of Top-k, scoring weights, risk thresholds, default time windows, clarification strategies, and acceptance criteria according to the actual system scale, number of business intelligence agents, tool pool size, response latency requirements, and security strategies.

[0142] Example 8:

[0143] As a further extension of the above embodiments, in other embodiments, the real-time power business tasks in this embodiment are not limited to data collection, alarms, screen diagnosis and auxiliary handling in the power grid control and operation system, but can also be applied to scenarios such as equipment operation and maintenance, protection fault diagnosis, operation mode verification, report generation, inspection analysis, defect judgment, dispatching knowledge retrieval and operation risk warning.

[0144] The business intelligence agents in this embodiment are not limited to intelligent data query agents, intelligent alarm agents, screen diagnosis agents, knowledge retrieval agents, and recommendation and handling agents. They may also include simulation calculation agents, file parsing agents, defect analysis agents, report generation agents, operation mode verification agents, equipment ledger agents, and third-party professional algorithm agents.

[0145] The tool invocation contract in this embodiment is not limited to Application Programming Interface (API), database interface, and Web service, but can also be applied to Model Context Protocol (MCP) service, Command-Line Interface (CLI) commands, Remote Procedure Call (RPC) interface, message queue, file interface, local script, vendor software development kit (SDK), plugin interface, image recognition service, knowledge graph query interface, rule engine interface, and professional algorithm service.

[0146] The candidate generation method in this invention can be implemented by rule matching, vector retrieval, domain knowledge graph matching, large model semantic scoring, historical success rate statistics, reinforcement learning strategies, manual configuration, or a combination of multiple methods; the pre-execution constraint verification can be implemented by static configuration verification, runtime probing, permission system query, tool contract parsing, return structure pattern matching, rule engine verification, or manual confirmation.

[0147] The tiered rollback method in this embodiment can be configured according to different business procedures. For low-risk read-only tasks, a conservative query can be performed using a default time window or default object range; for medium-risk analysis tasks, clarification and follow-up questions can be triggered first; for high-risk auxiliary handling or tasks that may affect production operations, automatic execution can be prohibited, and only suggestions or requests for manual confirmation, dual confirmation, or electronic signature confirmation can be generated.

[0148] This embodiment is not only applicable to power control and operation systems, but also to industrial scenarios such as oil and petrochemical, rail transit, water conservancy and hydropower, industrial control, energy management and manufacturing, which have requirements for real-time business analysis, multi-agent collaboration, tool pool invocation and high reliability.

[0149] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0150] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0151] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0152] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0153] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above, which are merely illustrative and not restrictive. Those skilled in the art, upon learning of the present invention, can make various modifications without departing from the spirit and scope of the claims, and all such modifications fall within the scope of protection of the present invention.

Claims

1. A method for multi-agent collaborative analysis and assisted handling, characterized in that, include: Obtain real-time power service tasks, parse the real-time power service tasks, and obtain the task context; Generate a structured task description based on the task context; Based on the structured task description, query the business intelligence agent capability catalog, tool call contract, pre-constraint template, and historical execution record; A candidate agent set is generated based on the structured task description, business agent capability catalog, tool invocation contract, and historical execution records; Based on the candidate agent set, the business agent capability catalog, and the tool invocation contract, determine the binding relationship between the candidate agent and one or more invokable tools, and generate the corresponding execution chain; Perform pre-execution constraint verification on the execution link based on the aforementioned pre-constraint template: If an execution link passes the pre-execution constraint verification, the execution link is sent to the corresponding business intelligence agent and tool gateway for execution, and the execution result is obtained; the execution result is subjected to field integrity verification, consistency verification, credibility verification and interpretation organization to determine whether the execution result is credible, and the credible execution result is recorded; If there is no execution link that passes the pre-execution constraint verification or the execution result does not meet the verification, the risk level of the task is determined according to the task context and a graded rollback is triggered.

2. The multi-agent collaborative analysis and assisted processing method according to claim 1, characterized in that, The real-time power business tasks include natural language query tasks, real-time measurement status query tasks, continuous over-limit alarm analysis tasks, protection action cause analysis tasks, control screen anomaly diagnosis tasks, trend analysis tasks, knowledge procedure retrieval tasks, auxiliary handling suggestion generation tasks, or sub-tasks dispatched by other intelligent agents.

3. The multi-agent collaborative analysis and assisted processing method according to claim 1, characterized in that, The task context includes natural language input, business events, alarm information, screen diagnostic requests, device objects, time windows, current operating status, task objectives, expected output, access permissions, risk levels, and manual confirmation strategies.

4. The multi-agent collaborative analysis and assisted processing method according to claim 3, characterized in that, The business intelligence agent capability catalog includes a description of each business intelligence agent's capabilities, risk costs, applicable business objects, set of callable tools, set of preconditions, output type, historical success rate, and disabling conditions. The business intelligence agents include intelligent data query agent, intelligent alarm agent, screen diagnosis agent, knowledge retrieval agent, recommendation and handling agent, behavior analysis agent, and report generation agent.

5. The multi-agent collaborative analysis and assisted processing method according to claim 4, characterized in that, The step of generating a candidate agent set based on the structured task description, business agent capability catalog, tool invocation contract, and historical execution records includes: Calculate the semantic similarity between the structured task description and the capability descriptions of each business agent; obtain historical success priors based on the historical execution records; generate a comprehensive matching score based on the task objectives and device entity types in the structured task description, the risk costs in the business agent capability catalog, semantic similarity, historical success priors, and context satisfaction; sort each business agent according to the comprehensive matching score to obtain a ranking result; select a preset number of candidate agents from the ranking result to form a candidate agent set.

6. The multi-agent collaborative analysis and assisted processing method according to claim 1, characterized in that, The pre-execution constraint verification includes verifying the context integrity, parameter value range, permission conditions, resource reachability, tool call contract, return structure, and risk boundary for each execution link.

7. The multi-agent collaborative analysis and assisted processing method according to claim 1, characterized in that, The step of sending the execution link to the corresponding business intelligence agent and tool gateway for execution and obtaining the execution result includes: Through the tool gateway, the tools to be executed in the execution chain are called sequentially according to the order of their arrangement. Each tool to be executed satisfies the tool type, input parameters, permission conditions and return structure requirements specified in the corresponding tool call contract. After each invocation of the currently executed tool and receipt of the intermediate result returned by the currently executed tool, the intermediate result is written to the task context to update the task context; Based on the updated task context, determine the input parameters or calling conditions of the next tool to be executed in the execution chain, and continue to call the next tool to be executed until all tools to be executed in the execution chain have been called; The execution result is generated based on the intermediate results returned by all pending tools and the updated task context.

8. The multi-agent collaborative analysis and auxiliary processing method according to claim 3, characterized in that, The step of determining the risk level of the task based on the task context and triggering a tiered rollback process includes: When a time window or device object is detected to be missing in the task context, a follow-up question is sent to the user for clarification. When the current running state is a normal read-only query, the query is executed using a conservative default window, and the assumptions used are marked in the output results. When there is no access permission, the automatic execution of control actions is denied; When the risk level exceeds the preset threshold, a manual confirmation prompt, a two-person confirmation prompt, or an auxiliary handling suggestion will be generated.

9. The multi-agent collaborative analysis and assisted processing method according to claim 1, characterized in that, The method further includes iteratively optimizing the business intelligence agent capability catalog, tool invocation contracts, and historical execution records based on the reliable execution results and feedback from the tiered rollback handling, until a preset iteration termination condition is met.

10. A multi-agent collaborative analysis and auxiliary processing system, characterized in that, include: The task access module is used to receive real-time power business tasks; The task parsing module parses the real-time power business task to obtain the task context and generates a structured task description based on the task context. The data query module is used to query the business intelligence agent capability catalog, tool call contract, pre-constraint template and historical execution record based on the structured task description; The candidate generation module is used to generate a set of candidate intelligent agents based on the structured task description, the business intelligent agent capability catalog, the tool invocation contract, and the historical execution record; The execution link generation module is used to determine the binding relationship between the candidate intelligent agent and one or more callable tools based on the candidate intelligent agent set, the business intelligent agent capability catalog and the tool invocation contract, and generate the corresponding execution link; The pre-execution constraint verification module is used to perform pre-execution constraint verification on the execution link according to the pre-constraint template. The collaborative execution module is used to distribute the verified execution links to the corresponding business intelligence agents and tool gateways for execution, and to obtain the execution results; The result verification module is used to perform field integrity verification, consistency verification, credibility verification, and interpretation organization on the execution result, determine whether the execution result is credible, and record the credible execution result; The graded rollback module is used to determine the risk level of an execution link that fails the pre-execution constraint verification or whose execution result does not meet the verification requirements, and to trigger graded rollback processing based on the task context.

11. A computer device, characterized in that, The device includes a processor, a memory, and a communication interface. The memory stores a computer program, which, when executed by the processor, causes the computer device to perform the multi-agent collaborative analysis and assisted processing method according to any one of claims 1 to 9.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the multi-agent collaborative analysis and assisted handling method as described in any one of claims 1 to 9.

Images