Flood control decision-making method and system based on multi-agent cooperation

Abstract

The invention discloses a flood control decision-making method and system based on multi-agent cooperation, and relates to the technical field of artificial intelligence and water conservancy information.The flood control decision-making method comprises the steps that a decision-making thinking chain template guided by domain business logic is constructed, and the flood control decision-making process originally depending on artificial experience is converted into an interpretable and traceable structured reasoning process; modularized disassembly is carried out on complex tasks in flood control business to form a standardized decision sequence, and a business framework is provided for the reasoning process of a large language model, so that each step of reasoning must be aligned with a preset decision target and call a specified knowledge demand; therefore, the common logic jump and constraint omission problems of a general large model in a professional scene are completely eradicated from the working principle, when a system processes user query, automatic matching is carried out, step-by-step advancing is carried out along the thinking chain template, and an intermediate result highly consistent with business operation is generated; effective alignment of artificial intelligence reasoning and domain expert thinking modes is realized.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and water conservancy information technology, specifically to a flood control decision-making method and system based on multi-agent collaboration. Background Technology

[0002] In recent years, the water conservancy industry has faced multiple challenges such as fluctuations in water resources in river basins, frequent extreme floods, and aging infrastructure. Existing intelligent auxiliary decision-making tools have lagged behind in emergency response and dispatch management, becoming a major bottleneck restricting flood control efficiency and safety. Although deep learning methods have achieved certain results in runoff prediction and flood risk assessment, the overall technical path is still at the stage of single model-driven, failing to fully realize semantic understanding and knowledge integration across data sources.

[0003] The emergence of large language models has provided new ideas for solving the above problems. Existing research has applied LLM to scenarios such as hydrological data aggregation, emergency protocol interpretation, and decision support. Some works have attempted to combine knowledge graphs and decision models to enhance reasoning capabilities. For example, the E-KELL system integrates LLM and flood control knowledge graphs to improve the accuracy and readability of emergency plan generation. A flood risk communication assistant based on GPT-4 has improved the public's understanding and acceptance of risk information. However, existing technologies still have key shortcomings: classic CoT (Cooperative Reasoning) prompts lack consistency and traceability in terms of business professionalism; tree-based multi-thinking methods have not solved the problems of business logic embedding and process consistency; and there is insufficient understanding of business chains and multi-step joint reasoning capabilities.

[0004] Therefore, there is an urgent need for a flood control decision-making method and system that can integrate domain knowledge, business logic, and intelligent reasoning. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a flood control decision-making method and system based on multi-agent collaboration. It transforms the flood control decision-making process, which originally relied on human experience, into an interpretable and traceable structured reasoning process by constructing a decision-making thought chain template guided by domain business logic. This modularly decomposes complex tasks in flood control operations, forming standardized decision sequences and providing a business framework for the reasoning process of large language models. Each step of the reasoning must align with the preset decision-making objectives and invoke specified knowledge requirements, thus eliminating the logical jumps and constraint omissions common in general-purpose large models in professional scenarios. When the system processes user queries, it automatically matches and progresses step-by-step along this thought chain template, generating intermediate results highly consistent with business operations. This effectively aligns artificial intelligence reasoning with the thinking patterns of domain experts, ensuring the business rationality and logical closure of the decision-making path.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: On one hand, a flood control decision-making method based on multi-agent collaboration, the specific steps of which are as follows: Design Phase: Based on domain business logic, the complex flood control task T is modularly decomposed into an ordered decision sequence. A standardized decision-making thought chain template is constructed, and the knowledge requirements driven by the decision sequence S are clearly defined. These are then mapped to the corresponding document collections to form a corresponding knowledge system and knowledge base; Agent construction phase: For each decision sequence Develop specialized intelligent agents to form an ensemble of intelligent agents. Each intelligent agent Bind a dedicated knowledge sub-repository within the knowledge base Business Arrangement Operating boundary Data Standards And expressed as: ; Agent scheduling and collaborative reasoning phase: First, the input task... The task decomposition tree is obtained through parsing and identification. For single-domain sub-tasks The response result is generated by calling a function. For complex tasks that span multiple domains and stages, construct a directed task graph. and according to the dependency edges The constraint sequence calls multiple agents to achieve hierarchical collaborative reasoning; Intelligent application stage: Receive user input questions and automatically initiate the reasoning process through a large language model combined with a preset decision-making thought chain template.

[0007] Furthermore, in the design phase, each decision sequence Including reasoning objectives Key constraints and decision output This process can be represented as: Based on the input task description, it can automatically match a decision-making thought chain template using the semantic understanding capabilities of LLM.

[0008] Furthermore, in the design phase, a decision-making thought chain is employed. To guide, each sequence is clearly defined. Required knowledge This information is then mapped to the corresponding document collections, forming a structured, semantically schedulable knowledge system and knowledge base, and then the reasoning objectives are established. Knowledge needs The document-related ternary mapping relationship is then updated to the decision sequence expression. It adopts the pyramid method of "document encoding + structure encoding" and combines the mainstream semantic model bge-m3 encoding model and bge-reranker-large reordering model to achieve a hybrid retrieval that combines full-text retrieval and vector retrieval in parallel.

[0009] Furthermore, in the intelligent agent construction phase, for complex business... The multidimensional characteristics, centered around the reasoning goal Constructing Domain Agents, denoted as Each agent Focus on specific decision sequences and reasoning target Design corresponding reasoning strategies To achieve specific actions, and to ensure that the agent possesses both independent domain capabilities and is easy for system scheduling, the agent's operational boundaries and input / output standards are explicitly introduced as parameters, and the agent's behavior is abstracted as follows: ,in, This refers to specific business orchestration, including mechanism model invocation, knowledge base invocation, and business boundaries and input / output format standards; The operational boundaries of a single intelligent agent include permissions and scope of invocation, triggering conditions, spatiotemporal constraints, and security restrictions. The requirements for input data, completion rules, output format, and response time should be uniformly described.

[0010] Furthermore, in the agent scheduling and collaborative reasoning phase, the task decomposition tree... First, semantic parsing and intent recognition are performed to obtain... Each subtask Corresponding reasoning target Decision output and knowledge needs ,Right now .

[0011] Furthermore, the aforementioned single-domain subtask By calling the function Automatically match and drive the corresponding specialized intelligent agents Execute this subtask to generate and output intermediate inference results. .

[0012] Furthermore, for complex cross-domain tasks, a directed task graph is constructed based on task dependencies. ,in, For a set of task nodes and , Let be the set of dependency edges between tasks and And it calls multiple agents according to the constraint order of dependency edge E to achieve hierarchical collaborative reasoning: ,in, Indicates the input task The overall result of the multi-agent collaborative reasoning process is an abstract description of the multi-agent collaborative reasoning in the entire complex flood control task, reflecting the final output of multiple agents collaborating to complete the task according to business logic. The union operation of sets represents the integration of the calling processes of all agents that satisfy dependencies, forming a complete collaborative reasoning chain. and It is a subtask in the task decomposition tree. For task node set The first in A prerequisite subtask It depends on The subsequent subtasks, It is a set of dependency edges between tasks. Subtasks and There are directed edges that depend on the order of completion, meaning that the task must be completed first. Only then can it be executed , Indicates calling the responsible subtask intelligent agents To execute subtasks And obtain the result after the agent executes the command. It focuses on sub-tasks A specialized intelligent agent possesses a dedicated knowledge base, business rules, and reasoning strategies for handling this sub-task. This indicates that the subtask has been completed. Based on this, call the subtask responsible intelligent agents To execute subtasks And obtain the results.

[0013] Furthermore, in the intelligent application stage, automated flood control decision-making is achieved based on user input, including thought chain inspiration, agent scheduling, and result output, wherein: During the thought chain inspiration stage, after the user inputs a question, the platform uses a large model combined with a pre-set decision thought chain to analyze the user's intent, clarify the core needs, and decompose the needs into a sequence of sub-tasks that conform to the logic of the thought chain. During the agent scheduling phase, agents are invoked through the Multi-Agent Coordination Protocol (MCP), and subtasks are assigned... With context state Mapped to a specific executable agent calling function: That is, while ensuring context consistency, tasks are assigned to appropriate agents. Based on this, automatic scheduling of multiple agents is completed according to the priority of sub-tasks; In the result output stage, the platform extracts key information from the user input, automatically completes parameter format conversion and necessary parameter completion according to the input standards of the corresponding intelligent agent, and after each intelligent agent completes the sub-task reasoning, it summarizes the intermediate results and integrates them into a complete decision plan according to the flood control business logic, and outputs the final answer.

[0014] On the other hand, a flood control decision-making system based on multi-agent collaboration includes: The flood control problem input module is used to collect user problems, clean and organize the input text, and extract core keywords and phrases by combining an enhanced model in the field of flood control. The flood control knowledge retrieval module is used to break down user questions into entity-relation-attribute triples, and then use a hybrid retrieval strategy to combine semantic weight and business weight to rearrange and filter relevant knowledge bases. The flood control thinking chain inspiration module is used to fine-tune and train the large language model based on flood control task thinking chain samples annotated by experts. It combines the knowledge base and user questions to generate multi-step reasoning thinking chains and verifies the logical consistency through the rule engine. The flood control multi-agent collaboration module is used to build multiple domain agents. Each agent has a built-in knowledge sub-base, business rules and reasoning strategies, and the sub-tasks are broken down according to the thinking chain and assigned to the corresponding agents for execution. The flood control answer output module is used to transform the results of multi-agent collaboration into a structured form that includes problem analysis, scheduling suggestions, and visual charts, and outputs them through web pages, mobile devices, flood control command screens, and API interfaces.

[0015] Compared with existing technologies, this flood control decision-making method based on multi-agent collaboration has the following advantages: I. This invention transforms the flood control decision-making process, which originally relied on human experience, into an interpretable and traceable structured reasoning process by constructing a decision-making thought chain template guided by domain business logic. It modularizes complex tasks in flood control operations, forming standardized decision sequences. This provides a business framework for the reasoning process of large language models, ensuring that each step of the reasoning aligns with the preset decision objectives and invokes specified knowledge requirements. This fundamentally eliminates the logical jumps and constraint omissions common in general-purpose large models in specialized scenarios. When the system processes user queries, it automatically matches and progresses along this thought chain template, generating intermediate results highly consistent with business operations. This effectively aligns artificial intelligence reasoning with the thinking patterns of domain experts, ensuring the business rationality and logical closure of the decision path.

[0016] Second, this invention achieves dynamic integration and on-demand retrieval of flood control multi-source knowledge and professional models by constructing a modular domain intelligent agent system and combining it with a task graph scheduling mechanism. Based on the sub-tasks decomposed from the decision-making chain, the corresponding specialized intelligent agents are activated. Each intelligent agent is deeply bound to its exclusive knowledge sub-base and reasoning strategy. For complex tasks involving multiple links and cross-domains, the system can automatically construct a directed task graph and strictly follow the business logic sequence defined by the dependency edges to schedule the work of multiple intelligent agent sequences through a cooperative function. This mechanism enables previously isolated capability units such as rainfall analysis, water level simulation, and scheduling decisions to form an organic collaborative network around a complex flood control decision-making task, realizing a paradigm shift from single model response to multi-agent collaborative operations. This ensures seamless connection and efficient flow of cross-domain knowledge and model computing power in complex business chains at the system level.

[0017] Therefore, the method of this invention not only ensures the traceability and interpretability of decisions through the thought chain mechanism, but also improves the efficiency of knowledge retrieval and task processing through multi-agent collaboration, ultimately achieving the accuracy and reliability of flood control decision recommendations.

[0018] Other advantages, objectives and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination or study, or may be learned from the practice of the invention. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0020] Figure 1 This is a flowchart of the design phase in Example 1; Figure 2 This is a flowchart of the operation during the agent construction phase in Example 1; Figure 3 This is a flowchart illustrating the operation of agent scheduling, collaborative reasoning, and intelligent applications in Example 1. Figure 4 This is an architecture diagram of a flood control decision-making system based on multi-agent collaboration. Detailed Implementation

[0021] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0022] Example 1 This embodiment addresses urban watershed flood control decision-making scenarios by applying a multi-agent collaborative flood control decision-making method. Through business logic decomposition and knowledge base construction in the design phase, specialized function development in the agent construction phase, and large-scale language model reasoning and multi-agent collaborative scheduling in the intelligent application phase, complex flood control tasks are transformed into standardized, traceable, and structured processes. This process integrates professional capabilities such as rainfall analysis, hydrological simulation, and scheduling decisions. Directed task graphs and collaborative functions ensure seamless integration of cross-domain knowledge and models, ultimately outputting a flood control decision-making scheme that meets business requirements. This effectively solves the problems of traditional decision-making relying on human experience, difficulty in integrating multi-source data, and opaque multi-reasoning logic.

[0023] (I) Design Phase: Business Decomposition and Knowledge Base Construction The design phase is the foundation of the entire flood control decision-making method. Its core objective is to transform the abstract and complex flood control task into an ordered and executable decision sequence, and to build a knowledge base to support reasoning, providing a framework and data support for the subsequent operation and reasoning process of the intelligent agent.

[0024] Modular Decomposition of Complex Flood Control Tasks: Flood control task T encompasses multiple stages, including rainfall monitoring, hydrological forecasting, and dispatch plan generation. It requires modular decomposition based on domain business logic. First, the core business chain of flood control decision-making is clarified: rainfall analysis, flood simulation, decision-making, and result output. Based on this, task T is decomposed into an ordered decision sequence. In this embodiment, Corresponding to the rainfall analysis subtask, Corresponding to the water situation simulation sub-task, For each scheduling decision subtask, each decision sequence It must include decision-making objectives. Decision output With knowledge demand The expression is , ,by For example, To obtain the spatiotemporal distribution of rainfall in the target watershed over the next 24 hours, For rainfall intensity distribution maps and total rainfall statistics reports, This includes real-time rainfall data, historical rainfall datasets for the same period, and parameters for regional rainfall forecasting models. of To predict reservoir water level changes and river flow within the basin, For the next 48 hours, a report on reservoir water level curves and river flow dynamics will be provided. This includes river hydrological data, current reservoir water level data, and flood evolution model parameters. of To generate the optimal solution for reservoir scheduling and river drainage, Suggestions for reservoir discharge volume and plans for the start-up and shutdown of drainage pumping stations. For reservoir scheduling rules, urban flood control standards, and pumping station operating parameters, by defining It allows for flexible replacement of knowledge requirements to adapt to flood control scenarios in different river basins, thereby improving the versatility of the method.

[0025] Knowledge Base Construction and Retrieval Strategy Design: The knowledge base is constructed using a pyramid method combining document encoding and structure encoding to ensure the hierarchical and interconnected organization of knowledge. The bottom layer of the pyramid is the raw document layer, which collects historical data documents, model parameter documents, and business specification documents in the field of flood control. The middle layer is the structured encoding layer, which organizes the bottom-level documents according to business objectives. Knowledge needs The document's ternary mapping relationship is encoded, for example, binding water level prediction with river hydrological data documents and flood evolution model parameter tables to form a structured index. The top layer is the knowledge service layer, providing a retrieval interface for the agent to call. To improve the accuracy and efficiency of knowledge retrieval, a hybrid retrieval strategy is adopted: First, a vector retrieval based on the semantic model BGE-M3 is performed, calculating the similarity between the user's question and the semantic vectors of documents in the knowledge base to initially filter out a set of semantically related documents; Second, a full-text retrieval is performed, supplementing documents that may have been missed by the vector retrieval through keyword matching; Third, the reranking model BGE-reranker-large is used to comprehensively score the results of the two retrievals, with scoring dimensions including semantic relevance, business matching degree, and data timeliness. Finally, the top 10 documents with the highest scores are selected as the optimal results to ensure the accuracy of the knowledge called by the agent.

[0026] (II) Agent Construction Phase: Specialized Agent Development The decision sequence that needs to be broken down in the design phase needs to be addressed during the agent construction phase. Develop specialized intelligent agents with unique knowledge, rules, and reasoning abilities to form an ensemble of intelligent agents. ,in Intelligent agents for rainfall analysis , Intelligent agent for water situation simulation , For scheduling decision-making intelligent agents Each intelligent agent The expression is That is, the function of an intelligent agent is determined by the corresponding decision sequence. with reasoning strategies A joint decision.

[0027] Rainfall Analysis Intelligent Agent Development: Bind a dedicated knowledge sub-repository within the knowledge base It includes real-time rainfall data, historical rainfall datasets, and rainfall forecasting model parameters. Its core function is to integrate multi-source rainfall data and analyze rainfall trends, and its inference strategy... This system includes a multi-source rainfall data fusion algorithm and spatiotemporal distribution analysis logic. It employs a weighted fusion algorithm to eliminate data bias from various sources, such as real-time weather station monitoring data, satellite remote sensing rainfall data, and radar detection data. The weights are dynamically adjusted based on data accuracy. Based on the fused real-time data and historical datasets, a rainfall forecasting model is invoked, and input... The model parameters are used to predict changes in rainfall intensity over the next 24 hours through time series analysis. Combined with GIS spatial interpolation algorithms, a spatiotemporal distribution heat map of rainfall is generated, and the final output is the decision-making outcome. .

[0028] Water situation simulation intelligent agent Development: Binding a dedicated knowledge sub-repository It includes river hydrological data, reservoir water level data, and flood evolution model parameters. Its function is to simulate the evolution of floods within a watershed and to develop inference strategies. Includes reservoir water level prediction logic and flood control calculation process, with Output rainfall data For input, combine Based on river hydrological data and current reservoir water level data, a water level prediction model is established. The inflow rate for future periods is calculated using the water balance equation. Combined with flood evolution model parameters, the reservoir water level change curve for the next 48 hours is predicted. When the predicted water level approaches the flood control limit, flood regulation calculations are initiated to simulate water level changes under different discharge rates, ensuring the calculation results comply with reservoir safety operation standards. Finally, the decision-making output includes a water level curve and a flow report. .

[0029] Scheduling Decision-Making Intelligent Agent Development Binding a dedicated knowledge sub-repository Its function is to generate the optimal scheduling scheme based on the results of hydrological simulation and to reason about the strategy. Includes solution generation logic and feasibility verification rules, to Output For input, combine The reservoir scheduling rules and urban flood control standards are analyzed using a multi-objective optimization algorithm. The algorithm aims to ensure reservoir safety, reduce urban flooding, and lower water transfer costs. It generates 3-5 candidate scheduling schemes, including parameters such as discharge volume and pump station start-up and shutdown times. A rule engine is then used to verify whether the candidate schemes meet business constraints, and two optimal schemes are selected as the decision output. .

[0030] (III) Agent Scheduling and Collaborative Reasoning Stage: Large Language Model Reasoning and Multi-Agent Collaboration Task semantic parsing and intent recognition: Reasoning is initiated through a large language model combined with a decision-making chain template. Based on the task type, the corresponding intelligent agent is invoked to achieve sub-task reasoning and result integration, ultimately outputting a flood control decision plan. Users submit questions through a flood control question input module. The input text is cleaned, and then a flood control domain enhancement model extracts core keywords and phrases. Subsequently, the large language model, combined with a pre-set decision-making chain template, performs semantic parsing and intent recognition on task T, generating a task decomposition tree. ,in, , , The correspondence is consistent with the design phase, and each subtask correspond , , , 1, 2, 3, Task Decomposition Tree Directly used for constructing directed task graphs The basis for invoking the intelligent agent is based on the dependencies between subtasks. Output yes Input, Output yes Input, construct a directed task graph ,in For a set of task nodes, For a set of dependent edges, This indicates that it needs to be completed first. Execute again ,Finish Only then can it be executed. Since user issues are complex, cross-domain tasks, they are addressed using a directed task graph. Middle Dependency Edge The constraint order is determined by calling multiple agents through a cooperative function to achieve hierarchical cooperative reasoning. The cooperative function expression is as follows: Coop Indicates the entire task The results of multi-agent collaborative reasoning. Call This indicates that the corresponding agent is invoked to execute the subtask and output the result. This is a set union operation used to integrate the invocation processes of all agents: Step 1: Call the rainfall analysis intelligent agent implement Through function Automatic matching and Retrieve a dedicated knowledge sub-repository from the knowledge base. First, a multi-source rainfall data fusion algorithm is run to integrate real-time and historical data. Then, rainfall spatiotemporal distribution analysis logic is used to generate... and will As The input data is transmitted to the system cache.

[0031] Step 2: Calling the water situation simulation intelligent agent implement In acquiring Then, based on the dependency edges via the Call function Call , Retrieve River hydrological data and reservoir water level data, combined The rainfall data is used to run the reservoir water level prediction logic and flood control calculation process to generate... If the predicted water level exceeds the safety threshold, the risk point will be automatically marked and synchronized to [the relevant authorities]. .

[0032] Step 3: Invoke the scheduling decision-making intelligent agent implement Based on the dependency edge Through function Call , Retrieve The scheduling rules and flood control standards in China, Based on water level and flow data, a multi-objective optimization algorithm is used to generate candidate scheduling schemes. Then, the optimal scheme is selected through feasibility verification rules, forming a... (This includes a scheduling plan that includes recommendations for flood discharge volume and pump station start-up and shutdown schedules).

[0033] Results Integration and Output take over Output , , Then, the results are integrated through the flood control answer output module: Rainfall data, Water situation forecast results and The system associates scheduling plans with data to generate structured decision reports that include problem analysis, scheduling suggestions, and visualization charts. These reports are then distributed through multiple channels: real-time water conditions and scheduling plans are displayed on the flood control command screen; risk warning information is pushed to staff via a mobile app; and data is synchronized to the city's flood control management platform via API for subsequent business decisions and data traceability.

[0034] In summary, this embodiment clarifies the standardized process and knowledge requirements for flood control decision-making through task decomposition and knowledge base construction in the design phase; it achieves modularization of capabilities in the fields of rainfall, water conditions, and dispatching through professional development in the agent construction phase; and it ensures the orderly execution of cross-domain tasks and the accuracy of results through large language model reasoning and multi-agent collaboration in the intelligent application phase. Throughout the process, the reasoning logic is traceable and the business process is reproducible, effectively solving the pain points of traditional flood control decision-making. Furthermore, the application of hybrid retrieval strategies and directed task graphs improves the efficiency of knowledge retrieval and the accuracy of multi-agent collaboration, providing scientific and efficient technical support for urban watershed flood control decision-making.

[0035] Example 2 Based on Embodiment 1, this embodiment provides a flood control decision-making method based on multi-agent collaboration, detailing the specific steps involved in making such a decision. Figure 1 As shown, the specific steps are as follows: (a) Design phase: such as Figure 1 As shown, a decision-making framework and knowledge base are constructed. Analyze the business logic in the flood control field and identify the core components of complex task T.

[0036] Task T is modularly broken down to form an ordered decision sequence.

[0037] Define decision objectives for each subtask.

[0038] Determine the knowledge requirements for each subtask.

[0039] A knowledge base is constructed using a pyramid method that combines document encoding and structure encoding, collecting original documents such as historical data, model parameters, and business specifications.

[0040] Establish a three-dimensional mapping relationship between business objectives, knowledge requirements, and documents, and bind knowledge requirements to documents in the knowledge base.

[0041] Configure a hybrid retrieval strategy: first execute vector retrieval and full-text retrieval using the semantic model, then use a re-ranking model to rearrange the results and select the optimal knowledge.

[0042] Construct a standardized decision-making thought chain template to clarify the execution order and reasoning logic of sub-tasks.

[0043] (ii) Agent construction phase: such as Figure 2 As shown, develop specialized functional units For each subtask in the design phase, develop a corresponding specialized intelligent agent.

[0044] Each intelligent agent is bound to a dedicated knowledge sub-base within the knowledge base.

[0045] Configure business rules for each agent.

[0046] Set a reasoning strategy for each agent.

[0047] Test the functionality of each agent to ensure that it can output results that meet the decision-making output requirements based on the knowledge subbase and reasoning strategy.

[0048] Integrate all intelligent agents to form a complete set of intelligent agents, ensuring that each intelligent agent can be invoked independently and respond collaboratively.

[0049] (III) Intelligent Scheduling and Collaborative Reasoning Stage: such as Figure 3 As shown, initiating reasoning and multi-agent collaboration User questions are received through the flood control question input module.

[0050] Clean the user-input text.

[0051] By using a large language model combined with a pre-set decision-making thought chain template, semantic analysis is performed on user questions to identify task intent.

[0052] Generate a task decomposition tree based on the intent, and clarify the number, type, and corresponding objectives, outputs, and knowledge requirements of the subtasks.

[0053] Based on the task decomposition tree, determine the subtask type: If it is a sub-task in a single domain, the corresponding intelligent agent will be automatically matched.

[0054] For complex cross-domain tasks, analyze the dependencies between subtasks.

[0055] For cross-domain tasks, a directed task graph is constructed based on dependencies to mark the execution order of subtasks.

[0056] Invoke agents in task order: Invoke the agent corresponding to the first subtask, enabling it to read its dedicated knowledge base, execute the inference strategy, and output intermediate results.

[0057] Once the previous agent has completed its task, its output is used as the input for the next subtask, and the corresponding agent is invoked.

[0058] Repeat this process until all subtasks have been completed.

[0059] The system receives and integrates the intermediate results from all agents.

[0060] The flood control answer output module transforms the integrated results into a structured report containing problem analysis, scheduling suggestions, and visual charts.

[0061] The final decision report is delivered through channels such as web pages, mobile devices, flood control command screens, and API interfaces.

[0062] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A flood control decision-making method based on multi-agent collaboration, characterized in that, The specific steps of this method are as follows: Design Phase: Based on domain business logic, the complex flood control task T is modularly decomposed into an ordered decision sequence. A standardized decision-making thought chain template is constructed, and the knowledge requirements driven by the decision sequence S are clearly defined. These are then mapped to the corresponding document collections to form a corresponding knowledge system and knowledge base; Agent construction phase: For each decision sequence Develop specialized intelligent agents to form an ensemble of intelligent agents. Each intelligent agent Bind a dedicated knowledge sub-repository within the knowledge base Business Arrangement Operating boundary Data Standards And expressed as: ; Agent scheduling and collaborative reasoning phase: First, the input task... The task decomposition tree is obtained through parsing and identification. For single-domain sub-tasks The response result is generated by calling a function. For complex tasks that span multiple domains and stages, construct a directed task graph. and according to the dependency edges The constraint sequence calls multiple agents to achieve hierarchical collaborative reasoning; Intelligent application stage: Receive user input questions and automatically initiate the reasoning process through a large language model combined with a preset decision-making thought chain template.

2. The flood control decision-making method based on multi-agent collaboration according to claim 1, characterized in that, In the design phase, each decision sequence Including reasoning objectives Key constraints and decision output This process can be represented as: Based on the input task description, it can automatically match a decision-making thought chain template using the semantic understanding capabilities of LLM.

3. The flood control decision-making method based on multi-agent collaboration according to claim 1, characterized in that, In the design phase, the decision-making thought chain To guide, each sequence is clearly defined. Required knowledge This information is then mapped to the corresponding document collections, forming a structured, semantically schedulable knowledge system and knowledge base, and then the reasoning objectives are established. Knowledge needs The document-related ternary mapping relationship is then updated to the decision sequence expression. It adopts the pyramid method of "document encoding + structure encoding" and combines the mainstream semantic model bge-m3 encoding model and bge-reranker-large reordering model to achieve a hybrid retrieval that combines full-text retrieval and vector retrieval in parallel.

4. The flood control decision-making method based on multi-agent collaboration according to claim 1, characterized in that, In the intelligent agent construction phase, for complex business... The multidimensional characteristics, centered around the reasoning goal Construct DomainAgents, denoted as Each agent Focus on specific decision sequences and reasoning target Design corresponding reasoning strategies To achieve specific actions, and to ensure that the agent possesses both independent domain capabilities and is easy for system scheduling, the agent's operational boundaries and input / output standards are explicitly introduced as parameters, and the agent's behavior is abstracted as follows: ,in, This refers to specific business orchestration, including mechanism model invocation, knowledge base invocation, and business boundaries and input / output format standards. This refers to the operational boundaries of a single intelligent agent, including permissions and scope of invocation, triggering conditions, spatiotemporal constraints, and security limitations. This refers to a unified description of input data requirements, completion rules, output format, and response time.

5. A flood control decision-making method based on multi-agent collaboration according to claim 1, characterized in that, In the agent scheduling and collaborative reasoning phase, for the task decomposition tree First, semantic parsing and intent recognition are performed to obtain... Each subtask Corresponding reasoning target Decision output and knowledge needs ,Right now .

6. A flood control decision-making method based on multi-agent collaboration according to claim 5, characterized in that, For the identified single-domain subtask By calling the function Automatically match and drive the corresponding specialized intelligent agents Execute this subtask to generate and output intermediate inference results. .

7. A flood control decision-making method based on multi-agent collaboration according to claim 5, characterized in that, For complex cross-domain tasks, a directed task graph is constructed based on task dependencies. ,in, For a set of task nodes and , Let be the set of dependency edges between tasks and And it calls multiple agents according to the constraint order of dependency edge E to achieve hierarchical collaborative reasoning: ,in, Indicates the input task The overall result of the multi-agent collaborative reasoning process. The union operation of sets represents the integration of the calling processes of all agents that satisfy dependencies, forming a complete collaborative reasoning chain. and It is a subtask in the task decomposition tree. For task node set The first in A prerequisite subtask It depends on The subsequent subtasks, It is a set of dependency edges between tasks. Subtasks and There are directed edges that depend on the order of completion, meaning that the task must be completed first. Only then can it be executed , Indicates calling the responsible subtask intelligent agents To execute subtasks And obtain the result after the agent executes the command. It focuses on sub-tasks A specialized intelligent agent possesses a dedicated knowledge base, business rules, and reasoning strategies for handling this sub-task. This indicates that the subtask has been completed. Based on this, call the subtask responsible intelligent agents To execute subtasks And obtain the results.

8. A flood control decision-making method based on multi-agent collaboration according to claim 1, characterized in that, In the intelligent application phase, automated flood control decision-making is achieved based on user input, including thought chain inspiration, agent scheduling, and result output. During the thought chain inspiration stage, after the user inputs a question, the platform uses a large model combined with a pre-set decision thought chain to analyze the user's intent, clarify the core needs, and decompose the needs into a sequence of sub-tasks that conform to the logic of the thought chain. During the agent scheduling phase, agents are invoked through the Multi-Agent Coordination Protocol (MCP), and subtasks are assigned... With context state Mapped to a specific executable agent calling function: That is, while ensuring context consistency, tasks are assigned to appropriate agents. Based on this, automatic scheduling of multiple agents is completed according to the priority of sub-tasks; In the result output stage, the platform extracts key information from the user input, automatically completes parameter format conversion and necessary parameter completion according to the input standards of the corresponding intelligent agent, and after each intelligent agent completes the sub-task reasoning, it summarizes the intermediate results and integrates them into a complete decision plan according to the flood control business logic, and outputs the final answer.

9. A flood control decision-making system based on multi-agent collaboration, applicable to the flood control decision-making method based on multi-agent collaboration as described in any one of claims 1-8, characterized in that, The system includes: The flood control problem input module is used to collect user problems, clean and organize the input text, and extract core keywords and phrases by combining an enhanced model in the field of flood control. The flood control knowledge retrieval module is used to break down user questions into entity-relation-attribute triples, and then use a hybrid retrieval strategy to combine semantic weight and business weight to rearrange and filter relevant knowledge bases. The flood control thinking chain inspiration module is used to fine-tune and train the large language model based on flood control task thinking chain samples annotated by experts. It combines the knowledge base and user questions to generate multi-step reasoning thinking chains and verifies the logical consistency through the rule engine. The flood control multi-agent collaboration module is used to build multiple domain agents. Each agent has a built-in knowledge sub-base, business rules and reasoning strategies, and the sub-tasks are broken down according to the thought chain and assigned to the corresponding agents for execution. The flood control answer output module is used to transform the results of multi-agent collaboration into a structured form that includes problem analysis, scheduling suggestions, and visual charts, and outputs them through web pages, mobile devices, flood control command screens, and API interfaces.

Images