Mountain torrent debris flow dynamic decision method and system based on multi-agent and LLM

Abstract

This application discloses a dynamic decision-making method and system for flash floods and debris flows based on multi-agent and LLM (Limited Language Model), belonging to the field of emergency management and intelligent decision-making technology. This method constructs an intelligent decision-making architecture based on a large model and multi-agent collaboration. It solves the execution challenges of complex disaster tasks through the division of labor and cooperation among multiple agents, and overcomes the decision-making limitations of traditional systems by leveraging the deep reasoning capabilities of a large language model. Specifically, the multi-agent collaborative framework decomposes disaster prevention and control decisions into sub-tasks such as intent parsing, contingency plan retrieval, spatial analysis, data acquisition, and analysis and prediction. Each functional agent focuses on its area of ​​expertise, improving the efficiency and accuracy of task execution. The agent communication protocol ensures efficient information interaction and collaborative scheduling between modules, realizing fully automated processing from task input to data output.

Description

Technical Field

[0001] This application belongs to the field of emergency management and intelligent decision-making technology, specifically involving a dynamic decision-making method and system for flash floods and debris flows based on multi-agent and LLM. Background Technology

[0002] Existing emergency management systems suffer from several technical limitations. In terms of plan retrieval, traditional systems rely on keyword matching, which struggles to understand the semantic intent of user queries, resulting in poor relevance of search results. Furthermore, plans are prone to losing contextual semantic information after being sliced. Regarding spatial correlation analysis, the systems lack the ability to deeply mine and intelligently analyze the complex spatial relationships between monitoring stations and hazardous areas. At the task planning level, they cannot automatically decompose data acquisition and analysis steps according to user needs. In query processing, the systems struggle to accurately translate users' natural language queries into executable database queries. Moreover, the systems have significant shortcomings in multi-source data fusion, failing to effectively integrate plan knowledge, spatial information, real-time monitoring data, and predictive analysis results, thus hindering the level of intelligent emergency decision-making. In addition, most current emergency support systems remain at a rudimentary stage of intelligence, relying on "rule matching + statistical analysis," lacking true semantic understanding and causal reasoning capabilities. Faced with complex disaster evolution processes, these systems cannot perform "if...then..." type hypothetical deductions like experts, cannot make forward-looking judgments by combining historical experience with real-time situations, and cannot infer potential risk propagation paths based on local anomaly data. This "no-reasoning" decision-making model severely restricts the ability to cope with uncertain events. Summary of the Invention

[0003] To address the problems of existing technologies, this application proposes a dynamic decision-making method and system for flash floods and debris flows based on multi-agent and LLM. By constructing a multi-agent collaborative framework, it realizes the division of labor and cooperation for complex tasks, and uses a large language model as the core reasoning engine to realize the semantic integration and deep causal inference of multi-source data, thereby improving the intelligence, accuracy and efficiency of disaster prevention and control decision-making.

[0004] This application is achieved through the following technical solution:

[0005] A dynamic decision-making method for flash floods and debris flows based on multi-agent and LLM approaches.

[0006] This dynamic decision-making method is based on a pre-constructed multi-agent collaborative framework and a large language model. The multi-agent collaborative framework includes an intent parsing agent, a retrieval agent, a spatial analysis agent, a data acquisition agent, and an analysis and prediction agent. It achieves information interaction and collaborative scheduling between each functional agent and the large language model through an agent communication protocol, including:

[0007] The intent parsing agent performs intent parsing on the decision-making task and decomposes it into multiple sub-tasks, and matches each sub-task with a corresponding functional agent.

[0008] The retrieval agent retrieves relevant emergency plan content from a pre-built emergency plan database based on the parsed decision intent.

[0009] The spatial analysis agent retrieves relevant spatial relationships from a pre-constructed spatial knowledge graph based on the parsed decision intent.

[0010] The data acquisition agent generates and executes SQL query statements based on the parsed decision intent to obtain real-time monitoring data;

[0011] The analytical and predictive agent performs statistical analysis and time-series prediction based on the real-time monitoring data to obtain the current disaster risk assessment results and prediction trends.

[0012] The decision-making tasks, emergency plan content, spatial relationships, real-time monitoring data, and current disaster risk assessment results and predicted trends are input into the large language model for semantic integration and multi-dimensional causal inference, generating a final structured dynamic processing suggestion report and outputting it.

[0013] In some implementations, the process of interpreting the intent of the decision-making task and decomposing it into multiple sub-tasks includes:

[0014] Load the pre-built business knowledge base and call the semantic understanding interface of the large language model to identify key semantic elements in the decision-making task, including geographical entities, time range, disaster type and task intent category;

[0015] For vague or incomplete expressions, semantic completion and integrity verification are performed by combining user role, current location or historical behavior context, and the large language model to obtain structured semantic parsing results;

[0016] Based on the semantic parsing results, a predefined task template is matched to decompose the original decision task into several executable sub-tasks, and the dependency relationships between the sub-tasks are constructed to form a task dependency graph.

[0017] Each subtask is equipped with a corresponding data requirement list and recommended analysis methods, and a corresponding functional agent is matched to issue task execution instructions.

[0018] In some implementations, the process of constructing the emergency response plan database includes:

[0019] The original contingency plan document is structured and parsed to extract text content, semantic structure, and metadata information;

[0020] Based on the semantic structure, the sliding window method combined with semantic boundary detection is used to segment the text content into semantically complete pre-plan slices;

[0021] The content of the plan slices is deduplicated, the terminology is standardized, and an index mapping relationship between the plan slices and the original documents is established.

[0022] For each of the aforementioned plan slices, a structured summary is generated and then spliced ​​with the slice;

[0023] The pre-trained semantic coding model is used to convert the plan slices into high-dimensional vector representations;

[0024] Establish a triplet storage structure that includes the content of the contingency plan slices, summary information, and high-dimensional vector representation.

[0025] In some implementations, retrieving relevant emergency plan content from a pre-built emergency plan database includes:

[0026] Convert the task instructions issued by the intent parsing agent into query vectors;

[0027] The query vector is compared one by one with all the plan slice vectors in the emergency plan database, and the top K with the highest similarity are selected from all the plan slice vectors as candidate content.

[0028] The candidate content is sorted a second time based on task requirements and slice summary to determine the final emergency plan content and upload it to the large language model.

[0029] In some implementations...

[0030] The spatial knowledge graph adopts a layered construction strategy, completing the design and implementation of the ontology layer and the relation layer in sequence. Each layer works together to form a knowledge system with spatial reasoning capabilities. The ontology layer serves as the core skeleton of the knowledge graph, used to clearly define all entity categories participating in spatial analysis and the core attributes corresponding to each entity.

[0031] The relationship layer is the core connecting the spatial entities in the local layer, clarifying the entity association logic and dividing it into two categories: spatial relationships and attribute relationships. The spatial relationships are constructed based on GIS topology theory and entity location, including: the monitoring buffer zone belongs to a specific geographical area, and the sub-basin contains monitoring stations and waterways; adjacency relationships: sub-basins are adjacent upstream, and waterways are connected downstream according to topographic elevation and water flow direction; location relationships: monitoring stations are located in sub-basins or beside waterways, and the source of spatiotemporal events falls in a specific sub-basin; the attribute relationships are based on spatial analysis logic, clarifying entity attribute dependencies, including: equipment-buffer zone association, topography-water flow association, watershed-waterway association, and event-diffusion association.

[0032] In some implementations, retrieving relevant spatial relationships from a pre-constructed spatial knowledge graph includes:

[0033] Starting with the geographic entities in the task instructions issued by the intent parsing agent, the corresponding geographic region entities are matched from the spatial knowledge graph;

[0034] By linking down the geographical area to match various monitoring stations and the risk areas monitored by each monitoring station, a "region → station → risk area" link is established.

[0035] Based on the upstream and downstream relationships of monitoring stations in terms of hydrology or geography, explore related stations within 1-3 jumps;

[0036] Expand the scope of the risk area and link it to key facilities;

[0037] Screening is based on risk level criteria;

[0038] The associated entities generated by multi-hop paths are aggregated to form a structured output.

[0039] In some implementations, generating and executing the SQL query statement includes:

[0040] Based on the task instructions issued by the intent parsing agent, the large language model is invoked to perform semantic completion and context expansion processing, and the integrity of the original task is verified to form a structured extended task.

[0041] Based on the extended task, entity abstraction is performed to extract key semantic entities;

[0042] Based on the extracted entity information, table relevance is evaluated from the real-time monitoring database to select a set of candidate tables.

[0043] Perform field relevance analysis on each candidate table in the candidate table set to determine the effective query fields in the candidate tables and establish a mapping relationship between fields and semantic entities;

[0044] Based on the foreign key relationships and spatial topology between the candidate tables, the optimal table join path is recommended;

[0045] Based on the comprehensive query type judgment results, a predefined standardized SQL template is invoked, and the large language model generates an SQL query statement.

[0046] Execute the generated SQL query to obtain real-time monitoring data; if the execution fails, proceed to the error handling process.

[0047] The returned data is formatted and then uploaded to the large language model.

[0048] In some implementations, the analytical predictive agent combines statistical methods and deep learning models to predict future disaster trends. It receives real-time monitoring data from the data acquisition agent and static risk factors from the spatial analysis agent, uses physical rules and data analysis models to assess disaster risks and predict the probability of disaster occurrence, peak time and impact range within future time windows, and outputs the current disaster risk level, future disaster trends and uncertainty intervals to the large language model.

[0049] In some implementations, generating the final structured dynamic processing recommendation report includes:

[0050] The large language model receives structured data uploaded by various functional agents, comprehensively considers three factors—the urgency of the information, the scope of its impact, and the credibility of the results—to calculate the importance of the content and rank it.

[0051] Evidence-based risk attribution analysis integrates multi-source observation data and background knowledge using a large language model to identify the root causes of disasters; reveals key factors driving trends and their pathways; performs multi-hop causal chain deduction based on spatial topological relationships and evolutionary patterns to identify chain reaction risks in advance; provides logical justification for disposal recommendations; and conducts intelligent analysis of predicted trends and key time nodes, marking risk inflection points.

[0052] The large language model selects the corresponding report template based on the risk level and disaster type;

[0053] The large language model adjusts the level of detail and the use of technical terms based on the user's role, generating a structured and dynamically processed suggestion report.

[0054] Secondly, this application proposes a dynamic decision-making system for flash floods and debris flows based on multi-agent and LLM, including:

[0055] A multi-agent collaborative module, comprising:

[0056] The intent analysis agent is used to parse the intent of the decision task and decompose it into multiple sub-tasks, match the corresponding functional agent for each sub-task and issue task execution instructions.

[0057] The retrieval agent is used to retrieve relevant emergency plan content from the pre-built emergency plan database based on the parsed decision intent and upload the retrieval results to the large language model inference module.

[0058] A spatial analysis agent is used to query relevant spatial relationships from a pre-built spatial knowledge graph based on the parsed decision intent and upload the query results to the large language model inference module.

[0059] The data acquisition agent is used to generate and execute SQL query statements based on the parsed decision intent, acquire real-time monitoring data and upload it to the large language model inference module.

[0060] An analytical and predictive agent is used to perform statistical analysis and time-series prediction based on the real-time monitoring data, obtain the current disaster risk assessment results and prediction trends, and upload the assessment results and prediction trends to the large language model inference module.

[0061] In addition, the large language model reasoning module uses the large language model to receive the decision-making tasks, emergency plan content, spatial correlations, real-time monitoring data, and current disaster risk assessment results and prediction trends uploaded by each functional agent, and performs semantic integration and multi-dimensional causal inference to generate and output the final structured dynamic processing suggestion report.

[0062] In the multi-agent collaboration module 201, each functional agent and the large language model inference module achieve information interaction and collaborative scheduling through an agent communication protocol.

[0063] This application proposes a dynamic decision-making method for flash floods and debris flows based on multi-agent and LLM (Language Model). It constructs an intelligent decision-making architecture that combines a large model with multi-agent collaboration. By leveraging the division of labor and cooperation among multiple agents, it addresses the challenges of executing complex disaster tasks and overcomes the limitations of traditional systems through the deep reasoning capabilities of a large language model. Specifically, the multi-agent collaborative framework decomposes disaster prevention and control decisions into sub-tasks such as intent parsing, contingency plan retrieval, spatial analysis, data acquisition, and analysis and prediction. Each functional agent focuses on its area of ​​expertise, improving the efficiency and accuracy of task execution. The agent communication protocol ensures efficient information exchange and collaborative scheduling between modules, achieving fully automated processing from task input to data output.

[0064] As the core reasoning engine, the large language model undertakes the key responsibilities of semantic integration and causal inference. On the one hand, it leverages its powerful semantic understanding capabilities to assist agents in completing tasks such as semantic completion and accurate intent recognition. On the other hand, it integrates multi-source heterogeneous data (pre-plan knowledge, spatial relationships, real-time monitoring data, and prediction results) output by multiple agents to conduct in-depth causal analysis and multi-hop inference, generating structured disposal suggestions containing complete reasoning chains. This architecture of "multi-agent collaborative execution + large model deep reasoning" effectively solves key technical bottlenecks in traditional emergency systems, such as weak semantic understanding, mechanized spatial analysis, rigid task planning, and lack of mechanistic support for decision-making. It achieves a leapfrog evolution from "passive response" to "proactive prediction," from "data display" to "causal insight," and from "mechanical execution" to "cognitive decision-making," significantly improving the efficiency and scientific nature of emergency response in complex disaster scenarios. It provides a new technical basis and engineering practice path for the intelligent prevention and control of multi-factor coupled disasters such as flash floods and mudslides.

[0065] Correspondingly, the dynamic decision-making system for flash floods and debris flows based on multi-agent and LLM proposed in this application also possesses the same technical effects as described above. Attached Figure Description

[0066] The accompanying drawings, which are included to provide a further understanding of the embodiments of this application and form part of this application, do not constitute a limitation on the embodiments of this application. In the drawings:

[0067] Figure 1 This is a flowchart of the dynamic decision-making method proposed in the embodiments of this application;

[0068] Figure 2 This is a block diagram illustrating the principle of the dynamic decision-making system proposed in the embodiments of this application;

[0069] Figure 3 This is a schematic diagram of the electronic device proposed in the embodiments of this application;

[0070] Figure 4 This is a schematic diagram of a computer-readable storage medium proposed in an embodiment of this application;

[0071] Figure reference numerals and corresponding component names:

[0072] 200-Dynamic decision-making system; 201-Multi-agent collaboration module; 2011-Intent analysis agent; 2012-Retrieval agent; 2013-Spatial analysis agent; 2014-Data acquisition agent; 2015-Analysis and prediction agent; 202-Large language model inference module; 300-Electronic device; 310-Memory; 320-Processor; 311-Computer program A; 400-Computer-readable storage medium; 411-Computer program B. Detailed Implementation

[0073] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the embodiments and accompanying drawings. The illustrative embodiments and descriptions of this application are only for explaining this application and are not intended to limit this application.

[0074] Existing disaster emergency response technologies suffer from key technical bottlenecks such as weak semantic understanding, mechanized spatial analysis, rigid task planning, and a lack of mechanistic support for decision-making. To address these issues, this application proposes a dynamic decision-making method for flash floods and debris flows based on multi-agent and LLM approaches. Figure 1 As shown, this dynamic decision-making method is based on a pre-built multi-agent collaborative framework and a large language model. The multi-agent collaborative framework includes an intent parsing agent, a retrieval agent, a spatial analysis agent, a data acquisition agent, and an analysis and prediction agent. It achieves information interaction and collaborative scheduling between each functional agent and the large language model through an agent communication protocol (A2A protocol), including the following steps:

[0075] Step 1: The decision task is parsed and decomposed into multiple sub-tasks by the intent parsing agent, and a corresponding functional agent is matched for each sub-task.

[0076] Step 2: The retrieval agent retrieves relevant emergency plan content from the pre-built emergency plan database based on the parsed decision intent.

[0077] Step 3: Based on the decision intent obtained from the parsing, the spatial analysis agent queries the pre-constructed spatial knowledge graph to obtain relevant spatial relationships;

[0078] Step 4: The data acquisition agent generates and executes an SQL query statement based on the parsed decision intent to obtain real-time monitoring data;

[0079] Step 5: The analysis and prediction agent performs statistical analysis and time series prediction based on real-time monitoring data to obtain the current disaster risk assessment results and prediction trends;

[0080] Step 6: Input the decision-making task, emergency plan content, spatial correlation, real-time monitoring data, and current disaster risk assessment results and prediction trends into the large language model for semantic integration and multi-dimensional causal inference, generate the final structured dynamic processing suggestion report and output it.

[0081] Furthermore, the intent parsing agent includes a semantic understanding part and a task decomposition part. The semantic understanding part is specifically implemented as follows: it loads a pre-built business knowledge base and calls the semantic understanding interface of the large language model to identify key semantic elements in the decision-making task, including geographical entities (such as "XX River Basin" or "XX County"), time range (such as "future 24 hours"), disaster type (such as "flash flood" or "mudslide"), and task intent category (such as "risk query", "early warning judgment", or "response plan generation"). For vague or incomplete expressions (such as "Is there any danger nearby?"), it combines the user role, current location, or historical behavior context to perform semantic completion and integrity verification through the large language model, and outputs structured semantic parsing results.

[0082] The task decomposition section breaks down complex decision-making tasks that have undergone semantic understanding into several executable subtasks. Specifically, based on the semantic parsing results, a predefined task template is matched to decompose the original task into several specific subtasks (e.g., "acquire rainfall data," "retrieve historical contingency plans," "analyze the impact of terrain slope"). Simultaneously, combining rules from the business knowledge base regarding disaster evolution logic and emergency response procedures, the dependencies between subtasks are constructed, forming a task dependency graph. Each subtask is then equipped with a corresponding data requirement list (e.g., real-time rainfall, soil moisture content, terrain slope, etc.) and recommended analysis methods (e.g., time-series trend analysis, spatial overlay analysis, physical model simulation, etc.), and finally, the corresponding functional agents (e.g., data acquisition agent, retrieval agent, spatial analysis agent, analysis and prediction agent, etc.) are matched.

[0083] For example, the input of the intent parsing agent is a disaster prevention and control decision-making task submitted by the user through the interface in natural language form (such as "Does a certain watershed need to activate a red alert in the next 24 hours?"). The intent parsing agent processes the task by: (1) performing semantic analysis to identify key information, such as region, time range, and disaster type; (2) decomposing the complex task into multiple specific sub-tasks, such as "acquiring rainfall data", "retrieving historical plans", and "analyzing the impact of terrain slope"; and (3) determining the logical order or dependency between these sub-tasks and assigning a corresponding functional agent to each sub-task. Finally, the intent parsing agent outputs a set of structured sub-task instructions and their dependencies (represented in the form of a task dependency graph), which is sent to other agents through the system's internal message mechanism, such as retrieval agents, spatial analysis agents, data acquisition agents, and analysis and prediction agents.

[0084] This application embodiment decomposes complex tasks into several executable sub-tasks. In subsequent applications, task parallel scheduling can also be adopted, including: parallel execution of independent tasks (mutually independent tasks can be executed in parallel, such as "real-time hydrological data acquisition" and "emergency plan retrieval" can be performed simultaneously); pipelined processing (subsequent tasks are started immediately after the completion of the preceding tasks, such as "statistical analysis" is started immediately after the completion of "real-time data acquisition"); and resource allocation optimization (allocating computing resources according to the computational complexity of the task, such as allocating more CPU / GPU resources to numerical prediction tasks).

[0085] Furthermore, the retrieval agent connects to a pre-built emergency response plan database, with input consisting of task requirements from the intent parsing agent (including keywords, geographical scope, disaster level, etc.). The retrieval agent transforms the task requirements into query vectors; searches the emergency response plan database for the most relevant emergency response plans; filters and sorts the results, removing duplicates and ranking them by relevance. The output is the content of the found relevant emergency response plans (including response levels, evacuation routes, and resource allocation plans), uploaded to the large language model in JSON format.

[0086] Furthermore, the process of building the emergency response plan database includes:

[0087] The original emergency response plan document is structured and parsed to extract text content, semantic structure, and metadata information (including plan name, plan type, etc.). Specifically, a document parsing engine can be used to structure and parse emergency response plans in PDF, Word, and other formats to extract text content, semantic structure, and metadata information.

[0088] Based on the semantic structure of the contingency plan document, a sliding window method combined with semantic boundary detection is used to segment the text content into semantically complete contingency plan slices. The length of each contingency plan slice is controlled between 512 and 1024 tokens. For example, the emergency plan is divided into contingency plan slices by segment.

[0089] The content of the plan slices is deduplicated, the terminology is standardized, and an index mapping relationship between the plan slices and the original documents is established.

[0090] For each contingency plan slice, a structured summary is generated using a large language model and then concatenated with the slice. The structured summary includes: a topic summary (within 50 characters), key entities (location, disaster type, measure type, etc.), description of applicable scenarios, priority markers, etc., to improve the accuracy of subsequent searches.

[0091] The pre-trained semantic coding model is used to convert the plan slices into high-dimensional vector representations; the vector dimension is set to 768 or 1024 to ensure the full expression of semantic information.

[0092] Establish a triplet storage structure for the plan slice content, summary information, and vector representation.

[0093] It should be noted that contingency plan texts are usually formulated by region, and their text content is highly similar. The main difference lies in the region and the specific warning values. Therefore, existing methods only use slicing to retrieve contingency plans, resulting in a lot of redundant information and low retrieval efficiency and reliability. In this embodiment, a structured summary (i.e., information that reflects the characteristics of the contingency plan document) is generated by using a large language model and then spliced ​​with the slices. This makes it easier to obtain task-related contingency plan content more quickly and accurately in subsequent retrieval processes.

[0094] The retrieval agent retrieves relevant emergency plan content from the emergency plan database. The specific process involves comparing the query vector with each of the plan slice vectors in the database. Specifically, the cosine of the angle between two vectors (i.e., similarity) can be calculated; the closer the value is to 1, the more relevant the content. The top K (e.g., the top 5 or top 10) slices with the highest similarity are selected as candidate content. The candidate content is then re-ranked based on task requirements and slice summaries to further improve accuracy. For example, the specific intent of the task; whether the plan slice summary fits the context; and whether it contains key entities. The final ranking score is a weighted average of these factors.

[0095] Furthermore, the spatial analysis agent connects to a pre-built spatial knowledge graph, with the input being a spatial query task from the intent parsing agent. This spatial analysis agent parses the spatial association information in the spatial query task and converts it into query statements that can be executed in the spatial knowledge graph; it extracts spatial association elements such as high-risk areas, confluence paths, and potential congestion points based on the query results. Output: Spatial association data, uploaded to the large language model in JSON or tabular format. The query process of the spatial analysis agent includes: starting with the geographic entity in the spatial query task, matching the corresponding geographic region entity from the spatial knowledge graph; linking downwards within the matched geographic region to various monitoring stations and the risk areas monitored by these stations, establishing a "region → station → risk area" link; considering the upstream and downstream relationships of monitoring stations in hydrology or geography, exploring associated stations within 1-3 hops; expanding the impact range of risk areas, associating them with key facilities such as emergency resources; filtering based on risk level and other conditions (e.g., retaining only medium-risk and above); aggregating associated entities generated by multi-hop paths, forming a structured output and feeding it back to the large language model, providing spatial dimension data support for the causal inference of the large language model. Furthermore, during the query process, the graph traversal depth can be limited according to actual needs, frequent query results can be cached, and complex queries can be split into subqueries that can be executed in parallel to optimize the query.

[0096] Furthermore, the spatial knowledge graph adopts a layered construction strategy, sequentially designing and implementing the ontology layer and the relation layer. Each layer collaborates to form a knowledge system with spatial reasoning capabilities. The ontology layer, as the core skeleton of the knowledge graph, is crucial in clearly defining all entity categories participating in spatial analysis and the core attributes corresponding to each entity, providing fundamental support for subsequent relation construction and rule-based reasoning. After screening and definition, the core spatial entity categories, corresponding core attributes, and typical examples involved in the embodiments of this application are shown in Table 1.

[0097] Table 1

[0098]

[0099] The relationship layer is the core connecting spatial entities in the local layer, clarifying the logic of entity associations. It is divided into two categories: spatial relationships (based on GIS topology and location features) and attribute relationships (based on spatial analysis logic). The core relationships are as follows:

[0100] Spatial relationships, based on GIS topology theory and entity location construction, include four core relationships: inclusion relationship: the monitoring buffer zone belongs to a specific geographical area, and the sub-basin includes the monitoring station and the river channel (determined by spatial range); adjacency relationship: sub-basins are adjacent upstream, and the river channels are connected downstream according to the topographic elevation and water flow direction; location relationship: the monitoring station is located in the sub-basin or next to the river channel, and the source of spatiotemporal events falls in a specific sub-basin (based on coordinate matching).

[0101] The attribute relationships, based on spatial analysis logic, clearly define entity attribute dependencies and include four core types of relationships: Equipment-Buffer Zone Association: Monitoring stations are uniquely associated with fixed-radius buffer zones, the radius of which is determined by the equipment type; Topography-Water Flow Association: Topographic data supports water flow path models, and slope directly affects water flow velocity; Watershed-Channel Association: Water from sub-watersheds flows out through specific channels, and channel parameters (width, roughness, etc.) determine the propagation speed of floods / pollutants; Event-Diffusion Association: Events such as floods propagate downstream along the channel, and the propagation time is calculated from channel parameters.

[0102] Furthermore, the data acquisition agent can automatically generate SQL queries and interface with different types of real-time monitoring databases. Input to the data acquisition agent: A description of the data requirement from the intent parsing agent (e.g., "Get hourly rainfall in the XX watershed over the past 6 hours"). The data acquisition agent generates appropriate SQL queries based on the data requirement; executes the SQL queries; and checks the completeness and timeliness of the returned data. Output to the data acquisition agent: Structured real-time monitoring data, uploaded to the large language model in time-series format.

[0103] Furthermore, the process of the data acquisition agent generating and executing SQL query statements includes:

[0104] It receives task instructions from the intent parsing agent, calls the large language model to perform semantic completion and context expansion processing, and performs integrity verification on the original task; for example, for tasks such as "checking water level", it automatically completes missing elements such as time range (e.g., current and the next 24 hours), geographical range (e.g., the user's jurisdiction), and facility type (e.g., reservoir, river) to form a structured extended task.

[0105] Based on the extended task, entity abstraction is performed to extract key semantic entities, including: time entities (such as "past 6 hours"), geographical entities (such as "XX County"), facility entities (such as "monitoring station" and "reservoir"), and monitoring parameter entities (such as "water level" and "rainfall").

[0106] Based on the extracted entity information, table relevance is evaluated from the metadata of the real-time monitoring database to select a set of candidate tables. Evaluation metrics include: entity field matching degree, semantic similarity, historical query frequency, and the strength of join relationships between tables.

[0107] Perform field relevance analysis on each table in the candidate table set to determine the effective query fields it contains, and establish a mapping relationship between fields and semantic entities;

[0108] Based on the foreign key relationships and spatial topology between candidate tables, the optimal table join path is recommended, prioritizing the query scheme with the fewest number of tables and the shortest join level to reduce query complexity.

[0109] Based on the query type determination result (such as simple query, multi-table JOIN, GROUP BY, subquery or complex combination), the predefined standardized SQL template is called, and the large language model generates SQL query statements that conform to the PostgreSQL syntax specification.

[0110] Execute the generated SQL statement to obtain real-time monitoring data; if the execution fails, enter the error handling process, including logging, exception marking and rollback strategy;

[0111] The returned data is formatted to ensure that the time series, numerical units, and null values ​​conform to the system standards before being uploaded to the large language model.

[0112] Furthermore, the analytical predictive agent combines statistical methods and deep learning models to predict future disaster trends. Inputs to the analytical predictive agent include real-time monitoring data from the data acquisition agent and static risk factors from the spatial analysis agent. The analytical predictive agent uses physical rules and data analysis models to assess disaster risk; it predicts the probability of disaster occurrence, peak time, and impact range within the next 6 to 72 hours. Outputs from the analytical predictive agent include the current disaster risk level (e.g., low, medium, high, extremely high), future trend curve, and uncertainty range, uploaded to the large language model in JSON format.

[0113] Furthermore, the statistical analysis and forecasting process includes:

[0114] Statistical analysis process: Analyze and predict the data received by the agent, obtain the real-time monitoring data uploaded by the agent, and perform the following statistical analysis:

[0115] (1) Calculation of basic statistics, including the calculation of statistical indicators such as mean, variance, skewness, kurtosis, and coefficient of variation;

[0116] (2) Time series feature analysis, including trend analysis (the least squares method is used to estimate parameters for linear trends, and polynomial regression or spline regression is used for nonlinear trends), seasonal decomposition (the time series data is decomposed into trend terms, seasonal terms and irregular terms), and periodic detection (autocorrelation analysis, spectrum analysis, wavelet analysis, etc.).

[0117] (3) Correlation analysis, including Pearson correlation coefficient (linear relationship), Spearman rank correlation (non-linear monotonic relationship), partial correlation analysis (controlling the influence of other variables), maximum information coefficient (detecting non-linear relationship), etc.

[0118] Temporal prediction process: The analysis and prediction agent uses an LSTM neural network model combined with the feature optimization capabilities of a large language model for temporal prediction. LSTM architecture design: The input layer contains multivariate temporal data with a window length of 24-72 hours; the hidden layer uses 2-3 layers of LSTM units, with 128-256 hidden units; the output layer is a fully connected layer, outputting predicted values ​​for the next 1-24 hours; a regularized loss function is used; the training strategy includes generating training samples through a sliding time window, monitoring the validation set loss to prevent overfitting, and employing a cosine annealing or exponential decay learning rate strategy.

[0119] The model incorporates physical constraints, including water level constraints (predicted water level not exceeding riverbed elevation), continuity constraints (upstream and downstream flow satisfying the continuity equation), energy constraints (total energy conservation constraints), and boundary conditions (boundary constraints such as upstream inflow and downstream water level). Simultaneously, the large language model optimizes the prediction features, for example, by adjusting the weights of the model's input features based on historical disaster cases and real-time situations, thereby improving prediction accuracy.

[0120] Optionally, an ensemble prediction model can be employed, combining LSTM neural network models, ARIMA models, and SVR models to weight and fuse the predictions generated by multiple models, resulting in a final fused prediction result. The weight coefficients of each model's prediction result are obtained by training on historical data, and the sum of the weight coefficients equals 1. The analysis and prediction agent uploads the statistical analysis results and time-series prediction results to the large language model, providing data support for subsequent causal inference.

[0121] Furthermore, semantic integration and multi-dimensional causal inference are performed through a large language model to generate a final structured dynamic processing suggestion report, including:

[0122] The large language model receives structured data uploaded by various functional agents, comprehensively considers three factors—the urgency of the information, the scope of its impact, and the credibility of the results—and uses a weighted summation method to calculate the importance of the content and rank it.

[0123] Evidence-based risk attribution analysis, using a large language model, integrates multi-source observational data and background knowledge to identify the root causes of disasters; reveals key factors driving trends and their pathways; performs multi-hop causal chain deduction based on spatial topological relationships and evolutionary patterns to identify chain reaction risks in advance (such as rainstorms triggering flash floods, flash floods leading to landslides, and landslides blocking river channels, thus exacerbating flood disasters); provides logical justifications for response recommendations, clarifying the connection between each recommendation and data, contingency plans, and spatial relationships; and intelligently assesses predicted trends and key time nodes, marking risk inflection points (such as the time when water levels are about to exceed the warning line).

[0124] The large language model selects the corresponding report template based on the risk level (general, significant, major, extremely major) and the disaster type (flash flood, mudslide, landslide, etc.);

[0125] The large language model adjusts the level of detail and the use of professional terminology based on the user's role (emergency commander, frontline rescuer, grassroots manager). For example, it provides macro-level decision-making suggestions and overall risk assessments for emergency commanders, and provides specific handling procedures and safety precautions for frontline rescuers.

[0126] The large language model in this application embodiment can adopt model architectures such as Qwen3-30B-A3B-Instruct-2507 and DeepSeek-R1. Taking the Qwen3-30B-A3B-Instruct-2507 model as an example, this model is a large language model built on an improved Decoder-only Transformer architecture. It adopts a Pre-Norm normalization strategy, RoPE (Rotation Position Encoding), and Dual Chunk Attention and sparse attention mechanisms that support 256K ultra-long contexts. Its training data covers massive multilingual and multi-domain corpora, including web pages, books, papers, code, and dialogue texts, and has undergone rigorous quality filtering, deduplication, and privacy anonymization. In terms of training methods, the model first learns a general language representation through large-scale self-supervised pre-training, and then combines instruction fine-tuning and human preference alignment techniques to improve its performance in complex tasks, logical reasoning, programming, and multilingual scenarios. It also optimizes inference efficiency through non-thinking mode, significantly reducing computational overhead while ensuring output quality. It should be noted that the embodiments of this application use current mainstream large language models (such as Qwen3-30B-A3B-Instruct-2507, DeepSeek-R1, etc.) as examples for description, but this is only for better explanation of the present invention and not for limiting the present invention. Those skilled in the art should understand that the multi-agent collaborative framework proposed in the embodiments of this application is essentially an interface with the inference and generation capabilities of large-scale pre-trained models. Therefore, any basic model that can provide similar capabilities, including but not limited to large language models, multimodal large models, and visual large models, can replace the specific models in the embodiments of this application to implement this application.

[0127] The dynamic decision-making method proposed in this application achieves distributed and efficient execution of complex disaster decision-making tasks through a multi-agent collaborative framework. Each functional agent focuses on its area of ​​expertise, improving the accuracy and efficiency of task processing. Using a large language model as the core reasoning engine, it achieves deep semantic integration and multi-dimensional causal inference of multi-source heterogeneous data, breaking through the "no reasoning" decision-making limitations of traditional systems. This method can not only accurately understand natural language decision-making intentions, intelligently match emergency plans, analyze geographic entity topological relationships, automatically plan data acquisition paths, and perform multi-dimensional statistical analysis and hydrological time series prediction, but also generate structured and personalized disposal suggestions containing complete reasoning chains, clarifying the logical basis of each measure, and improving the interpretability and credibility of the decision-making process. It effectively solves the key technical bottlenecks of traditional emergency systems, achieving a leapfrog evolution from "passive response" to "proactive prediction," from "data display" to "causal insight," and from "mechanical execution" to "cognitive decision-making," providing a new technical basis and engineering practice path for the intelligent prevention and control of multi-factor coupled disasters such as flash floods and debris flows.

[0128] Based on the same technical concept described above, this application also proposes a dynamic decision-making system for flash floods and debris flows based on multi-agent and LLM principles, such as... Figure 2 As shown, the dynamic decision-making system 200 includes:

[0129] Multi-agent cooperative module 201, comprising:

[0130] The Intent Analysis Agent 2011 is used to parse the intent of a decision-making task and decompose it into multiple sub-tasks. It then matches each sub-task with a corresponding functional agent and issues task execution instructions. The specific parsing process is as described in the dynamic decision-making method above and will not be repeated here. Optionally, the Intent Analysis Agent 2011 integrates a semantic understanding interface from a large language model to improve the accuracy of intent parsing.

[0131] The retrieval agent 2012 is used to retrieve relevant emergency plan content from a pre-built emergency plan database based on the parsed decision intent and upload the retrieval results to the large language model inference module 202. The construction and retrieval process of the emergency plan database are as described in the dynamic decision-making method above, and will not be repeated here.

[0132] The spatial analysis agent 2013 is used to query relevant spatial relationships from a pre-built spatial knowledge graph based on the parsed decision intent, and upload the query results to the large language model inference module 202. The construction of the spatial knowledge graph and the query process are as described in the dynamic decision-making method above, and will not be repeated here.

[0133] The data acquisition agent 2014 is used to generate and execute SQL query statements based on the parsed decision intent, acquire real-time monitoring data, and upload the real-time monitoring data to the large language model inference module 202. The process of generating and executing SQL query statements is as described in the dynamic decision-making method above, and will not be repeated here.

[0134] The predictive agent 2015 performs statistical analysis and time-series prediction based on real-time monitoring data to obtain the current disaster risk assessment results and predicted trends, and then uploads these results and trends to the large language model inference module 202. The statistical analysis and time-series prediction processes are as described in the dynamic decision-making method section above and will not be repeated here.

[0135] Additionally, the large language model reasoning module 202 utilizes the large language model to receive decision-making tasks, emergency plan content, spatial relationships, real-time monitoring data, and current disaster risk assessment results and prediction trends uploaded by various functional agents. It then performs semantic integration and multi-dimensional causal inference to generate and output a final structured dynamic processing suggestion report. The specific structure of the large language model and the fusion and reasoning process are as described in the dynamic decision-making method above, and will not be repeated here.

[0136] In the multi-agent collaboration module 201, each functional agent and the large language model reasoning module 202 achieve information interaction and collaborative scheduling through an agent communication protocol.

[0137] Based on the same technical concept described above, this application also proposes an electronic device, such as... Figure 3 As shown, the electronic device 300 includes: a memory 310, a processor 320, and a computer program A311 stored in the memory 310 and executable on the processor 320. When the processor 320 executes the computer program A311, it performs the following steps:

[0138] The intent parsing agent performs intent parsing on the decision-making task and decomposes it into multiple sub-tasks, and matches the corresponding functional agent to each sub-task.

[0139] Based on the decision intent obtained from the analysis, the retrieval agent retrieves relevant emergency plan content from a pre-built emergency plan database;

[0140] Based on the decision intent obtained from the parsing, the spatial analysis agent queries the pre-constructed spatial knowledge graph to obtain relevant spatial relationships;

[0141] The data acquisition agent generates and executes SQL query statements based on the parsed decision intent to acquire real-time monitoring data;

[0142] The analytical and predictive agent performs statistical analysis and time-series prediction based on real-time monitoring data to obtain the current disaster risk assessment results and prediction trends.

[0143] The decision-making tasks, emergency plan content, spatial relationships, real-time monitoring data, and current disaster risk assessment results and prediction trends are input into a large language model for semantic integration and multi-dimensional causal inference, generating a final structured dynamic processing suggestion report and outputting it.

[0144] Optionally, when the processor 320 executes the computer program A311, it can implement any of the embodiments in the corresponding examples of the above-described dynamic decision-making method.

[0145] It should be noted that the electronic device proposed in this application embodiment is a device used to implement the above-mentioned dynamic decision-making method. Therefore, based on the above-mentioned dynamic decision-making method proposed in this application embodiment, those skilled in the art can understand the specific implementation method and various variations of the electronic device in this application embodiment. Therefore, how the electronic device specifically implements the above-mentioned dynamic decision-making method will not be described in detail here. Any electronic device used by those skilled in the art to implement the above-mentioned dynamic decision-making method falls within the scope of protection of this application.

[0146] Based on the same technical concept described above, embodiments of this application also propose a computer-readable storage medium, such as... Figure 4 As shown, the computer-readable storage medium 400 stores a computer program B411, which, when executed by a processor, performs the following steps:

[0147] The intent parsing agent performs intent parsing on the decision-making task and decomposes it into multiple sub-tasks, and matches the corresponding functional agent to each sub-task.

[0148] Based on the decision intent obtained from the analysis, the retrieval agent retrieves relevant emergency plan content from a pre-built emergency plan database;

[0149] Based on the decision intent obtained from the parsing, the spatial analysis agent queries the pre-constructed spatial knowledge graph to obtain relevant spatial relationships;

[0150] The data acquisition agent generates and executes SQL query statements based on the parsed decision intent to acquire real-time monitoring data;

[0151] The analytical and predictive agent performs statistical analysis and time-series prediction based on real-time monitoring data to obtain the current disaster risk assessment results and prediction trends.

[0152] The decision-making tasks, emergency plan content, spatial relationships, real-time monitoring data, and current disaster risk assessment results and prediction trends are input into a large language model for semantic integration and multi-dimensional causal inference, generating a final structured dynamic processing suggestion report and outputting it.

[0153] Optionally, when the computer program B411 is executed by the processor, it can implement any of the embodiments corresponding to the above-described dynamic decision-making method.

[0154] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0155] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application 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.

[0156] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. 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... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0157] 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.

[0158] 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.

[0159] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of this application. It should be understood that the above description is only a specific embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A dynamic decision-making method for flash floods and debris flows based on multi-agent and LLM approaches, characterized in that, This dynamic decision-making method is based on a pre-constructed multi-agent collaborative framework and a large language model. The multi-agent collaborative framework includes an intent parsing agent, a retrieval agent, a spatial analysis agent, a data acquisition agent, and an analysis and prediction agent. It achieves information interaction and collaborative scheduling between each functional agent and the large language model through an agent communication protocol, including: The intent parsing agent performs intent parsing on the decision-making task and decomposes it into multiple sub-tasks, and matches each sub-task with a corresponding functional agent. The retrieval agent retrieves relevant emergency plan content from a pre-built emergency plan database based on the parsed decision intent. The spatial analysis agent retrieves relevant spatial relationships from a pre-constructed spatial knowledge graph based on the parsed decision intent. The data acquisition agent generates and executes SQL query statements based on the parsed decision intent to obtain real-time monitoring data; The analytical and predictive agent performs statistical analysis and time-series prediction based on the real-time monitoring data to obtain the current disaster risk assessment results and prediction trends. The decision-making tasks, emergency plan content, spatial relationships, real-time monitoring data, and current disaster risk assessment results and prediction trends are input into the large language model for semantic integration and multi-dimensional causal inference, generating a final structured dynamic processing suggestion report and outputting it.

2. The dynamic decision-making method for flash floods and debris flows based on multi-agent and LLM as described in claim 1, characterized in that, The aforementioned intention parsing and decomposition of the decision-making task into multiple sub-tasks includes: Load the pre-built business knowledge base and call the semantic understanding interface of the large language model to identify key semantic elements in the decision-making task, including geographical entities, time range, disaster type and task intent category; For vague or incomplete expressions, semantic completion and integrity verification are performed by combining user role, current location or historical behavior context through the large language model to obtain structured semantic parsing results; Based on the semantic parsing results, a predefined task template is matched to decompose the original decision task into several executable sub-tasks, and the dependency relationships between the sub-tasks are constructed to form a task dependency graph. Each subtask is equipped with a corresponding data requirement list and recommended analysis methods, and a corresponding functional agent is matched to issue task execution instructions.

3. The dynamic decision-making method for flash floods and debris flows based on multi-agent and LLM as described in claim 1, characterized in that, The process of constructing the emergency response plan database includes: The original contingency plan document is structured and parsed to extract text content, semantic structure, and metadata information; Based on the semantic structure, the sliding window method combined with semantic boundary detection is used to segment the text content into semantically complete pre-plan slices; The content of the plan slices is deduplicated, the terminology is standardized, and an index mapping relationship between the plan slices and the original documents is established. For each of the aforementioned plan slices, a structured summary is generated and then spliced ​​with the slice; The pre-trained semantic coding model is used to convert the plan slices into high-dimensional vector representations; Establish a triplet storage structure that includes the content of the contingency plan slices, summary information, and high-dimensional vector representation.

4. The dynamic decision-making method for flash floods and debris flows based on multi-agent and LLM as described in claim 3, characterized in that, The process of retrieving relevant emergency plan content from a pre-built emergency plan database includes: Convert the task instructions issued by the intent parsing agent into query vectors; The query vector is compared one by one with all the plan slice vectors in the emergency plan database, and the top K with the highest similarity are selected from all the plan slice vectors as candidate content. The candidate content is sorted a second time based on task requirements and slice summary to determine the final emergency plan content and upload it to the large language model.

5. The dynamic decision-making method for flash floods and debris flows based on multi-agent and LLM as described in claim 1, characterized in that, The spatial knowledge graph adopts a layered construction strategy, completing the design and implementation of the ontology layer and the relation layer in sequence. Each layer works together to form a knowledge system with spatial reasoning capabilities. The ontology layer serves as the core skeleton of the knowledge graph, used to clearly define all entity categories participating in spatial analysis and the core attributes corresponding to each entity. The relationship layer is the core connecting the spatial entities in the local layer, clarifying the entity association logic and dividing it into two categories: spatial relationships and attribute relationships. The spatial relationships are constructed based on GIS topology theory and entity location, including: the monitoring buffer zone belongs to a specific geographical area, and the sub-basin contains monitoring stations and waterways; adjacency relationships: sub-basins are adjacent upstream, and waterways are connected downstream according to topographic elevation and water flow direction; location relationships: monitoring stations are located in sub-basins or beside waterways, and the source of spatiotemporal events falls in a specific sub-basin; the attribute relationships are based on spatial analysis logic, clarifying entity attribute dependencies, including: equipment-buffer zone association, topography-water flow association, watershed-waterway association, and event-diffusion association.

6. The dynamic decision-making method for flash floods and debris flows based on multi-agent and LLM as described in claim 5, characterized in that, The process of retrieving relevant spatial relationships from a pre-constructed spatial knowledge graph includes: Starting with the geographic entities in the task instructions issued by the intent parsing agent, the corresponding geographic region entities are matched from the spatial knowledge graph; By linking down the matching geographic regions, various monitoring stations and the risk areas monitored by each monitoring station are included, a "region → station → risk area" link is established. Based on the upstream and downstream relationships of monitoring stations in terms of hydrology or geography, explore related stations within 1-3 jumps; Expand the scope of the risk area and link it to key facilities; Screening is based on risk level criteria; The associated entities generated by multi-hop paths are aggregated to form a structured output.

7. A dynamic decision-making method for flash floods and debris flows based on multi-agent and LLM as described in claim 1, characterized in that, The generation and execution of the SQL query statement includes: Based on the task instructions issued by the intent parsing agent, the large language model is invoked to perform semantic completion and context expansion processing, and the integrity of the original task is verified to form a structured extended task. Based on the extended task, entity abstraction is performed to extract key semantic entities; Based on the extracted entity information, table relevance is evaluated from the real-time monitoring database to select a set of candidate tables. Perform field relevance analysis on each candidate table in the candidate table set to determine the effective query fields in the candidate tables and establish a mapping relationship between fields and semantic entities; Based on the foreign key relationships and spatial topology between the candidate tables, the optimal table join path is recommended; Based on the comprehensive query type judgment results, a predefined standardized SQL template is invoked, and the large language model generates an SQL query statement. Execute the generated SQL query to obtain real-time monitoring data; if the execution fails, proceed to the error handling process. The returned data is formatted and then uploaded to the large language model.

8. A dynamic decision-making method for flash floods and debris flows based on multi-agent and LLM as described in claim 1, characterized in that, The analytical predictive agent combines statistical methods and deep learning models to predict future disaster trends. It receives real-time monitoring data from the data acquisition agent and static risk factors from the spatial analysis agent. It uses physical rules and data analysis models to assess disaster risks and predict the probability of disaster occurrence, peak time, and impact range within future time windows. It outputs the current disaster risk level, future disaster trends, and uncertainty interval to the large language model.

9. A dynamic decision-making method for flash floods and debris flows based on multi-agent and LLM according to any one of claims 1-8, characterized in that, The generation of the final structured dynamic processing recommendation report includes: The large language model receives structured data uploaded by various functional agents, comprehensively considers three factors—the urgency of the information, the scope of its impact, and the credibility of the results—to calculate the importance of the content and rank it. Evidence-based risk attribution analysis integrates multi-source observation data and background knowledge using a large language model to identify the root causes of disasters; reveals key factors driving trends and their pathways; performs multi-hop causal chain deduction based on spatial topological relationships and evolutionary patterns to identify chain reaction risks in advance; provides logical justification for disposal recommendations; and conducts intelligent analysis of predicted trends and key time nodes, marking risk inflection points. The large language model selects the corresponding report template based on the risk level and disaster type; The large language model adjusts the level of detail and the use of technical terms based on the user's role, generating a structured and dynamically processed suggestion report.

10. An intelligent decision-making system for the prevention and control of flash floods and debris flows, characterized in that, include: A multi-agent collaborative module, comprising: The intent analysis agent is used to parse the intent of the decision task and decompose it into multiple sub-tasks, match the corresponding functional agent for each sub-task and issue task execution instructions. The retrieval agent is used to retrieve relevant emergency plan content from the pre-built emergency plan database based on the parsed decision intent and upload the retrieval results to the large language model inference module. A spatial analysis agent is used to query relevant spatial relationships from a pre-built spatial knowledge graph based on the parsed decision intent and upload the query results to the large language model inference module. The data acquisition agent is used to generate and execute SQL query statements based on the parsed decision intent, acquire real-time monitoring data and upload it to the large language model inference module. An analytical and predictive agent is used to perform statistical analysis and time-series prediction based on the real-time monitoring data, obtain the current disaster risk assessment results and prediction trends, and upload the assessment results and prediction trends to the large language model inference module. In addition, the large language model reasoning module uses the large language model to receive the decision-making tasks, emergency plan content, spatial correlations, real-time monitoring data, and current disaster risk assessment results and prediction trends uploaded by each functional agent, and performs semantic integration and multi-dimensional causal inference to generate and output the final structured dynamic processing suggestion report. In the multi-agent collaboration module 201, each functional agent and the large language model inference module achieve information interaction and collaborative scheduling through an agent communication protocol.

Images