Disaster early warning system and method based on large language model multi-agent

Abstract

The application discloses a disaster early warning system and method based on a large language model multi-agent, which comprises a situation awareness agent, an execution agent, a field expert agent group, a system shared state module and a tool library module, the situation awareness agent is used for monitoring data and triggering a warning workflow signal, the execution agent is used for decomposing and routing tasks, the field expert agent group comprises an emergency management agent and a plurality of expert agents in different fields, and is used for executing professional tasks in different fields and writing back task analysis results in different fields to a unified state, the situation awareness agent, the execution agent and the field expert agent group call corresponding tools in the tool library module and interact information through the system shared state module, and early warning information is automatically generated according to a preset template by the emergency management agent. The application can improve the accuracy and consistency of early warning research and judgment, and shorten the time from risk discovery to forming an executable action plan.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence and natural disaster emergency management, and particularly relates to a disaster early warning system and method based on large language model multi-agents. Background Art

[0002] Currently, in the early warning and emergency management of natural disasters, especially meteorological disasters such as typhoons and heavy rains, the commonly adopted technology is the multi-source information integration and threshold early warning technology based on numerical prediction models, monitoring station networks, and geographic information systems. A typical approach is that meteorological, water conservancy, natural resources and other departments respectively establish information systems in their own industries. Through numerical weather prediction models, flood evolution models, geological disaster susceptibility models, etc., the rainfall, wind field, water level, landslide risk, etc. in the future period are predicted. Then, according to the pre-established index system and classification standards (such as rainfall, wind force, water level exceeding a certain threshold), the corresponding early warning level is triggered, and the early warning information is pushed through the internal business system of the department or the release channels for the public.

[0003] In order to improve the coordination efficiency, some regions have built comprehensive disaster early warning and emergency command platforms, and tried to converge the data of multiple departments such as meteorology, water conservancy, emergency, and natural resources on the same platform to achieve the centralized display, consultation and joint consultation of early warning information. Such platforms usually integrate the data of each department's business system through interfaces, provide unified map display, plan management, task flow and consultation record functions. Some systems also introduce rule engines or workflow engines to manage the key links such as early warning release, consultation start, and plan start in a process-oriented manner.

[0004] In addition, with the development of artificial intelligence technology, in the existing technologies, there has also been an exploration of using machine learning models or knowledge graph technology to evaluate disaster risks and intelligently interpret early warning information. For example, training loss prediction models based on historical disaster situations and meteorological data, or constructing a disaster scenario knowledge base to assist in judgment. There are also public technologies that propose using conversational artificial intelligence to provide auxiliary functions such as policy query, plan query, and document writing for emergency personnel. However, most of these technologies still take a single model or a single task as the core, mainly solve the intelligent problems in a certain link, and have not yet formed a full-process coordination framework that runs through "monitoring - prediction - evaluation - decision - action".

[0005] Objectively speaking, the above-mentioned existing technologies have the following problems and deficiencies in practical applications:

[0006] (1) Departmental fragmentation and semantic inconsistency: Although the integrated platform integrates data from multiple departments on the interface, the underlying layer is still a simple “scraping” of multiple independent systems. The indicator systems, professional terms and data structures used by each department are quite different. There is a lack of unified semantic representation and shared state model, which makes it difficult for cross-departmental automatic reasoning and collaborative analysis in a unified semantic space. It still relies heavily on manual consultation.

[0007] (2) Rigid business processes and insufficient response time: Existing early warning processes usually follow a fixed sequence, requiring multi-level experts and managers to review and sign off at each level. Even if some systems introduce workflow engines, they are mainly digitized and form-based versions of existing manual processes, making it difficult to flexibly adjust the analysis path and task allocation according to the development of the disaster and the needs of temporary decision-making. This results in a long time from risk identification to the formation of an executable action plan, which is not conducive to seizing the "golden early warning time".

[0008] (3) Rigid early warning rules and difficulty in depicting the impact of complex disasters: Existing technologies mostly adopt a tiered early warning method based on thresholds. They mainly rely on single or a few physical indicators such as rainfall, wind speed, and water level to determine the early warning level. It is difficult to fully consider the amplification or weakening effect of different regions' exposure and vulnerability of disaster-bearing bodies and differences in infrastructure on the disaster impact. It is insufficient in depicting the complex and cascading disaster scenarios such as typhoon superimposed with rainstorms, flash floods, geological disasters, and traffic interruptions, which affects the quantitative assessment of the specific impact range and severity.

[0009] (4) Limited intelligent decision support capabilities, making it difficult to automatically orchestrate cross-departmental tasks on demand: Most existing rule-based or single-model-based decision support systems can only provide fixed response suggestions under preset scenarios, lacking the ability to automatically decompose tasks according to natural language requirements, call different departmental models or tools, and perform multiple rounds of reasoning and cross-verification. When faced with complex, cross-departmental integrated tasks such as "providing a cross-city and county preventive transfer plan and a cross-regional material dispatch plan for a certain area within the next 24 hours," existing technologies often cannot complete these tasks automatically through the system, still requiring a large amount of manual coordination and repeated communication.

[0010] (5) Existing conversational AI applications remain at the level of partial assistance: Although there have been attempts to use large models or chatbots in the field of emergency response, they are mostly limited to static auxiliary functions such as policy Q&A, knowledge retrieval or document generation. They lack deep coupling with real-time monitoring and forecasting data, professional simulation models and departmental business systems, and also lack multi-agent collaboration mechanisms. It is difficult to undertake the core work of cross-departmental early warning and judgment and action plan generation for institutions while ensuring controllability and traceability.

[0011] Therefore, existing technologies cannot effectively solve the problem of how to use artificial intelligence technology to flexibly, controllably, and traceably coordinate and support the entire early warning process in a unified semantic and state space under the coupled scenarios of multiple departments, multiple data sources, and multiple disasters. It is difficult to meet the actual needs of refined disaster risk assessment and rapid formation of actionable plans. Summary of the Invention

[0012] To overcome the shortcomings of the above-mentioned technologies, this invention provides a disaster early warning system and method based on a large language model and multiple agents. This invention constructs a disaster early warning system comprising situational awareness agents, execution agents, and domain expert agents. It achieves unified orchestration of multi-source monitoring and forecasting data and professional models from meteorology, water resources, natural resources, emergency management, transportation, and power sectors. Under a shared semantic space and unified state representation, it automatically completes key steps such as disaster risk identification, early warning level determination, impact assessment, and cross-departmental action plan generation. This improves the accuracy and consistency of early warning analysis, shortens the time from risk discovery to the formation of an executable action plan, and enhances the flexibility, controllability, and traceability of the early warning and emergency response processes.

[0013] Terminology Explanation:

[0014] 1. API: Application Programming Interface.

[0015] The technical solution adopted by this invention to overcome its technical problems is:

[0016] A disaster early warning system based on a large language model and multiple agents includes:

[0017] The situational awareness agent is used to continuously monitor multi-source observation and forecast data of the target area, update the environmental status and identify anomalies, and generate an early warning workflow signal when preset triggering conditions are met.

[0018] The executing agent is at least used to receive situational awareness agents or external task requests, decompose and route tasks, and conduct cross-agent discussions and negotiations based on task results to reach a consensus.

[0019] A group of domain expert intelligent agents, including an emergency management intelligent agent and several expert intelligent agents from different domains, is used to receive and execute tasks assigned by the execution intelligent agents and write back the task analysis results of each domain to a unified state.

[0020] The system shared state module is used to provide each intelligent agent with a shared system state, a unified information interaction hub, and a semantic space;

[0021] The tool library module provides the necessary tools for each agent.

[0022] Among them, the situational awareness agent, the execution agent, and the domain expert agent group call the corresponding tools in the tool library module and interact with information through the system shared status module. Finally, the emergency management agent automatically generates early warning information according to the preset template.

[0023] Furthermore, expert agents in different domains include at least one or more of the following agents:

[0024] Meteorological intelligent agents, housing and construction intelligent agents, power intelligent agents, natural resource intelligent agents, water conservancy intelligent agents, maritime intelligent agents, transportation intelligent agents, marine economy intelligent agents, and agricultural intelligent agents. These are just some commonly used intelligent agents, and the group is not limited to them. In addition to emergency management intelligent agents, domain expert intelligent agent groups can select at least one or more intelligent agents from various fields, depending on the specific circumstances.

[0025] Furthermore, the situational awareness agent, the execution agent, and the domain expert agent group are built on the same large language model and uniformly adopt the MRKL–ReAct reasoning framework. The reasoning process includes the knowledge activation stage, the tool selection stage, and the result reflection stage.

[0026] Furthermore, the information flow among the situational awareness agents, the execution agents, and the domain expert agents is achieved through the following routing pattern:

[0027] Sequential routing: The output of the previous domain expert agent serves as a trigger condition to automatically trigger the analysis of the downstream domain expert agent;

[0028] Conditional routing: Determines whether to automatically activate the corresponding workflow based on whether the risk indicator exceeds the threshold;

[0029] Parallel routing: Domain expert intelligent agents perform parallel analysis on different disaster dimensions, and the analysis results are synchronized in the system's shared state module.

[0030] Furthermore, the toolkit modules used to construct the situational awareness agents, execution agents, and domain expert agent swarms include the following tools:

[0031] General tools and data access tools: These should include at least data cleaning, quality control, anomaly detection, threshold monitoring, spatial interpolation, and statistical analysis tools.

[0032] Task orchestration and conflict resolution tools: including at least task decomposition strategies, routing and matching algorithms, and multi-round negotiation and conflict resolution mechanisms, used to transform natural language tasks into executable sequences of subtasks;

[0033] Domain-specific analysis tools: used to configure corresponding professional models and APIs, which are invoked by a group of domain expert intelligent agents through a unified tool interface.

[0034] Furthermore, the domain expert intelligent agents included in the domain expert intelligent agent swarm and the tools used by each domain expert intelligent agent are all configured in a modular manner; the system shared state adopts a versioned graph structure, which saves the current value and records the historical trajectory, allowing each intelligent agent to access and update within the scope of its permissions, and supports collaborative reasoning while maintaining the division of roles.

[0035] This invention also discloses a disaster early warning method based on a large language model multi-agent system, applied to the aforementioned disaster early warning system based on a large language model multi-agent system, comprising the following steps:

[0036] By continuously monitoring multi-source observation and forecast data of the target area through situational awareness agents, the environmental status is updated and anomalies are identified, and it is determined whether the preset early warning triggering conditions are met.

[0037] When the warning trigger condition is met or an external task request is received, the executing agent parses the task and decomposes it into several sub-tasks.

[0038] Based on the semantic matching between the subtask description and the capability description of each domain expert agent, the executing agent automatically assigns the subtask to the matching domain expert agent for execution.

[0039] Based on role prompts and information from the system's shared state module, the domain expert agents assigned to tasks call the corresponding tools in the tool library module to execute sub-tasks in parallel, and then write the task analysis results from each domain back to the unified state.

[0040] The executing agent calls upon expert agents from various domains to conduct cross-agent discussions and negotiations on the results of preset tasks in order to reach a consensus. Based on the consensus, the emergency management agent automatically generates early warning information according to a preset template.

[0041] Furthermore, the tasks required for a disaster early warning system based on a large language model and multiple agents include at least assessing the warning level, analyzing disaster risks, generating disaster reports, and developing action plans.

[0042] Furthermore, the executing agent invokes expert agents from various domains to conduct cross-agent discussions and negotiations on the results of the preset task in order to reach a consensus, specifically including:

[0043] The executing agent calls on expert agents from various domains to conduct cross-agent discussions and negotiations on the results of preset tasks. The preset task results include at least generating a disaster report and formulating an action plan. If inconsistencies or logical conflicts are found in the descriptions, targeted sub-tasks are re-initiated or supplementary explanations are requested, forming a multi-round negotiation mechanism of "generation-verification-correction" to ensure that the final conclusion is consistent at the semantic level, that is, a consistent conclusion is formed and written back to the shared state.

[0044] Furthermore, based on the consensus reached, after the emergency management intelligent agent automatically generates early warning information according to a preset template, the process also includes the following steps:

[0045] The generated early warning information is submitted to the emergency command center for manual review and official release, and the relevant decision records in the system's shared status module are updated based on the review feedback.

[0046] The beneficial effects of this invention are:

[0047] (1) Achieve cross-departmental unified semantics and state space, breaking down information silos: By constructing a unified semantic representation and a shared system state module, multi-source heterogeneous data and intermediate inference results from meteorology, water conservancy, natural resources, emergency management, transportation, and electricity are represented in the same framework, and each intelligent agent works collaboratively in a unified context, overcoming the collaborative difficulties caused by the lack of unified indicator systems and terminology systems among departments in the existing technology.

[0048] (2) Improve the accuracy and reliability of early warning assessment: Adopt a multi-agent architecture consisting of situational awareness agents, execution agents and domain expert agents, introduce multi-source model and tool calls, cross-agent semantic reconciliation and discussion negotiation self-reflection mechanism, the conclusions of each domain can be mutually verified and corrected, and compared with systems based on a single model or fixed rules, it has higher consistency and stability in early warning level determination, impact assessment and action suggestions.

[0049] (3) Shorten the time from risk identification to action plan formation: This invention automatically decomposes and routes complex natural language tasks by executing intelligent agents, and calls multiple domain expert intelligent agents and their professional tools in parallel, reducing the links of multiple rounds of manual consultation and manual sorting, thus significantly shortening the overall time from risk triggering to generating early warning suggestions and cross-departmental action plans, which is conducive to seizing the critical time window for disaster response.

[0050] (4) Enhance the flexibility and scalability of early warning and emergency response processes: Through the dynamic task decomposition and conditional routing mechanism driven by the large language model, this invention can adaptively adjust the analysis path and task division according to the real-time situation and decision-making needs, rather than being limited to a pre-fixed linear process; at the same time, expert agents and tools in various fields are configured in a modular way, which makes it easy to access new disaster types, departments or business systems in different regions, and has good scalability and migration capabilities.

[0051] (5) Improve the interpretability and traceability of the system: The present invention explicitly records the reasoning trajectory of the intelligent agent, the tool call sequence, the intermediate state update and the multi-round negotiation process, which can provide a clear source and logical chain for the early warning conclusion and action suggestions, making it easier for post-event review, responsibility identification and continuous optimization of the model and process, and making up for the shortcomings of some existing intelligent systems with a high degree of "black box".

[0052] (6) Support integration with existing business systems and serve long-term capacity building: The system integrates with existing monitoring and forecasting systems, business platforms and contingency plan management systems through standardized interfaces. It can be used for rolling early warning and decision support in daily operation, and can also be used to reenact and compare typical disaster processes using historical data after the fact, providing technical support for contingency plan revision, capacity assessment and optimization of institutional collaboration mechanisms. Attached Figure Description

[0053] Figure 1 This is a schematic diagram illustrating the principle of the disaster early warning system based on a large language model multi-agent according to an embodiment of the present invention.

[0054] Figure 2 This is a comparison chart of test results for the disaster early warning system based on a large language model and multiple agents described in this embodiment of the invention on the task of judging the warning level.

[0055] Figure 3 This is a comparison chart of test results for the disaster early warning system based on a large language model and multiple agents, as described in an embodiment of the present invention, in the task of generating early warning content.

[0056] Figure 4 This is a comparison chart of test results for the disaster early warning system based on a large language model multi-agent described in this embodiment of the invention in assisting human experts to generate real-time early warning information. Detailed Implementation

[0057] To facilitate a better understanding of the present invention by those skilled in the art, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The following are merely exemplary and do not limit the scope of protection of the present invention.

[0058] This invention discloses a disaster early warning system based on a large language model and multiple agents, such as... Figure 1 As shown, it includes situational awareness agents, execution agents, and a group of domain expert agents. The disaster early warning system based on a large language model multi-agent system also includes a system shared state module and a tool library module.

[0059] The situational awareness agent is used to continuously monitor multi-source observation and forecast data of the target area, update environmental conditions and identify anomalies, and generate an early warning workflow signal when preset trigger conditions are met. Specifically, the multi-source observation and forecast data of the target area includes structured and unstructured data. The structured data includes at least meteorological data, water conservancy data, road network topology data, and population data, while the unstructured data includes at least domain-specific knowledge data, emergency response plan data, and legal and regulatory data.

[0060] The executing agent, acting as the "scheduling center" of the early warning system for disasters, is at least used to receive situational awareness agents or external task requests, decompose and route tasks. Specifically, the task is first parsed and then decomposed into several sub-tasks according to the actual situation. The sub-tasks are then routed and assigned to the matching domain expert agents in the domain expert agent group. The executing agent also conducts cross-agent discussions and negotiations on the task results processed by the domain expert agents to form a consensus.

[0061] The domain expert agent ensemble includes an emergency management agent and several expert agents from different domains. These agents are at least used to receive and execute tasks assigned by the execution agent, and to write back the task analysis results from each domain to a unified state. The expert agents from different domains include at least one or more of the following: meteorological agent, housing and construction agent, power agent, natural resources agent, water conservancy agent, maritime agent, transportation agent, marine economy agent, and agricultural agent, but are not limited to these listed entities.

[0062] The system shared state module provides each agent with a shared system state, a unified information interaction hub, and a semantic space. This shared system state is used to centrally describe at least the environmental background, disaster intensity distribution, infrastructure exposure, emergency resource distribution, historical decision records, and situational evolution trends. The agents mentioned here include all those described earlier, namely, situational awareness agents, execution agents, and a group of domain expert agents. In this embodiment, the input of each agent includes the shared system state; that is, the input portion of each agent contains the shared system state so that each agent can receive the shared state of the entire system in real time.

[0063] The tool library module is used to provide the necessary tools for each agent. All tools are exposed to the corresponding agents through a unified tool call and knowledge access interface, and the tool selection and parameter configuration process is explicitly recorded.

[0064] Among them, the situational awareness agent, the execution agent, and the domain expert agent group call the corresponding tools in the tool library module and interact with information through the system shared status module. Finally, the emergency management agent automatically generates early warning information according to the preset template.

[0065] In this embodiment, prompts based on roles, functions, and workflows are defined for each agent in the LangSmith platform, enabling them to better complete their tasks.

[0066] In this embodiment, the tool library module includes the following tools:

[0067] (1) General tools and data access tools: including at least data cleaning, quality control, anomaly detection, threshold monitoring, spatial interpolation, and statistical analysis tools. The general tools and data access tools are tools used by situational awareness agents, and these tools can also be used by domain expert agent groups.

[0068] (2) Task orchestration and conflict resolution tools: including at least task decomposition strategies, routing and matching algorithms, and multi-round negotiation and conflict resolution mechanisms, used to transform natural language tasks into executable sub-task sequences. The task orchestration and conflict resolution tools are the tools used by the executing agent.

[0069] (3) Domain-specific analysis tools: These are used to configure corresponding professional models and APIs, such as numerical weather prediction models, flood or flash flood models, geological hazard susceptibility models, traffic accessibility assessment models, and material scheduling and optimization solution tools, which are called by the domain expert intelligent agent swarm through a unified tool interface. The domain-specific analysis tools are the tools used by the domain expert intelligent agent swarm.

[0070] Preferably, the domain expert intelligent agent group includes several domain expert intelligent agents and the tools used by each domain expert intelligent agent are configured in a modular manner, which facilitates the access to new disaster types, departments or business systems in different regions and has good scalability and migration capabilities.

[0071] In this embodiment, based on the determination of the tool library module, when constructing the above-mentioned intelligent agents, the same large language model is used as the inference engine, and role-based configurations are constructed for each intelligent agent: (1) Large model base and prompt configuration: Each intelligent agent is based on the same large language model and constructs different responsibilities through role-specific system prompts: the situational awareness intelligent agent prompts emphasize continuous monitoring and threshold triggering; the execution intelligent agent prompts emphasize task coordination, planning and integration; the domain expert intelligent agent group prompts encode domain knowledge, collaboration protocols and output format constraints. (2) Unified inference framework and ontology knowledge injection: the three types of intelligent agents uniformly adopt the MRKL-ReAct inference framework. The inference process includes the knowledge activation stage, the tool selection stage and the result reflection stage. In the knowledge activation stage, the corresponding domain knowledge, historical cases and institutional processes are retrieved; in the tool selection stage, appropriate tools and parameters are selected according to the current system state; in the result reflection stage, the initial results are corrected and supplemented, thereby realizing an interpretable and traceable analysis process. (3) Shared state and memory range control: The system state is implemented using a versioned graph structure, which not only saves the current value but also records the historical trajectory, allowing each agent to access and update within the scope of their permissions, and supports collaborative reasoning while maintaining the division of roles.

[0072] Furthermore, in the coordinated construction of multiple agents, this embodiment combines the agents into a hierarchical collaborative working system through a unified state diagram and information flow control mechanism.

[0073] (1) Information flow control and routing modes were designed: that is, three routing modes were designed through a unified state diagram and information flow control mechanism: 1) Sequential routing: the output of the previous domain expert agent is used as the trigger condition to automatically trigger the analysis of the downstream domain expert agent. For example, after the meteorological agent completes the rainfall forecast, the water conservancy agent is triggered to perform flood modeling. 2) Conditional routing: whether the corresponding workflow is automatically activated is determined according to whether the risk indicator exceeds the threshold. For example, when the risk indicator exceeds the threshold, special workflows such as evacuation planning and material scheduling are automatically activated. 3) Parallel routing: the domain expert agent group performs parallel analysis on different disaster dimensions. The analysis results are synchronized in the system's shared state module and then synthesized by the agents.

[0074] (2) Task decomposition and subtask matching: The executing agent performs hierarchical task decomposition based on the trigger signal of the situational awareness agent or the external task situation (i.e., artificial natural language) request, combined with the current system state and resource constraints, and breaks down the complex task into subtasks with input, target and output format constraints; then, based on the semantic matching between the task description and the capability description of each domain expert agent, the subtask is automatically assigned to the matching (i.e. the most suitable) domain expert agent for execution.

[0075] (3) Cross-agent semantic reconciliation and multi-round negotiation: The executing agent discusses and negotiates the results submitted by expert agents from different domains. When inconsistencies or logical conflicts are found, a targeted sub-task is re-initiated or a request for supplementary explanation is made, forming a multi-round negotiation mechanism of "generation-verification-correction" so that the final conclusion reaches a consensus at the semantic level and is written back to the shared state.

[0076] (4) Generation of early warning and action plans for government agencies: Based on collaborative reasoning, the emergency management intelligent agent automatically generates early warning conclusions such as early warning level determination, impact assessment and action suggestions by department and region according to the preset template, outputs situation reports and early warning drafts that meet the standards, and then submits them to the emergency command center for manual review and formal release.

[0077] To verify the effectiveness of the disaster early warning system based on a large language model multi-agent proposed in this invention, this embodiment constructs a systematic testing and evaluation scheme, including but not limited to: (1) Warning level determination test: Select several typhoon scenarios including historical typhoon events and numerical simulations, run the complete warning process, evaluate the consistency between the system's warning level determination and the current expert team's decision, and statistically analyze the accuracy and decision delay under different prompting strategies and different tool configurations. (2) Impact forecast and situation report synthesis test: Under multiple typhoon scenarios, compare different processes such as "multi-agent + semantic reconciliation" (report generated by multi-agent after discussion and optimization), "multi-agent without semantic reconciliation" (report generated by multi-agent without discussion and optimization), "single agent" (report written by a single agent) and "human baseline" (report written by human experts), and evaluate the clarity, structure, decision relevance and analytical completeness of the produced reports by both experts and AI (Artificial Intelligence), while recording the report generation time. (3) System early warning draft generation and end-to-end process testing: For typical historical events (such as strong typhoon process), reproduce the whole link process from observation to forecasting to impact assessment to early warning suggestions to action plan, and examine the operational stability, task arrangement correctness and timeliness of the system described in this invention on the whole early warning chain.

[0078] The above tests verified that the disaster early warning system based on a large language model and multiple agents proposed in this invention can achieve rapid, collaborative, and traceable disaster early warning and emergency decision support in complex, multi-department, and multi-hazard scenarios, laying a technical foundation for its subsequent promotion and application in different regions and disaster types.

[0079] This invention also discloses a disaster early warning method based on a large language model multi-agent system, applied to the aforementioned disaster early warning system based on a large language model multi-agent system, comprising the following steps:

[0080] S1. By continuously monitoring multi-source observation and forecast data of the target area through the situational awareness agent, the environmental status is updated and anomalies are identified, and it is determined whether the preset early warning triggering conditions are met.

[0081] S2. When the warning trigger condition is met or an external task request is received, the executing agent parses the task and decomposes it into several sub-tasks.

[0082] S3. Based on the semantic matching between the subtask description and the capability description of each domain expert agent, the executing agent automatically assigns the subtask to the matching domain expert agent for execution.

[0083] S4. Based on role prompts and information from the system's shared state module, the domain expert agents assigned to tasks call the corresponding tools in the tool library module to execute sub-tasks in parallel, and then write back the task analysis results of each domain to the unified state.

[0084] S5. The executing agent invokes expert agents from various domains to conduct cross-agent discussions and negotiations on the results of the preset tasks. In this embodiment, the tasks required for the disaster early warning system based on a large language model multi-agent system include at least assessing the warning level, analyzing disaster risks, generating disaster reports, and formulating action plans. Additional tasks can be added as needed. However, cross-agent discussions and negotiations are not required when completing tasks such as assessing the warning level and analyzing disaster risks. The preset task results include at least generating disaster reports and formulating action plans. If inconsistencies or logical conflicts are found, targeted sub-tasks are re-initiated or supplementary explanations are requested, forming a multi-round negotiation mechanism of "generation-verification-correction" to ensure that the final conclusion is semantically consistent, i.e., a consistent conclusion is formed and written back to the shared state. Based on the consistent conclusion, the emergency management agent automatically generates early warning information according to a preset template.

[0085] In a preferred embodiment, after the emergency management intelligent agent automatically generates early warning information according to a preset template based on a consensus conclusion, the method further includes the following steps: submitting the generated early warning information to the emergency command center for manual review and formal release, and updating the relevant decision records in the system's shared status module based on the review feedback.

[0086] The following example, using emergency management during a typhoon in a coastal province, illustrates the early warning system and method for disasters based on a large language model and multiple agents as described in this invention.

[0087] The experimental environment for this embodiment is a Windows operating system (CPU: Intel(R) Core(TM) i9-14900HX (2.20 GHz); GPU: NVIDIA GeForce RTX 4070 Laptop 8G). The program design is based on Python 3.11, using the Langgraph library to build a multi-agent system and workflow, the LangSmith platform to build prompt words for the agent system, the Langchain and pydantic libraries to build tool library modules, and OpenAI's API to call the large language model.

[0088] This embodiment constructs a disaster early warning system comprising 12 intelligent agents to conduct practical testing of multi-agent systems in the field of disaster early warning. The 12 intelligent agents are a situational awareness agent, an execution agent, and 10 domain expert agents (emergency management agent, meteorological agent, housing and construction agent, power agent, natural resources agent, water conservancy agent, maritime agent, transportation agent, marine economy agent, and agricultural agent). The 10 domain expert agents are responsible for their respective domain's professional tasks. The situational awareness agent is responsible for data interaction with the real world, triggering early warning workflows, and transmitting and displaying results. The execution agent is responsible for task decomposition and routing, thereby driving the system's deduction and operation.

[0089] This embodiment selected three early warning processes during a typhoon as test scenarios: early warning level assessment based on observation and forecast data, disaster report writing, and assisting human experts in generating real-time early warning information. These tests verified the effectiveness of the disaster early warning system based on a large language model and multi-agent architecture described in this embodiment in terms of accuracy, stability, and practicality.

[0090] 1. Early Warning Level Assessment

[0091] This embodiment selected 100 typhoons (including typhoons generated by a typhoon simulator and historical real typhoons, at a ratio of 9:1) as the test scenario. Figure 2 As shown, during each typhoon, the disaster early warning system generates warning levels (from low to high: 0, IV, III, II, I) hourly based on observation and forecast data according to preset judgment rules, and compares them with the expert judgment results. To this end, this embodiment sets a single agent as the benchmark, meaning that all tasks of domain experts are completed by a single agent. It also compares the results of using professional judgment tools (cooperative tools) with multiple agents under different prompts (i.e., different numbers of warning level judgment examples seen by the domain expert agents, corresponding to zero samples, one sample, and two samples, respectively). The number of samples for zero samples, one sample, and two samples are 0, 1, and 2, respectively.

[0092] Figure 2 In this context, the closer the color is to the true value, the more accurate the analysis. From... Figure 2 As can be seen, the judgment results (with tools) in this embodiment are highly consistent with the judgment results (true values) of human experts in terms of time sequence. As the number of samples seen by each domain expert agent increases, the judgment effect continuously improves. Table 1 shows that although language-based judgment (zero sample, one sample, two samples) lags slightly behind domain expert agents directly calling local tools, its overall accuracy is high. Table 1 also shows that the multi-agent technology based on a large language model can significantly improve judgment efficiency, reducing time by more than 85%. Furthermore, relying on the collaboration of multiple domain expert agents, the accuracy rate of warning level judgment reaches over 94%, basically meeting the needs of emergency management. In Table 1, a single judgment refers to the judgment of the warning level at a specific moment in time.

[0093] Table 1. Test results of the disaster early warning system based on a large language model and multi-agent system on the task of early warning level assessment.

[0094]

[0095] 2. Disaster Report Writing

[0096] This embodiment tested a disaster report writing task during 10 historical typhoons. The most crucial aspect of disaster report writing is the discussion and consultation after each domain expert agent completes its individual disaster analysis. Each domain expert agent reads the analysis reports from other domain expert agents and provides suggestions based on its own understanding and data. After receiving these suggestions, each domain expert agent optimizes its previous disaster analysis results, which are then integrated into a human-friendly report. This embodiment tests the report content by scoring the information relevant to emergency responders during the disaster. The comparison objects include reports written by human experts, reports written by a single agent, reports generated by multiple agents without discussion and optimization, and reports generated by multiple agents after discussion and optimization.

[0097] Figure 3 In this context, "human" refers to a report written by human experts, "single agent" refers to a report written by a single agent, "multi-agent without discussion and optimization" refers to a report generated by multiple agents without discussion and optimization, and "multi-agent with discussion and optimization" refers to a report generated by multiple agents after discussion and optimization. Figure 3As can be seen, the disaster report generated in this embodiment (i.e., the report generated by multiple agents through discussion and optimization) surpasses reports written by human experts, reports written by a single agent, and reports generated by multiple agents without discussion and optimization in all indicators except time (data presentation, clarity, level of detail, comprehensiveness, feasibility of implementation, impact assessment, and response measures), and the improvement is more than 25% compared to reports written by human experts. A comparison with reports generated by multiple agents without discussion and optimization demonstrates that the suggestions and optimizations made by multiple agents effectively improve the quality of the analysis report.

[0098] 3. Assist human experts in generating real-time early warning information.

[0099] The disaster early warning system described in this embodiment can not only perform tasks independently, but also interact with human experts to generate real-time warning information when needed. The reliability of the disaster early warning system's task decomposition, agent routing, and tool selection was tested in four task categories (information retrieval, data computation, visualization, and decision-making). For each task category, 100 common tasks encountered during emergency response were selected to compare the results given by human experts with the results automatically generated by the system, thereby verifying the system's reliability.

[0100] from Figure 4 As can be seen, the disaster early warning system described in this embodiment achieves higher accuracy rates in task decomposition, agent routing, and tool selection across various tasks (information retrieval, data computation, visualization, and decision-making) than the preset standard (set at 80%). This indicates that the disaster early warning system of this embodiment possesses high reliability and accuracy in assisting human experts in generating early warning information.

[0101] The above description only outlines the basic principles and preferred embodiments of the present invention. Those skilled in the art can make many changes and modifications based on the above description, and these changes and modifications should fall within the protection scope of the present invention.

Claims

1. A disaster early warning system based on a large language model and multiple agents, characterized in that, include: The situational awareness agent is used to continuously monitor multi-source observation and forecast data of the target area, update the environmental status and identify anomalies, and generate an early warning workflow signal when preset triggering conditions are met. The executing agent is at least used to receive situational awareness agents or external task requests, decompose and route tasks, and conduct cross-agent discussions and negotiations based on task results to reach a consensus. A group of domain expert intelligent agents, including an emergency management intelligent agent and several expert intelligent agents from different domains, is used to receive and execute tasks assigned by the execution intelligent agents and write back the task analysis results of each domain to a unified state. The system shared state module is used to provide each intelligent agent with a shared system state, a unified information interaction hub, and a semantic space; The tool library module provides the necessary tools for each agent. Among them, the situational awareness agent, the execution agent, and the domain expert agent group call the corresponding tools in the tool library module and interact with information through the system shared status module. Finally, the emergency management agent automatically generates early warning information according to the preset template.

2. The disaster early warning system based on a large language model and multiple agents according to claim 1, characterized in that, Expert agents in different domains include at least one or more of the following agents: Meteorological intelligent agents, housing and construction intelligent agents, power intelligent agents, natural resources intelligent agents, water conservancy intelligent agents, maritime intelligent agents, transportation intelligent agents, marine economy intelligent agents, and agricultural intelligent agents.

3. The disaster early warning system based on a large language model and multiple agents according to claim 1, characterized in that, Situational awareness agents, execution agents, and domain expert agents are all built on the same large language model and uniformly adopt the MRKL–ReAct inference framework. The inference process includes a knowledge activation stage, a tool selection stage, and a result reflection stage.

4. The disaster early warning system based on a large language model and multiple agents according to claim 1, characterized in that, The information flow among situational awareness agents, executive agents, and the group of domain expert agents is achieved through the following routing pattern: Sequential routing: The output of the previous domain expert agent serves as a trigger condition to automatically trigger the analysis of the downstream domain expert agent; Conditional routing: Determines whether to automatically activate the corresponding workflow based on whether the risk indicator exceeds the threshold; Parallel routing: Domain expert intelligent agents perform parallel analysis on different disaster dimensions, and the analysis results are synchronized in the system's shared state module.

5. The disaster early warning system based on a large language model and multiple agents according to claim 1, characterized in that, The toolkit modules used to build swarms of situation-aware agents, execution agents, and domain expert agents include the following tools: General tools and data access tools: These should include at least data cleaning, quality control, anomaly detection, threshold monitoring, spatial interpolation, and statistical analysis tools. Task orchestration and conflict resolution tools: including at least task decomposition strategies, routing and matching algorithms, and multi-round negotiation and conflict resolution mechanisms, used to transform natural language tasks into executable sequences of subtasks; Domain-specific analysis tools: used to configure corresponding professional models and APIs, which are invoked by a group of domain expert intelligent agents through a unified tool interface.

6. The disaster early warning system based on a large language model and multiple agents according to claim 1, characterized in that, The domain expert agent swarm comprises several domain expert agents and the tools used by each domain expert agent are configured in a modular manner; the system's shared state adopts a versioned graph structure, saving the current value and recording historical trajectories, which can be accessed and updated by each agent within its authorized scope, supporting collaborative reasoning while maintaining the division of roles.

7. A disaster early warning method based on a large language model and multiple agents, characterized in that, The disaster early warning system based on a large language model multi-agent as described in any one of claims 1-6 includes the following steps: By continuously monitoring multi-source observation and forecast data of the target area through situational awareness agents, the environmental status is updated and anomalies are identified, and it is determined whether the preset early warning triggering conditions are met. When the warning trigger condition is met or an external task request is received, the executing agent parses the task and decomposes it into several sub-tasks. Based on the semantic matching between the subtask description and the capability description of each domain expert agent, the executing agent automatically assigns the subtask to the matching domain expert agent for execution. Based on role prompts and information from the system's shared state module, the domain expert agents assigned to tasks call the corresponding tools in the tool library module to execute sub-tasks in parallel, and then write the task analysis results from each domain back to the unified state. The executing agent calls upon expert agents from various domains to conduct cross-agent discussions and negotiations on the results of preset tasks in order to reach a consensus. Based on the consensus, the emergency management agent automatically generates early warning information according to a preset template.

8. The disaster early warning method based on a large language model multi-agent according to claim 7, characterized in that, The tasks that a disaster early warning system based on a large language model and multiple agents needs to complete include at least assessing the warning level, analyzing disaster risks, generating disaster reports, and formulating action plans.

9. The disaster early warning method based on a large language model and multiple agents according to claim 8, characterized in that, The executing agent invokes expert agents from various domains to conduct cross-agent discussions and negotiations on the results of a pre-defined task, in order to reach a consensus. Specifically, this includes: The executing agent calls on expert agents from various domains to conduct cross-agent discussions and negotiations on the results of preset tasks. The preset task results include at least generating disaster reports and formulating action plans. If inconsistencies or logical conflicts are found, targeted sub-tasks are re-initiated or supplementary explanations are requested, forming a multi-round negotiation mechanism of "generation-verification-correction" to ensure that the final conclusion is consistent at the semantic level, that is, a consistent conclusion is formed and written back to the shared state.

10. The disaster early warning method based on a large language model multi-agent according to any one of claims 7-9, characterized in that, Based on the consensus reached, after the emergency management intelligent agent automatically generates early warning information according to a preset template, the following steps are also included: The generated early warning information is submitted to the emergency command center for manual review and official release, and the relevant decision records in the system's shared status module are updated based on the review feedback.

Images