Disaster response plan intelligent generation method and device, equipment and medium

Abstract

This application relates to the field of artificial intelligence technology, and discloses a method, apparatus, device, and medium for intelligent generation of disaster response plans. The method includes: acquiring disaster event information; inputting the disaster event information into a pre-trained plan information processing model to generate disaster event attributes based on the disaster event information, obtaining corresponding disaster attribute response information; acquiring opinions from multiple users regarding the disaster attribute response information; generating corresponding candidate response plans based on the opinions; inputting the candidate response plans into the plan information processing model to evaluate the candidate response plans, obtaining corresponding plan evaluation response information; and determining the final disaster response plan based on the plan evaluation response information. The embodiments of this application can improve the logic and readability of the generated disaster emergency response plans.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method, apparatus, equipment and medium for intelligent generation of disaster response plans. Background Technology

[0002] With the development of artificial intelligence technology, Large Language Models (LLMs) have demonstrated powerful natural language processing and knowledge reasoning capabilities. Among related technologies, LLMs have been applied to various emergency decision-making exploration scenarios, providing intellectual support and scientific response suggestions. However, existing LLMs still have significant limitations in knowledge processing within the disaster emergency response domain. Their training corpora are mostly based on general semantics and open-domain text, lacking the ability to map structured emergency attribute data and contextual knowledge specific to certain disaster types. This often leads to semantic biases and logical inconsistencies in emergency knowledge reasoning, risk assessment, and response plan generation, affecting the logic and readability of disaster emergency plans. Summary of the Invention

[0003] The purpose of this application is to provide a method, apparatus, equipment, and medium for intelligently generating disaster response plans, aiming to improve the logic and readability of the generated disaster emergency response plans.

[0004] This application provides a method for intelligently generating disaster response plans, including:

[0005] Obtain disaster event information;

[0006] The disaster event information is input into a pre-trained contingency plan information processing model to generate disaster event attributes based on the disaster event information, and obtain corresponding disaster attribute response information; the contingency plan information processing model is a model based on a large language model, which is trained based on sample event information and corresponding sample event attribute information and sample contingency plans and corresponding sample evaluation response information.

[0007] Obtain feedback from multiple users regarding their responses to the disaster attributes;

[0008] Based on the aforementioned feedback, corresponding candidate response plans are generated;

[0009] The candidate response plans are input into the plan information processing model to evaluate the candidate response plans and obtain the corresponding plan evaluation response information.

[0010] Based on the response information to the aforementioned contingency plan evaluation, the final disaster response plan will be determined.

[0011] In some embodiments, generating disaster event attributes based on the disaster event information includes:

[0012] Based on a preset similarity threshold, information search is performed on the disaster event information in the corresponding domain database to obtain the corresponding disaster attribute query information;

[0013] The disaster attribute query information is integrated into a preset prompt word template to construct prompt words;

[0014] Based on the disaster attribute query information and the prompt words, contextual semantic association and text content generation are performed to obtain the disaster attribute response information.

[0015] In some embodiments, the training method for the contingency plan information processing model includes:

[0016] The large language model to be trained is subjected to training operations until the preset training termination conditions are met, thereby obtaining the pre-planned information processing model; the training operations include:

[0017] The sample event information is input into the current large language model to generate disaster event attributes based on the sample event information, and obtain the corresponding sample attribute response information;

[0018] Determine whether the first information deviation meets the preset deviation condition; the first information deviation is the deviation between the sample event attribute information and the sample attribute response information;

[0019] If not, based on the first information deviation, the model parameters of the current large language model are adjusted to obtain the large language model for the next training operation;

[0020] If so, obtain the sample plan and corresponding sample evaluation information for the response information of the sample attributes;

[0021] The sample proposal is input into the current large language model to evaluate the sample proposal and obtain the corresponding sample evaluation response information.

[0022] Based on the second information bias, the model parameters of the current large language model are adjusted to obtain the large language model for the next training operation; the second information bias is the deviation between the sample evaluation information and the sample evaluation response information.

[0023] In some embodiments, during the training operation, the model parameters of the large language model are frozen, and a trainable low-rank matrix is ​​inserted into the corresponding linear layer of the large language model. Adjusting the model parameters of the current large language model includes:

[0024] The low-rank matrix is ​​adjusted to obtain the updated low-rank matrix;

[0025] When the training termination condition is met, the model parameters of the current large language model and the updated low-rank matrix are fitted to obtain the pre-plan information processing model.

[0026] In some embodiments, before inputting the disaster event information into the pre-trained contingency plan information processing model, the method further includes:

[0027] The sample event information is input in parallel into several preliminary information processing models to perform disaster event attribute generation operations based on the sample event information, thereby obtaining the sample event attribute information; the preliminary information processing models are mainly based on different types of large language models;

[0028] Information mapping processing is performed on the sample event information and the sample event attribute information to make them correspond to each other.

[0029] In some embodiments, generating corresponding candidate response plans based on the opinion information includes:

[0030] The opinions described are clustered to obtain several opinion clusters;

[0031] The core opinion information in the opinion cluster is determined, and the relative swing probability and self-learning probability of the user corresponding to other opinion information are calculated; the relative swing probability is the probability that the user changes their opinion information, and the self-learning probability is the probability that the user supports the core opinion information.

[0032] The opinion information, the relative swing probability, and the autonomous learning probability are input into a preset contingency plan generation model to perform a contingency plan generation operation based on the opinion information, the relative swing probability, and the autonomous learning probability, thereby obtaining the candidate response contingency plan.

[0033] In some embodiments, prior to clustering the opinion information, the method further includes:

[0034] The opinion information is Z-number transformed to obtain opinion information in Z-number format;

[0035] The opinion information in Z-number form is subjected to normal cloud transformation to obtain opinion information in normal cloud form.

[0036] This application embodiment also provides a disaster response plan intelligent generation device, including:

[0037] The first module is used to obtain disaster event information;

[0038] The second module is used to input the disaster event information into a pre-trained contingency plan information processing model to generate disaster event attributes based on the disaster event information and obtain corresponding disaster attribute response information; the contingency plan information processing model is a model based on a large language model, which is trained based on sample event information and corresponding sample event attribute information and sample contingency plans and corresponding sample evaluation response information.

[0039] The third module is used to obtain the opinions input by multiple users regarding the disaster attribute response information;

[0040] The fourth module is used to generate corresponding candidate response plans based on the opinion information;

[0041] The fifth module is used to input the candidate response plan into the plan information processing model to evaluate the candidate response plan and obtain the corresponding plan evaluation response information.

[0042] The sixth module is used to determine the final disaster response plan based on the evaluation response information of the plan.

[0043] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described intelligent disaster response plan generation method.

[0044] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described intelligent disaster response plan generation method.

[0045] The beneficial effects of this application are as follows: By introducing a pre-trained contingency plan information processing model, this model generates corresponding disaster attribute response information based on disaster event information. Then, it obtains opinions from multiple users regarding these responses and generates corresponding candidate response plans based on these opinions. The contingency plan information processing model evaluates these candidate response plans, and finally, based on the evaluation responses, determines the final disaster response plan. Although this contingency plan information processing model is primarily based on a general language model, its training process specifically incorporates sample event information, corresponding sample event attribute information, sample contingency plans, and corresponding sample evaluation responses. This allows the model to overcome the shortcomings of general language models in knowledge processing within the disaster emergency domain, possessing stronger disaster situation understanding and contingency plan evaluation capabilities. Generating corresponding candidate response plans based on opinions from responses to disaster attribute information effectively compensates for potential cognitive blind spots in single models, integrating user experience into the initial generation stage of the plan. The generated candidate response plans are then input back into the contingency plan information processing model for evaluation, and the final disaster response plan is determined based on the evaluation responses. This iterative "model-human-model" feedback loop mechanism ensures that the plan can be continuously optimized and improved during the generation process, providing decision-makers with specific directions for improvement, thereby avoiding major defects that may exist in the final plan and significantly improving the logic and readability of the disaster emergency plan. Attached Figure Description

[0046] Figure 1 This is an application environment diagram of the intelligent generation method for disaster response plans provided in the embodiments of this application.

[0047] Figure 2 This is a flowchart of the disaster emergency response plan generation method provided in the embodiments of this application.

[0048] Figure 3 This is a flowchart of a method for generating disaster event attributes based on disaster event information, provided in an embodiment of this application.

[0049] Figure 4 This is a flowchart of a method for performing training operations provided in an embodiment of this application.

[0050] Figure 5 This is a flowchart of a method for generating candidate response plans provided in an embodiment of this application.

[0051] Figure 6 This is a schematic diagram of the disaster emergency response plan generation device provided in the embodiments of this application.

[0052] Figure 7 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0054] It should be noted that although functional modules are divided in the device schematic diagram and a logical order is shown in the flowchart, in some cases, the steps shown may be performed in a different order than the module division in the device or the order in the flowchart. The terms "first," "second," etc., in the specification, claims, and drawings are used to distinguish similar objects and are not used to describe a specific order or sequence.

[0055] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0056] The intelligent disaster response plan generation method provided in this application can be executed by a computer device, which can be a terminal device or a server. Terminal devices include, but are not limited to, mobile phones, computers, smart home appliances, vehicle terminals, and aircraft. The server can be a standalone physical server, a server cluster consisting of multiple physical servers, a distributed system, or a cloud server. Furthermore, the information, data, and signals involved in this application's embodiments are all authorized by the relevant parties or have received full authorization from all parties involved, and the collection, use, and processing of related data comply with the relevant laws, regulations, and standards of the relevant countries and regions.

[0057] In disaster emergency decision support systems, the knowledge training corpus of large language models mainly comes from general semantics and open-domain texts, lacking the ability to map structured emergency attribute data and contextual knowledge specific to disaster types. This leads to semantic biases and logical inconsistencies when the model performs disaster knowledge reasoning, risk assessment, and response plan generation tasks, thus affecting the logical rigor and readability of disaster emergency plans. Specifically, the general nature of the knowledge training corpus prevents the model from effectively identifying key attribute relationships in disaster events, resulting in broken reasoning chains or semantic ambiguities during the generation of response suggestions. Furthermore, this problem directly weakens the scientific rigor and practicality of the emergency decision support system. For example, in a certain earthquake disaster, disaster information was input into an existing large language model. The model generated disaster attribute response information based on general training corpus, but because it did not integrate structured emergency attributes such as local geological structure data and population distribution information, the generated evacuation route suggestions did not match the actual terrain, causing logical contradictions in the plan content and reducing readability. Specifically, the model failed to accurately map the topological relationship between the epicenter and evacuation sites, resulting in suggested evacuation routes passing through known fault zones. Furthermore, the risk assessment lacked coherence in its projections of secondary disasters, leading to confusion in emergency command instructions.

[0058] If the aforementioned problems are not addressed, the process of generating disaster emergency response plans will continue to be hampered by semantic biases and logical inconsistencies, leading to a decline in the reliability of emergency decision support systems, compromising the applicability of response plans, and increasing uncertainty in disaster response. As a preferred implementation method, the long-term existence of this problem will systematically constrain the efficiency of disaster emergency response, affecting the rational allocation of emergency resources and the timely execution of scientific decisions.

[0059] Based on this, embodiments of this application provide a method, apparatus, device, and medium for intelligent generation of disaster response plans. By combining disaster attribute response information and user opinion information generated by a plan information processing model trained in the field of disaster response, the key characteristics of disaster events can be accurately captured, and the logic and readability of the generated disaster emergency plan can be improved.

[0060] Figure 1 This diagram illustrates the application environment of the intelligent disaster response plan generation method provided in this embodiment. (See also...) Figure 1This method is applied to an intelligent disaster response plan generation system. The system includes a terminal 110 and a server 120. The terminal 110 and server 120 are connected via a network. The terminal 110 can be a desktop terminal or a mobile terminal; a mobile terminal can be at least one of a mobile phone, tablet, or laptop. The server 120 can be a standalone server or a server cluster consisting of several servers. The terminal 110 sends disaster event information to the server 120. The server 120 acquires the disaster event information, inputs it into a pre-trained plan information processing model to generate disaster event attributes based on the disaster event information, obtains corresponding disaster attribute response information, acquires opinions from multiple users regarding the disaster attribute response information, generates corresponding candidate response plans based on the opinions, inputs the candidate response plans into the plan information processing model to evaluate the candidate response plans, and obtains corresponding plan evaluation response information. The plan information processing model is a model primarily based on a large language model, trained on sample event information and corresponding sample event attribute information, sample plans, and corresponding sample evaluation response information.

[0061] It should be understood that Figure 1 The application scenarios shown are merely examples. In practical applications, the intelligent disaster response plan generation method provided in this application embodiment can also be applied to other scenarios. For example, the above-mentioned intelligent disaster response plan generation method can be directly applied to terminal 110. Terminal 110 is used to obtain disaster event information, input the disaster event information into a pre-trained plan information processing model, generate disaster event attributes based on the disaster event information, obtain corresponding disaster attribute response information, obtain opinion information input by multiple users regarding the disaster attribute response information, generate corresponding candidate response plans based on the opinion information, input the candidate response plans into the plan information processing model, evaluate the candidate response plans, and obtain corresponding plan evaluation response information.

[0062] See Figure 2 In one embodiment, a method for intelligently generating disaster response plans is provided. The subject executing the method can be a terminal or a server, including but not limited to steps S201 to S206.

[0063] Step S201: Obtain disaster event information.

[0064] Disaster event information refers to all available data and descriptions related to a specific disaster event, such as event type, time of occurrence, location, scope of impact, and potential risks. This information forms the basis for developing response plans.

[0065] Acquiring disaster event information aims to collect all necessary data related to a specific disaster event. For example, this can be done manually by operators based on on-site reports or news coverage, inputting information such as the type of disaster, time of occurrence, location, and initial impact. Alternatively, it can be achieved by connecting to specific data interfaces, such as meteorological monitoring systems or earthquake monitoring networks, to periodically or in real-time acquire raw sensor data or event reports. Both methods provide the foundational data for subsequent disaster attribute generation.

[0066] In some embodiments, disaster event information can be obtained by using Python Scrapy to crawl websites such as Weibo, news websites, and official announcements, mimicking user browsing behavior by inputting word cookie addresses, and collecting big data on public behavior related to a specific emergency event on social media. After obtaining the disaster event information, a cleaning operation can be performed on the initial disaster event information set to remove text data that is empty, contains garbled text, etc., and also removes English letters, numbers, symbols, URLs, and emojis to ensure the accuracy and reliability of subsequent analysis. After completing data preprocessing and obtaining a word segmentation list, semantic classification, word frequency analysis, and the TF-IDF algorithm are used to extract the attributes and weights of keywords. Semantic classification performs semantic analysis on the words in the text, dividing words into different categories, such as nouns, verbs, and adjectives. The Boson NLP sentiment analysis dictionary is introduced as a training bag-of-words to obtain the sentiment score of the disaster event information. Then, a language term set is used to conduct multi-dimensional sentiment analysis representation on the sentiment score, and the disaster event information is classified according to the sentiment score.

[0067] Step S202: Input the disaster event information into the pre-trained contingency plan information processing model to generate disaster event attributes based on the disaster event information and obtain the corresponding disaster attribute response information.

[0068] The emergency response information processing model is primarily based on a large language model, trained on sample event information and corresponding sample event attribute information, as well as sample emergency plans and corresponding sample evaluation response information. In essence, this model is built upon a large language model, with its core function being to process various types of information related to disaster response plans. This model undergoes specialized pre-training to understand disaster scenarios, generate disaster attributes, and evaluate response plans. The training process utilizes sample event information, corresponding sample event attribute information, sample emergency plans, and corresponding sample evaluation response information, thereby equipping it with professional knowledge and reasoning capabilities specific to the disaster emergency domain. Sample event information refers to the raw data of historical disaster events used to train the model. This data corresponds to actual disaster events, providing a foundation for the model to learn disaster characteristics. Sample event attribute information refers to disaster event attribute data corresponding to the sample event information, which has been manually labeled or confirmed by experts. This data serves as the "standard answer" for the model in the disaster event attribute generation task, guiding the model's training and optimization. Sample contingency plans refer to historical disaster response plans used to train the contingency plan information processing model. These plans are formulated for specific sample events and provide examples for the model to learn the structure and content of the plans. Sample evaluation response information refers to the evaluation results of the contingency plans corresponding to the sample contingency plans, which have been evaluated by humans or experts. This evaluation information serves as the "standard answer" for the model in the contingency plan evaluation task, guiding the model's training and enabling it to accurately assess the effectiveness and rationality of the plans.

[0069] Disaster event attribute generation refers to the process by which the contingency plan information processing model analyzes and extracts key attributes of an input disaster event, such as disaster severity, affected population, infrastructure damage, and secondary disaster risk. This process aims to transform the raw, unstructured description of the disaster event into a structured set of attributes that can be used for subsequent decision-making.

[0070] Disaster attribute response information refers to the detailed description or summary of various attributes of a disaster event output by the contingency plan information processing model after generating the disaster event attributes. This information is usually presented in the form of natural language text, aiming to clearly convey the model's understanding of the disaster event.

[0071] After the acquired disaster event information is fed into the emergency response information processing model, the model can use pre-defined keyword matching rules to directly extract core disaster-related words from the input disaster event information, such as "flood," "earthquake," and "fire," and associate these words with a pre-defined list of disaster attributes. Alternatively, the model can use a simple text classification algorithm to categorize the disaster event information into pre-defined disaster types and output descriptions of common attributes under that type. During training, this emergency response information processing model can employ a general-purpose large language model and use a corpus containing a large amount of disaster-related text for basic training, enabling it to acquire a preliminary understanding of disaster-related text.

[0072] Step S203: Obtain the opinions input by multiple users regarding the disaster attribute response information.

[0073] Feedback information refers to the comments, suggestions, or supplementary explanations offered by multiple users after reviewing responses regarding disaster attributes, based on their professional knowledge, experience, or specific perspectives. This information reflects human understanding of the disaster situation and consideration of potential response measures.

[0074] After the disaster attribute response information is generated, it is presented to multiple users. These users can be emergency management experts, officials, or community representatives. They can directly input supplements, corrections, or questions about the response information in free text. For example, users can point out that a key affected area is missing from the disaster attribute response information generated by the model, or express different opinions on the severity of a certain risk assessment. This feedback is collected by the implementing body and serves as important input for the subsequent generation of candidate response plans.

[0075] Step S204: Based on the feedback information, generate corresponding candidate response plans.

[0076] Candidate response plans refer to disaster response plans that are initially generated based on user feedback and have not yet been finalized. These plans are preliminary results that integrate opinions from multiple parties and require further evaluation and refinement.

[0077] After collecting user feedback, the implementing body needs to transform this feedback into several preliminary candidate response plans. For example, a simple voting mechanism can be used to identify the most frequently mentioned action suggestions in the user feedback and combine them into a preliminary draft plan. Alternatively, the implementing body can pre-set a rule-based expert system that triggers corresponding plan templates based on keywords in the user feedback and fills in relevant information, thereby generating several candidate response plans.

[0078] Step S205: Input the candidate response plan into the plan information processing model to evaluate the candidate response plan and obtain the corresponding plan evaluation response information.

[0079] Contingency plan evaluation refers to the process by which a contingency plan information processing model analyzes and assesses candidate response plans. This process aims to identify potential flaws, risks, or deficiencies in the plans and to determine their feasibility, effectiveness, and completeness.

[0080] The contingency plan evaluation response information refers to the assessment results of the candidate response plans output by the contingency plan information processing model after completing the contingency plan evaluation. This information typically includes the advantages, disadvantages, risk warnings, and improvement suggestions of the contingency plan.

[0081] After the generated candidate response plans are fed back into the response plan information processing model, the model performs preliminary text analysis based on the structure and content knowledge it learned during training. For example, the model can check the fluency and logical coherence of the candidate response plans and identify any obvious contradictions or omissions. The model can then output a text-based evaluation report, pointing out potential problems or areas for improvement for each candidate response plan.

[0082] Step S206: Based on the response information to the plan evaluation, determine the final disaster response plan.

[0083] A disaster response plan is a formal plan that has been determined to be executable and used to guide actual disaster response actions after multiple rounds of evaluation, revision, and validation. This plan is the final product of the entire development methodology.

[0084] Upon receiving the response to the contingency plan evaluation, the implementing body can review and revise the candidate contingency plans based on this information. For example, if the evaluation response points out a deficiency in a certain aspect, the implementing body can adjust the content of the plan to remedy these deficiencies. Alternatively, the implementing body can set a simple threshold: if the number of negative evaluation keywords in the evaluation response exceeds a preset value, the candidate contingency plan is marked as requiring significant modification; otherwise, it is considered basically usable. Through this process, the final disaster response plan is determined.

[0085] The following example will provide a more detailed explanation of the above technical solution:

[0086] Suppose a sudden flood disaster occurs at "Location A". First, the system will obtain disaster event information, such as receiving reports of continuous heavy rainfall and rapidly rising river levels at "Location A" through a real-time data interface connected to the meteorological bureau, and obtaining textual descriptions of water accumulation in some low-lying areas through a news media interface.

[0087] Next, this disaster event information is input into a pre-trained contingency plan information processing model. Based on the disaster knowledge learned during its training, the model analyzes this information and generates disaster event attributes. For example, the model might identify attributes such as "Disaster Type: Flooding," "Affected Area: Location A," "Main Impacts: Rising River Levels, Urban Flooding," and "Potential Risks: People Trapped, Traffic Disruptions," and generate a detailed disaster attribute response, such as: "Location A has experienced severe flooding, mainly caused by continuous heavy rainfall. River levels have exceeded warning lines, and low-lying urban areas are flooded, which is expected to lead to traffic paralysis and the risk of people being trapped."

[0088] Subsequently, the disaster-related response information was presented to multiple users, such as User A from the local emergency command center, User B from the water resources department, and User C, a community representative. User A might add comments like, "We need to pay attention to the upstream reservoir's discharge situation." User B might point out, "There are multiple chemical plants in this area, and the risk of secondary disasters needs to be assessed." User C might report, "There are a large number of elderly people living alone in the community, making evacuation difficult." The system collects these opinions and information.

[0089] Based on the collected feedback, the system begins to generate corresponding candidate response plans. For example, the system may simply summarize the opinions of users A, B, and C to form a preliminary draft plan, which may include items such as: "activating the upstream reservoir flood discharge early warning mechanism", "dispatching a professional team to assess the risks of the chemical plant", and "organizing community volunteers to assist elderly people living alone in evacuation".

[0090] The candidate response plan is then input into the plan information processing model for evaluation. The model analyzes each item of the plan and, combined with its understanding of flood disasters, generates an evaluation response. For example, the model might point out: "The measures for risk assessment of chemical plants in the plan are too general and lack specific implementation details; the resource allocation for the evacuation of elderly people living alone is not clearly defined."

[0091] Finally, based on the evaluation and response information of the contingency plan, decision-makers (such as the head of the emergency command center) can manually review and revise the candidate response plans. According to the model's evaluation, decision-makers may instruct relevant departments to refine the specific steps of the chemical plant risk assessment and specify which resources to allocate for the evacuation of elderly people living alone. After this round of revisions, the final disaster response plan is determined, making it more targeted, feasible, and comprehensive.

[0092] Based on the above examples, the technical concept of this embodiment demonstrates a significant technical contribution. Existing large language models have limitations in knowledge processing in the field of disaster emergency response. Their knowledge training corpora are mostly based on general semantics and open-domain text, lacking the ability to map structured emergency attribute data and contextual knowledge for specific disaster types. This often leads to semantic deviations and logical inconsistencies in the model's emergency knowledge reasoning, risk assessment, and response plan generation, affecting the logic and readability of disaster emergency plans.

[0093] This embodiment introduces a pre-trained contingency plan information processing model. While primarily based on a general language model, its training process specifically incorporates sample event information, corresponding sample event attribute information, sample contingency plans, and corresponding sample evaluation responses. This enables the model to overcome the shortcomings of general language models in knowledge processing within the disaster emergency domain, possessing stronger disaster situation understanding and contingency plan evaluation capabilities. For example, in the aforementioned flood disaster example, the model can accurately identify the correlation between "rising river water levels" and "urban flooding," generating professional disaster attribute response information rather than a general text summary. Furthermore, this embodiment obtains opinions from multiple users regarding the disaster attribute response information and uses these opinions as the basis for generating candidate response plans. This effectively compensates for the cognitive blind spots that may exist in a single model, integrating the experience of human experts and local realities into the initial generation stage of the contingency plan. In the example, the supplementary opinions from users A, B, and C, such as "pay attention to upstream reservoir discharge" and "assess the risks of chemical plants," directly enrich the considerations of the contingency plan, which is difficult to fully cover by relying solely on the model. Furthermore, this embodiment inputs the generated candidate response plans back into the plan information processing model for evaluation, and determines the final disaster response plan based on the evaluation responses. This iterative "model-human-model" feedback loop mechanism ensures that the plan can be continuously optimized and improved during the generation process. In the example, the model's evaluation of "general risk assessment measures for chemical plants" and "unclear evacuation resources for elderly people living alone" provides decision-makers with specific directions for improvement, thereby avoiding potential major flaws in the final plan and significantly improving its logic and readability. Compared to the semantic biases and logical inconsistencies that may arise from directly generating plans using existing large language models, this embodiment, by introducing multiple opinions and iterative model evaluation, can generate more accurate, comprehensive, and executable disaster response plans.

[0094] In some embodiments described above, this application proposes inputting disaster event information into a pre-trained contingency plan information processing model to generate disaster event attributes based on the disaster event information, thereby obtaining corresponding disaster attribute response information. However, in practice, directly inputting the raw disaster event information into a large language model may lead to the model's inability to accurately understand the deeper meaning of the event, or the generated information may be insufficiently focused and comprehensive, thus affecting the accuracy and effectiveness of subsequent contingency plan generation. Therefore, this application further proposes a method for generating disaster event attributes based on disaster event information.

[0095] See Figure 3 In one embodiment, the method for generating disaster event attributes based on disaster event information includes, but is not limited to, steps S301 to S303.

[0096] Step S301: Based on a preset similarity threshold, search for disaster event information in the corresponding domain database to obtain the corresponding disaster attribute query information.

[0097] Step S302: Integrate the disaster attribute query information into the preset prompt word template to construct prompt words.

[0098] Step S303: Based on the disaster attribute query information and prompt words, perform contextual semantic association and text content generation to obtain disaster attribute response information.

[0099] By setting a similarity threshold, the accuracy and relevance of search results can be ensured, avoiding the introduction of irrelevant information. Methods based on keyword matching and semantic similarity calculation can be employed. First, keywords are extracted from disaster event information. Then, word vector models (e.g., Word2Vec, GloVe, or BERT embeddings) are used to calculate the semantic similarity between these keywords and entries in the domain database. Entries with similarity scores higher than a preset threshold are selected as disaster attribute query information. Alternatively, knowledge graph-based query methods can be used. Disaster event information is parsed into entities and relationships. Then, path queries or relationship reasoning are performed in a constructed disaster domain knowledge graph to obtain event-related attribute nodes and relationships. These are then filtered based on preset confidence or relevance scores as similarity thresholds.

[0100] Constructing prompts aims to organize structured or semi-structured information retrieved from domain databases in a way that is easily understood and processed by a large language model, forming effective instructions to guide the model in generating high-quality disaster attribute response information. Prompt templates provide a standardized input format, ensuring the model can accurately identify and utilize query information. A prompt template can be preset as: "Please analyze and generate the attributes of the disaster event in detail based on the following disaster event information and related background knowledge: [Disaster Event Information]; Related Background Knowledge: [Disaster Attribute Query Information]." The content within square brackets is filled with actual information at runtime. Prompt templates can also be designed in a more guiding question-and-answer format, such as: "Disaster Event Description: [Disaster Event Information]. Please answer the key attributes of this event based on the following domain knowledge, including but not limited to: event type, location, impact range, potential risks, emergency response level, etc. Domain Knowledge: [Disaster Attribute Query Information]."

[0101] Based on contextual semantic association and text content generation of disaster attribute query information and prompts, this approach leverages the powerful text understanding and generation capabilities of a large language model to transform preprocessed and guided input (disaster attribute query information and prompts) into structured or semi-structured disaster attribute response information. Through contextual semantic association, the model can understand the relationship between the query information and prompts, generating logically coherent and accurate responses. The large language model receives pre-constructed prompts containing disaster event information and disaster attribute query information. Based on its internal knowledge and training data, combined with the prompt instructions, the model performs deep semantic analysis on the input, identifying key entities, events, and relationships, and generating text containing detailed attributes such as event type, occurrence time, location, impact range, number of affected people, economic losses, and secondary disaster risks. During the generation process, an attention mechanism can be introduced, allowing the model to focus more on the parts of the prompts most relevant to the current generated content when generating each word, thereby improving the accuracy and relevance of the generated content. For example, when generating "impact range," the model will focus on data such as geographical location and population density in the disaster attribute query information.

[0102] The proposed solution, when generating disaster event attributes based on disaster event information, firstly, performs a precise information search of the original disaster event information in the relevant domain database based on a preset similarity threshold, thereby obtaining disaster attribute query information highly relevant to the current disaster event. This step effectively filters out key and valuable background knowledge from massive amounts of information, avoiding the problems of model comprehension bias or insufficient information that may result from directly inputting the original information. Subsequently, this filtered disaster attribute query information is cleverly integrated into a preset prompt word template to construct prompt words with clear guidance. This approach allows the large language model to obtain not only the original event information but also refined and organized professional domain knowledge when receiving input, greatly enhancing the depth and breadth of the model's understanding of the event. Finally, the emergency response information processing model performs in-depth contextual semantic association and text content generation based on this rich disaster attribute query information and structured prompt words. In this way, the model can more accurately identify the key elements of the disaster event and generate logically clear, detailed, and highly relevant disaster attribute response information, thus laying a solid foundation for subsequent emergency response plan generation.

[0103] As a specific implementation method, when the system receives disaster event information such as "Continuous heavy rainfall occurred in XX County, XX City, XX Province, causing the water level of the XX River to rise sharply and some low-lying areas to flood," the system first searches for information in a domain database containing geographic information, historical disaster data, meteorological data, population distribution, etc., based on a preset similarity threshold (e.g., 0.8). Through keyword matching and semantic similarity calculation, the system can retrieve disaster disaster records of the region in the past five years, hydrological characteristics of the XX River, the county's flood control plan, and disaster attribute query information such as population density and infrastructure distribution in the affected areas. Subsequently, this query information is integrated into a preset prompt word template. For example, the template can be designed as: "Please analyze and generate the attributes of the disaster event in detail based on the following disaster event description and related domain knowledge, including but not limited to event type, location, scope of impact, potential risks, emergency response level, etc. Event description: [Disaster event information]. Domain knowledge: [Disaster attribute query information]." The actual disaster event information and the retrieved disaster attribute query information are filled into the template to construct a complete prompt word. Finally, the contingency plan information processing model receives the prompt word and, based on its powerful contextual semantic association and text generation capabilities, outputs detailed disaster attribute response information, such as: "Event type: flood disaster; Location: XX County, XX City, XX Province; Affected area: XX River Basin and surrounding low-lying areas; Potential risks: farmland flooding, traffic interruption, landslides, drinking water pollution; Emergency response level: It is recommended to activate Level II emergency response."

[0104] The above technical solution, when generating disaster event attributes, no longer relies solely on raw disaster event information. Instead, it incorporates precise query information from a domain database and effectively guides the large language model by combining carefully constructed prompts. This significantly improves the accuracy, comprehensiveness, and professionalism of the disaster attribute response information, avoiding generalization or inaccurate outputs due to insufficient information or misunderstandings. This refined attribute generation method provides a more solid and reliable foundation of data for subsequent candidate response plan generation, thereby improving the efficiency and effectiveness of the entire intelligent disaster response plan generation method.

[0105] In some other embodiments, this application proposes an intelligent disaster response plan generation method. This method includes acquiring disaster event information, inputting the disaster event information into a pre-trained plan information processing model to generate disaster event attributes based on the disaster event information, and obtaining corresponding disaster attribute response information. The plan information processing model is a model based on a large language model, trained based on sample event information and corresponding sample event attribute information, sample plans, and corresponding sample evaluation response information. Subsequently, it acquires opinion information from multiple users regarding the disaster attribute response information and generates corresponding candidate response plans based on this opinion information. Next, it inputs the candidate response plans into the plan information processing model to evaluate the candidate response plans, obtaining corresponding plan evaluation response information, and finally determines the final disaster response plan based on the plan evaluation response information. However, in the implementation process, how to effectively train a large language model to simultaneously and accurately generate disaster event attributes and evaluate plans, while ensuring the convergence of the training process and the model performance, is a technical problem that needs to be solved. In response, this application further proposes a training method for a contingency plan information processing model, which includes: performing training operations on a large language model to be trained until a preset training termination condition is met, thereby obtaining the contingency plan information processing model.

[0106] Training can be conducted using supervised learning, which involves continuously inputting training data and adjusting model parameters based on the difference between the model's output and the true labels. Training termination conditions can include reaching a preset number of training epochs, the model's performance metrics on the validation set no longer improving over several consecutive epochs, or the loss function converging below a preset threshold. Furthermore, training can be combined with reinforcement learning, for example, using reward mechanisms to guide the model to generate more relevant disaster attribute responses and contingency plan evaluation responses. Training termination conditions can also include reaching a training time limit or the change in model parameters falling below a certain minimum value.

[0107] See Figure 4 In one embodiment, the method for performing the training operation includes, but is not limited to, steps S401 to S406.

[0108] Step S401: Input the sample event information into the current large language model to generate disaster event attributes based on the sample event information and obtain the corresponding sample attribute response information.

[0109] By inputting known sample event information, the current large language model attempts to generate response information that it believes corresponds to the disaster attributes. The sample event information can include structured or unstructured text data such as disaster type, time of occurrence, location, scope of impact, and potential risks.

[0110] After receiving sample event information, the current large language model utilizes its internal language understanding and generation capabilities to output a text describing the disaster attributes, i.e., sample attribute response information. When inputting sample event information, specific prompting engineering techniques can be employed, such as encapsulating the sample event information in a pre-defined template, to guide the current large language model to more accurately generate the required disaster attribute response information.

[0111] Step S402: Determine whether the first information deviation meets the preset deviation conditions.

[0112] If not, proceed to step S403; if yes, proceed to step S404.

[0113] The first information bias is the deviation between the sample event attribute information and the sample attribute response information.

[0114] Determine whether the first information deviation meets the preset deviation conditions. The first information deviation is the deviation between the sample event attribute information and the sample attribute response information. This step is used to quantify the difference between the current model's performance in the disaster event attribute generation task and the real situation.

[0115] By comparing the sample attribute response information generated by the current large language model with the actual sample event attribute information, the first deviation information can be obtained, which can be used to guide subsequent model parameter adjustments.

[0116] First information bias can be measured by calculating text similarity metrics, such as BLEU and ROUGE, to assess the degree of word overlap and semantic matching between model-generated text and real text. First information bias can also be determined through manual evaluation or rule-based matching. For example, a set of key attribute words can be defined, and the model-generated response information can be checked to see if these words are included, and compared with the attribute information of sample events.

[0117] The preset deviation conditions could be that these indicators reach a certain threshold, the loss function falls below a certain value, or the matching rate of key attributes reaches a certain percentage.

[0118] Step S403: Based on the first information bias, adjust the model parameters of the current large language model to obtain the large language model for the next training operation.

[0119] When the current large language model performs poorly on the attribute generation task, it is necessary to optimize the internal parameters of the current large language model based on the first information bias so that it can generate more accurate attribute response information in the next iteration.

[0120] Model parameter tuning can be performed using gradient descent and its variants. This involves calculating the gradient of the loss function with respect to the model parameters and updating the parameters in the opposite direction of the gradient to reduce bias. Tuning strategies can also include learning rate scheduling and regularization techniques to prevent overfitting and accelerate convergence.

[0121] Step S404: Obtain the sample plan and corresponding sample evaluation information for the response information of the sample attributes.

[0122] When the current large language model performs well on the attribute generation task, the training process moves to the next stage: training the proposal evaluation task. At this point, it is necessary to obtain sample proposals related to the generated sample attribute response information and their corresponding real evaluation information.

[0123] Sample contingency plans can be textual descriptions of response strategies, action steps, and resource allocation for specific disaster attributes. Sample evaluation information can include expert ratings, strengths and weaknesses analyses, and feasibility assessments of the plan. This information is typically pre-stored in the training dataset and correlated with the sample event attribute information. In some cases, sample contingency plans and sample evaluation information can also be generated through simulation or semi-automatic methods to expand the training dataset.

[0124] Step S405: Input the sample proposal into the current large language model to evaluate the sample proposal and obtain the corresponding sample evaluation response information.

[0125] Current large language models, upon receiving sample proposals, utilize their learned knowledge and reasoning abilities to generate sample evaluation responses. Taking the sample proposal as input, the current large language model analyzes its completeness, rationality, feasibility, and effectiveness, outputting a text-based evaluation—the sample evaluation response. Similarly, prompting engineering can be employed, such as providing explicit evaluation criteria or dimensions when inputting the sample proposal, guiding the model to generate structured or specifically formatted sample evaluation responses.

[0126] Step S406: Based on the second information bias, adjust the model parameters of the current large language model to obtain the large language model for the next training operation.

[0127] The second information bias is the difference between the sample evaluation information and the sample evaluation response information.

[0128] When the performance of the current large language model on the contingency plan evaluation task deviates from the sample evaluation information, i.e., the second information bias, it is necessary to further optimize the parameters of the current large language model based on this bias so that it can evaluate the contingency plan more accurately.

[0129] The calculation method for the second information bias is similar to that for the first information bias. It can use text similarity metrics or specific evaluation metrics to measure the difference between the model-generated sample evaluation responses and the actual sample evaluation responses. Model parameter adjustment also employs gradient descent and its variants, updating the model parameters based on the gradient of the loss function calculated using the second information bias.

[0130] This application employs a phased, iterative optimization training method to enable a contingency plan information processing model to simultaneously master two core capabilities: disaster event attribute generation and contingency plan evaluation. During training, sample event information is first input into the large language model to be trained, causing it to generate sample attribute response information. Then, the model's performance on the attribute generation task is judged by calculating the first information deviation between this response information and the actual sample event attribute information. If the deviation does not meet preset conditions, the model parameters of the large language model are adjusted based on this deviation to improve its attribute generation capability. When the model reaches a certain level in the attribute generation task, the training process enters the contingency plan evaluation stage. At this point, sample contingency plans related to the generated sample attribute response information and their corresponding actual sample evaluation information are acquired, and the sample contingency plans are input into the current large language model for evaluation, obtaining sample evaluation response information. Next, the second information deviation between this response information and the actual sample evaluation information is calculated, and the model parameters of the large language model are further adjusted based on this deviation to optimize its contingency plan evaluation capability. The entire training operation continues until the preset training termination conditions are met, resulting in a contingency plan information processing model that performs excellently in both disaster event attribute generation and contingency plan evaluation. This phased training strategy enables the model to learn and optimize for different types of tasks, avoiding performance bias that may result from training for a single objective, and ensuring the model's comprehensiveness and accuracy in complex disaster response scenarios.

[0131] The following is a concrete example to illustrate the process. The training of the contingency plan information processing model can follow this procedure. First, prepare an initial large language model with a Transformer architecture, such as a fine-tuned version based on GPT-3 or LLaMA. During the attribute generation phase of training, sample event information containing information such as disaster type, occurrence time, and location is input into the model. For example, inputting "In August 2023, a magnitude 5.0 earthquake occurred in a certain location, with a focal depth of 10 kilometers, causing damage to some houses." The model will generate a sample attribute response, such as "Earthquake disaster, moderate intensity, limited impact range, the main risk is the structural safety of buildings." Next, this generated response is compared with preset real sample event attribute information, for example, using BERT-Score to calculate semantic similarity as the first information bias. If the similarity is lower than a preset threshold, the model's weights and biases are adjusted according to the loss function using the backpropagation algorithm. When the loss in the attribute generation phase converges or reaches a preset number of rounds, the contingency plan evaluation phase begins. At this point, sample contingency plans related to the aforementioned earthquake disaster attributes are obtained, such as "activate emergency response, dispatch rescue teams, assess building damage, and establish temporary resettlement sites," along with corresponding expert evaluation information, such as "the plan is complete, but lacks consideration for secondary disasters." This sample plan is input into the model, which generates sample evaluation responses, such as "the plan is basically feasible, but does not mention the risk of secondary disasters." Similarly, the second information deviation between this response and the expert evaluation information is calculated, and the model parameters are adjusted accordingly. This process is iterated until the model's attribute generation accuracy and contingency plan evaluation F1 score on the validation set both reach the preset standards, ultimately yielding a contingency plan information processing model that can be used for actual disaster response.

[0132] Through the above technical solution, this application provides a systematic and efficient method for training a contingency plan information processing model. This method, through phased and iterative parameter adjustments, enables the large language model to perform deep learning and optimization for the two core tasks of disaster event attribute generation and contingency plan evaluation. This training strategy effectively solves the problem of model performance imbalance that may result from a single training objective, ensuring that the model can accurately identify and generate disaster attributes and comprehensively and objectively evaluate candidate contingency plans when processing complex disaster information. This significantly improves the overall performance and generalization ability of the contingency plan information processing model, laying a solid foundation for the subsequent generation of high-quality disaster response plans, thereby improving the efficiency and accuracy of disaster response.

[0133] In some embodiments, during training operations, the model parameters of the large language model are frozen, and a trainable low-rank matrix is ​​inserted into the corresponding linear layer of the large language model. The model parameters of the current large language model are adjusted, including: adjusting the low-rank matrix to obtain an updated low-rank matrix; and when the training termination condition is met, fitting the model parameters of the current large language model and the updated low-rank matrix to obtain the pre-planned information processing model.

[0134] Freezing the model parameters of a large language model means keeping the core parameters (e.g., weights and biases learned during pre-training) unchanged and not involved in gradient updates. This can be achieved by setting the `requires_grad` attribute of the model parameters to `False`, or by including only the parameters that need updating in the optimizer. Inserting trainable low-rank matrices into the corresponding linear layers of a large language model refers to introducing additional, lower-rank matrices into specific linear transformation layers of the pre-trained large language model (e.g., query, key-value projection layers in attention mechanisms, or fully connected layers in feedforward networks). These low-rank matrices are typically parameterized with smaller dimensions, such as by the product of two smaller matrices, thus providing additional trainable parameters to the model without significantly increasing the total number of model parameters. This insertion can be done by superimposing a low-rank update matrix on the original weight matrix or by introducing it through parallel paths.

[0135] The low-rank matrices are adjusted during training, with only the parameters of these newly inserted low-rank matrices being updated based on the gradients calculated from the loss function. This means the core structure and most parameters of the model remain unchanged, with only a small number of parameters specifically adapted to the task being modified. This adjustment is typically achieved through backpropagation algorithms and optimizers (such as Adam and SGD) to minimize the difference between the predicted output and the true label. The current large language model parameters and the updated low-rank matrices are then fitted together. Once training reaches a preset termination condition, the adjusted low-rank matrices are integrated with the original frozen large language model parameters. This integration can be a physical merging, such as adding the updated low-rank matrix to the original weight matrix to form a new weight matrix for inference; or a logical association, where the core parameters of the large language model and the low-rank matrices work together during inference to complete the forward propagation calculation. This fitting operation allows the trained model to be deployed and used as a whole.

[0136] This application's approach constructs an efficient and stable fine-tuning mechanism by freezing the main parameters of a large language model and cleverly inserting trainable low-rank matrices into specific linear layers within it. During training, only these low-rank matrices are adjusted, while the vast majority of the large language model's parameters remain unchanged. This design allows the model to fully utilize its general language understanding and generation capabilities accumulated during pre-training when adapting to specific disaster response plan generation tasks, avoiding the catastrophic forgetting problem that might occur with full fine-tuning. Simultaneously, the significantly reduced number of parameters required for training significantly lowers the computational resource requirements and training time. When the training process meets the preset termination conditions, the adjusted low-rank matrices are fitted to the original frozen large language model parameters, forming a unified plan information processing model. This fitted model retains the powerful generalization ability of the large language model while acquiring specialized knowledge for disaster response plan generation tasks, enabling more accurate and efficient generation of disaster event attributes and plan evaluation. This efficient parameter fine-tuning strategy, combined with the training operations described above, can effectively improve the performance of the pre-plan information processing model in a specific domain even with limited resources, solving the challenges of high resource consumption and knowledge forgetting faced by directly fine-tuning large models.

[0137] As a specific implementation method, the above-mentioned technical means can be implemented with reference to the Low-Rank Adaptation (LoRA) technique. During training, the query and value projection matrices of the attention mechanism in the pre-trained large language model can be selected as the target linear layer. For each original weight matrix W, it can be decomposed into W + ΔW, where ΔW is a low-rank matrix composed of the product of two smaller matrices B and A (i.e., ΔW = BA). During training, the original weight matrix W is frozen, and only the parameters of matrices B and A are trainable. For example, if the dimension of the original weight matrix W is dxk, then B can be dxr, and A can be rxk, where r is much smaller than d and k (e.g., r = 8 or 16). When adjusting the low-rank matrix, only the parameters of matrices B and A are updated with gradients. When the training termination condition is met, the updated ΔW (i.e., BA) can be added back to the original W, forming a new weight matrix W' = W + BA, which is used in the subsequent inference stage. Alternatively, the updated low-rank matrices B and A can be deployed as independent modules in parallel with the frozen W, and the effect of W + BA can be dynamically calculated during inference.

[0138] By employing the aforementioned technical solutions, when fine-tuning a pre-trained large language model for disaster response plan generation, the computational resources and storage space required for training can be significantly reduced, accelerating the training speed. Since the main parameters of the large language model are frozen, the general language knowledge and capabilities learned during the pre-training phase are effectively preserved, avoiding catastrophic forgetting problems that may occur during task-specific fine-tuning. This ensures the model's generalization ability and stability in generating disaster attribute response information and evaluating candidate response plans. Simultaneously, by introducing a trainable low-rank matrix, the model can efficiently learn disaster response domain-specific knowledge and patterns, resulting in a final plan information processing model that maintains generality while possessing greater professionalism and accuracy, thereby improving the quality and efficiency of disaster response plan generation.

[0139] In some embodiments described above, this application proposes a method for generating disaster response plans using a pre-trained contingency plan information processing model. However, during the training process of the contingency plan information processing model, the method of obtaining sample event attribute information may have limitations. For example, if only a single model or method is relied upon to generate sample event attribute information, the information may be incomplete or lack diversity, thereby affecting the training effect of the contingency plan information processing model and the quality of the final generated plan. Therefore, this application further proposes a method for constructing a mapping relationship between sample event information and sample event attribute information.

[0140] In some embodiments, before inputting disaster event information into a pre-trained contingency plan information processing model, the method further includes: inputting sample event information into several preliminary information processing models in parallel to perform a disaster event attribute generation operation based on the sample event information to obtain sample event attribute information; and performing information mapping processing on the sample event information and the sample event attribute information to make them correspond to each other.

[0141] Parallel input of sample event information into several preliminary information processing models refers to sending the same sample event information to multiple different processing units for processing simultaneously. This can be achieved through distributed computing frameworks, such as using Apache Spark or Hadoop to distribute sample event information to multiple computing nodes, with each node running a preliminary information processing model; or, on a single high-performance server, using multi-threading or multi-processing technology to simultaneously start multiple instances of preliminary information processing models to process the sample event information separately.

[0142] The preliminary information processing model primarily consists of large language models of different types. This model is used to generate disaster event attributes based on sample event information, thus obtaining the sample event attribute information. It serves as the data preparation step before training the contingency plan information processing model. These models can be large language models with different architectures or parameter configurations, such as BERT, the GPT series, and T5, each with its own focus on natural language understanding and generation tasks. They can also be hybrid models combining domain-specific knowledge graphs or rule engines to enhance the ability to identify and extract disaster event attributes. Different types of large language models have different pre-training data, model structures, or fine-tuning strategies, therefore they may produce different outputs or focus on extracting different dimensions of information when processing the same input. For example, they can be models based on the Transformer architecture but with different numbers of layers, attention heads, or hidden layer dimensions; they can also be models specifically pre-trained or fine-tuned for different languages ​​or specific domains (such as medical, legal, and disaster management).

[0143] Information mapping of sample event information and sample event attribute information aims to ensure a clear and accurate correspondence between sample event information and the sample event attribute information generated by the preliminary information processing model. This ensures that the input can be correctly associated with the expected output during subsequent training. This can be achieved by establishing a database or data structure to associate and store the unique identifier of each sample event information with the sample event attribute information generated by all corresponding preliminary information processing models; or by establishing this one-to-many mapping relationship at the data storage level through metadata tags or file naming conventions, ensuring data consistency and traceability. Information mapping ensures that each sample event information can accurately find its corresponding sample event attribute information generated by multiple preliminary models, thus providing rich and accurate supervision signals for the training of the contingency plan information processing model. This can be achieved through data cleaning, deduplication, and merging operations to integrate attribute information from different preliminary models into the same sample event record; or by manual review or semi-automated tools to verify the accuracy of the mapping relationship, especially when there are differences in the outputs of the preliminary models.

[0144] This application, by inputting sample event information in parallel into several preliminary information processing models and utilizing these models, primarily based on different types of large language models, to perform disaster event attribute generation, enables in-depth analysis and attribute extraction of the same sample event information from multiple dimensions and perspectives. Each preliminary information processing model, due to its different type, architecture, or training focus, may capture different key elements or provide different interpretations of the sample event information. For example, one model may excel at identifying disaster type and timing, while another may focus more on assessing potential impacts and resource requirements. This combination of parallel processing and diverse models yields more comprehensive, accurate, and robust sample event attribute information. Subsequently, information mapping processing is performed on the sample event information and these diverse sample event attribute information to ensure that each original event can be accurately associated with all its related attribute information generated by different models. This mechanism provides high-quality, multi-dimensional supervision signals for the training of subsequent emergency response information processing models, significantly improving the models' ability to understand complex disaster scenarios, generate accurate disaster attributes, and evaluate response plans, thereby effectively solving the problem of information bias that may result from a single model.

[0145] The following is a concrete example. Suppose we have a sample event description: "In August 2023, a mudslide occurred in a mountainous area, causing road closures and trapping several tourists who urgently needed rescue." To generate comprehensive sample event attribute information, this text can be input into three preliminary information processing models in parallel. For example, the first preliminary information processing model could be a disaster event element extraction model based on GPT-4 fine-tuning, focusing on identifying the disaster type, time, location, and direct consequences; the second preliminary information processing model could be a risk assessment model based on Llama 2 fine-tuning, focusing on identifying potential secondary disasters, characteristics of affected populations, and emergency needs; and the third preliminary information processing model could be a resource scheduling model based on Falcon fine-tuning, focusing on identifying the required rescue forces, types of supplies, and traffic conditions. After parallel input, the first model might output "Disaster type: mudslide; Time: August 2023; Location: a mountainous area; Direct consequences: road closures, tourists trapped." The second model might output "Potential risks: secondary landslides, food shortages; Affected population: tourists; Emergency needs: life-saving, medical assistance." The third model might output "Required resources: professional rescue teams, medical teams; Required supplies: food, water, first-aid medicines; Traffic conditions: road closures." Subsequently, the original sample event information is associated and mapped with all the attribute information generated by these three models, for example, by storing it in a structured data record to ensure that each attribute accurately belongs to the sample event, forming a rich and multi-dimensional training sample.

[0146] Through the aforementioned technical solutions, this application obtains a more comprehensive, accurate, and diverse training dataset. This high-quality training data significantly improves the robustness and accuracy of the emergency response information processing model in understanding disaster events, generating disaster attributes, and subsequently evaluating response plans. By integrating the advantages of different types of large language models, the bias or information omissions that may exist in a single model are avoided, thereby making the final generated disaster response plan more targeted, scientific, and practical, effectively improving the efficiency and effectiveness of disaster response.

[0147] In some of the embodiments described above in this application, a method for generating corresponding candidate response plans is proposed based on user feedback information regarding disaster attributes. However, in practice, user opinions may be diverse or even conflicting. Directly generating plans based on these original opinions may result in low-quality plans, difficulty in effectively integrating the viewpoints of different users, and failure to fully consider the dynamic changes and potential tendencies of user opinions, thereby affecting the practicality and acceptability of the final plan. Therefore, this application further proposes a method for generating candidate response plans.

[0148] See Figure 5 In one embodiment, the method for generating candidate response plans includes, but is not limited to, steps S501 to S503.

[0149] Step S501: Cluster the opinion information to obtain several opinion clusters.

[0150] Clustering of various opinions aims to summarize and organize a large amount of potentially scattered user opinions, identify opinion groups with similar themes, viewpoints, or tendencies, and thus lay the foundation for subsequent core opinion extraction and user behavior analysis.

[0151] This clustering operation can employ the K-means clustering algorithm, which iteratively assigns data points to the nearest cluster centers and updates the cluster centers until convergence. When applied to textual opinion information, it is typically necessary to first convert the textual opinions into vector representations and then cluster these vectors. Alternatively, hierarchical clustering algorithms can be used, which construct a clustering tree and progressively merge or split clusters until a stopping condition is met.

[0152] Step S502: Determine the core opinion information in the opinion cluster, and calculate the relative swing probability and autonomous learning probability of the user corresponding to other opinion information.

[0153] Core opinions can be determined by calculating the centrality of opinions within each cluster, for example, by selecting the opinion closest to the centroid of the cluster. Alternatively, text summarization or keyword extraction can be performed on all opinions within each cluster, and then the opinion that best represents the theme of that cluster can be selected as the core opinion.

[0154] The relative swing probability is the probability that a user will change their opinion, while the self-learning probability is the probability that a user will support the core opinion. These two probabilities reflect the dynamics and malleability of user opinions, and are crucial for understanding user behavior, predicting opinion evolution, and generating more adaptive contingency plans. The relative swing probability can be calculated based on factors such as the user's historical opinion change records, user ratings or feedback on different opinions, and user activity in discussions. The self-learning probability can be calculated based on the user's level of agreement with the core opinion, the user's social relationship with the core opinion's publisher, and the user's mastery of relevant domain knowledge. Furthermore, by constructing user behavior models and utilizing Bayesian networks or Markov chain models, combined with users' historical interaction data, the likelihood of users changing their opinions or accepting new opinions in specific situations can be predicted.

[0155] Step S503: Input the opinion information, relative swing probability, and autonomous learning probability into the preset contingency plan generation model, and perform contingency plan generation operation based on the opinion information, relative swing probability, and autonomous learning probability to obtain candidate response contingency plans.

[0156] The contingency plan generation model is an intelligent model that comprehensively considers the content of user opinions, the dynamic trend of user opinions, and the user's acceptance of core viewpoints. This model can be a deep learning-based sequence generation model, such as the Transformer or GPT series models, receiving structured opinion information (text), relative swing probabilities (numerical values), and self-learning probabilities (numerical values) as input. Through its encoder-decoder architecture, it generates contingency plan text that conforms to disaster response logic and user intent. Alternatively, it can be a hybrid model based on rules and expert knowledge. A rule engine filters, combines, or modifies preset contingency plan templates based on opinion information and probability values, and then an expert system evaluates and optimizes the generated preliminary contingency plan.

[0157] This application's solution overcomes the problem of low-quality contingency plans that may result from directly processing raw opinions by refining the input of user opinions. Specifically, it first clusters the various opinions, grouping those with similar themes or tendencies into several opinion clusters, thereby effectively organizing and understanding diverse user viewpoints. Based on this, the most representative core opinions are identified from each opinion cluster, which helps to extract key propositions and avoid information redundancy. Simultaneously, to capture the dynamism and malleability of user opinions, the relative swing probability and self-learning probability of users corresponding to other opinions are calculated, quantifying the likelihood of users changing their opinions or supporting core opinions. This probabilistic information provides deeper insights into user behavior for contingency plan generation. Finally, the opinion information, after clustering, core opinion extraction, and dynamic probability quantification, along with the relative swing probability and self-learning probability, is input into a pre-set contingency plan generation model. This model comprehensively considers the content of opinions, the stability of user opinions, and the acceptance of consensus, thereby performing a more intelligent and comprehensive contingency plan generation operation, resulting in candidate response plans that effectively integrate collective wisdom and adapt to dynamic user changes. This approach ensures that the generated candidate proposals are not only based on current opinions, but also take into account the potential evolution of those opinions and user acceptance, significantly improving the practicality and adoption rate of the proposals.

[0158] As a specific implementation method, after receiving feedback from multiple users regarding disaster attributes, this feedback can be processed first. For example, when receiving feedback about a forest fire in a certain area, opinions such as "suggesting the deployment of helicopters for firefighting" and "establishing firebreaks as soon as possible" can be vectorized into text and clustered using the K-means algorithm to obtain opinion clusters such as "airborne firefighting strategy" and "ground isolation strategy." Within the "airborne firefighting strategy" cluster, "deploying large firefighting helicopters for aerial water spraying" can be identified as the core feedback, as it semantically best represents the main viewpoint of this cluster. Subsequently, for other users, their relative swing probability and self-learning probability can be calculated based on their historical behavioral data. For example, if a user has frequently changed their opinions on firefighting plans in the past, their relative swing probability might be 0.7; if the user shows high agreement with the expert-proposed airborne firefighting plan, their self-learning probability might be 0.8. Finally, this processed feedback, the calculated relative swing probability, and the self-learning probability are input into a pre-defined contingency plan generation model. The model can be a finely tuned version of a large language model. It can comprehensively analyze these inputs and generate a candidate response plan that includes "prioritizing the deployment of large firefighting helicopters with nighttime operation capabilities and combining this with the establishment of ground firebreaks to form a three-dimensional firefighting system."

[0159] Through the aforementioned technical solution, this application effectively addresses the issues of fragmented and conflicting opinions, as well as the difficulty in capturing dynamic changes in user opinions, that may arise when generating contingency plans directly based on original user feedback. By clustering feedback information and extracting core opinions, the core viewpoints of different users can be clearly identified and integrated, avoiding disorganized contingency plan content. Simultaneously, the introduction of relative swing probability and self-learning probability ensures that the contingency plan generation process fully considers the stability of user opinions and the acceptance of consensus, thereby generating more adaptive and forward-looking candidate response plans. This not only improves the quality and practicality of candidate plans but also makes the subsequent plan evaluation and final plan determination processes more efficient and accurate, ultimately enhancing the overall efficiency of disaster response plan generation and decision support capabilities.

[0160] In some embodiments, before clustering the opinion information, the method further includes: performing Z-number transformation on the opinion information to obtain opinion information in Z-number form; and performing normal cloud transformation on the Z-number form opinion information to obtain opinion information in normal cloud form.

[0161] Z-number is a mathematical tool for handling uncertainty and ambiguity, combining a numerical constraint with a reliability measure of that constraint. By converting raw opinion information into Z-number form, the inherent ambiguity and uncertainty in opinions can be effectively captured, and their reliability quantified. For example, a user's description of the severity of a disaster (e.g., "very severe") can be converted into a fuzzy number, along with a reliability measure indicating the user's confidence in this judgment (e.g., "relatively certain"). Implementation methods can include: one approach is to leverage expert knowledge or predefined rules to map opinions in natural language form to predefined fuzzy sets and reliability levels, then mathematically representing them as Z-numbers; another approach is to combine natural language processing techniques to extract sentiment and confidence levels from textual opinions and convert them into two components of a Z-number.

[0162] The normal cloud model is a mathematical model capable of transforming uncertainty between qualitative concepts and quantitative values. It describes the fuzziness and randomness of a concept using three parameters: expectation (Ex), entropy (En), and hyperentropy (He). Converting opinion information in Z-number form into normal cloud form can further transform opinions with fuzzy and reliability information into a more structured and quantifiable representation, thus better reflecting the overall uncertainty distribution of opinions. Implementation methods can include: one approach is to derive the Ex, En, and He parameters of the normal cloud model based on the fuzzy and reliability components of the Z-number using specific algorithms. The center value of the fuzzy number can be used as Ex, its dispersion can affect En, and the reliability measure can affect He. Another approach is to utilize statistical methods or machine learning models to learn the mapping relationship between the Z-number and the normal cloud parameters to adapt to the transformation needs of different types of opinion data.

[0163] Before clustering user-input opinions, this application first performs a Z-number transformation, converting the raw, potentially subjective and uncertain opinions into a Z-number form that includes fuzzy constraints and reliability metrics. This transformation allows the system to quantify and preserve the inherent imprecision in the opinions, rather than simply treating them as precise data. Subsequently, these Z-number opinions are further transformed using a normal cloud model. The normal cloud model maps the fuzziness and reliability expressed by the Z-number to a cloud droplet distribution with expectation, entropy, and hyperentropy parameters, thus comprehensively characterizing the fuzziness and randomness of the opinions at a quantitative level. Through this two-stage transformation, the raw, potentially ambiguous opinion information is systematically transformed into a standardized, computable form that reflects both its semantic content and accurately expresses its degree of uncertainty. This preprocessing mechanism ensures that subsequent opinion clustering operations (such as identifying core opinions, calculating relative swing probabilities, and self-learning probabilities) can be performed on a more robust and information-rich representation, significantly improving the accuracy and reliability of the clustering results, and thus laying a solid foundation for generating high-quality candidate response plans.

[0164] For example, suppose a user expresses the opinion "The loss may be significant, but I'm not entirely sure" regarding the potential impact of a disaster. Before clustering the opinion information, a Z-number transformation is first performed on this opinion. "The loss may be significant" can be represented as a fuzzy number A, for example, its membership function peaks between 80 and 95 for a loss level of 0 to 100; while "I'm not entirely sure" can be represented as a reliability metric B, for example, its membership function peaks between 0.5 and 0.7 in a reliability range of 0 to 1. Subsequently, this Z-number form of opinion information is transformed into a normal cloud. Based on the center value and distribution range of the fuzzy number A, the expected value Ex and entropy En of the normal cloud can be calculated; simultaneously, combined with the reliability metric B, the hyperentropy He of the normal cloud can be further determined. Finally, the opinion is transformed into a normal cloud described by the parameters Ex, En, and He, which not only reflects the core judgment of "significant loss" but also quantifies its fuzziness and the user's uncertainty. This normal cloud-like opinion information is then fed into the contingency plan generation model for clustering, allowing the clustering algorithm to fully consider the inherent uncertainty of the opinions, rather than treating them as precise single values.

[0165] By introducing Z-number transformation and normal cloud transformation before clustering opinion information, this application effectively addresses the ambiguity, uncertainty, and subjectivity prevalent in user opinions. This preprocessing mechanism transforms raw, potentially vague opinion information into a structured, quantifiable, and comprehensively reflective normal cloud form that reflects its uncertainty characteristics. This enables the subsequent opinion clustering process to more accurately identify core opinions and more precisely calculate users' relative swing probabilities and self-learning probabilities. Therefore, the reliability and effectiveness of candidate response plans generated based on this refined opinion information are significantly improved, providing a more solid and accurate data foundation for ultimately determining high-quality disaster response plans.

[0166] See Figure 6 This application also provides a disaster emergency response plan generation device, which can implement the above-mentioned disaster emergency response plan generation method. The device includes:

[0167] The first module, 601, is used to obtain disaster event information;

[0168] The second module 602 is used to input disaster event information into a pre-trained contingency plan information processing model to generate disaster event attributes based on the disaster event information and obtain corresponding disaster attribute response information. The contingency plan information processing model is a model based on a large language model, which is trained based on sample event information and corresponding sample event attribute information, sample contingency plans and corresponding sample evaluation response information.

[0169] The third module 603 is used to obtain the opinions input by multiple users regarding the response information on disaster attributes;

[0170] Module 4, 604, is used to generate corresponding candidate response plans based on the feedback information.

[0171] The fifth module 605 is used to input candidate response plans into the plan information processing model to evaluate the candidate response plans and obtain the corresponding plan evaluation response information.

[0172] Module 606 is used to determine the final disaster response plan based on the response information from the plan evaluation.

[0173] The specific implementation method of the disaster emergency plan generation device is basically the same as the specific implementation method of the disaster emergency plan generation method described above, and will not be repeated here.

[0174] Figure 7 This is a schematic diagram of the hardware structure of the electronic device provided in the embodiments of this application.

[0175] The following reference Figure 7 To describe an electronic device 700 according to such an embodiment of the present disclosure. Figure 7 The electronic device 700 shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments disclosed herein.

[0176] like Figure 7 As shown, the electronic device 700 is presented in the form of a general-purpose computing device. The components of the electronic device 700 may include, but are not limited to: at least one processing unit 710, at least one storage unit 720, a bus 730 connecting different system components (including storage unit 720 and processing unit 710), a display unit 740, etc.

[0177] The storage unit stores program code, which can be executed by the processing unit 710, causing the processing unit 710 to perform the steps described in the above-described disaster emergency plan generation method section of this specification according to various exemplary embodiments of this disclosure.

[0178] Storage unit 720 may include a readable medium in the form of a volatile storage unit, such as random access memory (RAM) 7201 and / or cache memory 7202, and may further include a read-only memory (ROM) 7203.

[0179] The storage unit 720 may also include a program / utility 7204 having a set (at least one) program module 7205, such program module 7205 including but not limited to: an operating system, one or more application programs, other program modules and program data, each or some combination of these examples may include an implementation of a network environment.

[0180] Bus 730 can represent one or more of several types of bus structures, including a memory cell bus or memory cell controller, a peripheral bus, a graphics acceleration port, a processing unit, or a local bus using any of the various bus structures.

[0181] Electronic device 700 can also communicate with one or more external devices 700' (e.g., keyboard, pointing device, Bluetooth device, etc.), and with one or more devices that enable a user to interact with electronic device 700, and / or with any device that enables electronic device 700 to communicate with one or more other computing devices (e.g., router, modem, etc.). This communication can be performed via input / output (I / O) interface 750. Furthermore, electronic device 700 can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN), and / or public networks, such as the Internet) via network adapter 760. Network adapter 760 can communicate with other modules of electronic device 700 via bus 730. It should be understood that, although not shown in the figures, other hardware and / or software modules can be used in conjunction with electronic device 700, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0182] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0183] The disaster emergency response plan generation method, apparatus, equipment, and medium provided in this application introduce a pre-trained plan information processing model. This model generates corresponding disaster attribute response information based on disaster event information, then obtains opinions from multiple users regarding the disaster attribute response information, and generates corresponding candidate response plans based on these opinions. The plan information processing model evaluates the generated candidate response plans, and finally determines the disaster response plan based on the plan evaluation response information. Because this plan information processing model is primarily based on a general language model, but its training process specifically incorporates sample event information, corresponding sample event attribute information, sample plans, and corresponding sample evaluation response information, this plan information processing model can overcome the shortcomings of general language models in disaster emergency knowledge processing, possessing stronger disaster situation understanding and plan evaluation capabilities. Generating corresponding candidate response plans based on opinions input regarding disaster attribute response information effectively compensates for the cognitive blind spots that may exist in a single model, integrates user experience into the initial plan generation stage, and then inputs the generated candidate response plans back into the plan information processing model for plan evaluation, determining the final disaster response plan based on the evaluation response information. This iterative "model-human-model" feedback loop mechanism ensures that the plan can be continuously optimized and improved during the generation process, providing decision-makers with specific directions for improvement, thereby avoiding major defects that may exist in the final plan and significantly improving the logic and readability of the disaster emergency plan.

[0184] From the above description of the embodiments, those skilled in the art will readily understand that the exemplary embodiments described herein can be implemented by software or by combining software with necessary hardware. Therefore, the technical solutions according to the embodiments of this disclosure can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) or on a network, including several instructions to cause a computing device (such as a personal computer, server, or network device, etc.) to execute the methods described above according to the embodiments of this disclosure.

[0185] The program product may employ any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples (a non-exhaustive list) of readable storage media include: electrical connections having one or more wires, portable disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0186] Computer-readable storage media may include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable storage medium may also be any readable medium other than a readable storage medium that can transmit, propagate, or transfer a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the readable storage medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof.

[0187] Those skilled in the art will understand that the above modules can be distributed in the device as described in the embodiments, or they can be modified accordingly and located in one or more devices that are unique to this embodiment. The modules in the above embodiments can be combined into one module, or they can be further divided into multiple sub-modules.

[0188] Exemplary embodiments of this disclosure have been specifically shown and described above. It should be understood that this disclosure is not limited to the detailed structures, arrangements, or implementations described herein; rather, this disclosure is intended to cover various modifications and equivalent arrangements contained within the spirit and scope of the appended claims.

Claims

1. A method for intelligent generation of a disaster response plan, characterized by, include: Obtain disaster event information; The disaster event information is input into a pre-trained contingency plan information processing model to generate disaster event attributes based on the disaster event information, and obtain corresponding disaster attribute response information; the contingency plan information processing model is a model based on a large language model, which is trained based on sample event information and corresponding sample event attribute information and sample contingency plans and corresponding sample evaluation response information. Obtain feedback from multiple users regarding their responses to the disaster attributes; Based on the aforementioned feedback, corresponding candidate response plans are generated; The candidate response plans are input into the plan information processing model to evaluate the candidate response plans and obtain the corresponding plan evaluation response information. Based on the response information to the aforementioned contingency plan evaluation, the final disaster response plan will be determined. The generation of disaster event attributes based on the disaster event information includes: Based on a preset similarity threshold, information search is performed on the disaster event information in the corresponding domain database to obtain the corresponding disaster attribute query information; The disaster attribute query information is integrated into a preset prompt word template to construct prompt words; Based on the disaster attribute query information and the prompt words, contextual semantic association and text content generation are performed to obtain the disaster attribute response information; The step of generating corresponding candidate response plans based on the opinion information includes: The opinions described are clustered to obtain several opinion clusters; The core opinion information in the opinion cluster is determined, and the relative swing probability and self-learning probability of the user corresponding to other opinion information are calculated; the relative swing probability is the probability that the user changes their opinion information, and the self-learning probability is the probability that the user supports the core opinion information. The opinion information, the relative swing probability, and the autonomous learning probability are input into a preset contingency plan generation model to perform a contingency plan generation operation based on the opinion information, the relative swing probability, and the autonomous learning probability, thereby obtaining the candidate response contingency plan.

2. The method according to claim 1, wherein The training method for the contingency plan information processing model includes: The large language model to be trained is subjected to training operations until the preset training termination conditions are met, thereby obtaining the pre-planned information processing model; the training operations include: The sample event information is input into the current large language model to generate disaster event attributes based on the sample event information, and obtain the corresponding sample attribute response information; Determine whether the first information deviation meets the preset deviation condition; the first information deviation is the deviation between the sample event attribute information and the sample attribute response information; If not, based on the first information deviation, the model parameters of the current large language model are adjusted to obtain the large language model for the next training operation; If so, obtain the sample plan and corresponding sample evaluation information for the response information of the sample attributes; The sample proposal is input into the current large language model to evaluate the sample proposal and obtain the corresponding sample evaluation response information. Based on the second information bias, the model parameters of the current large language model are adjusted to obtain the large language model for the next training operation; the second information bias is the deviation between the sample evaluation information and the sample evaluation response information.

3. The disaster countermeasure plan intelligent generation method according to claim 2, characterized by, During the training operation, the model parameters of the large language model are frozen, and a trainable low-rank matrix is ​​inserted into the corresponding linear layer of the large language model. Adjusting the model parameters of the current large language model includes: The low-rank matrix is ​​adjusted to obtain the updated low-rank matrix; When the training termination condition is met, the model parameters of the current large language model and the updated low-rank matrix are fitted to obtain the pre-plan information processing model.

4. The intelligent generation method for disaster response plans according to claim 1, characterized in that, Before inputting the disaster event information into the pre-trained contingency plan information processing model, the method further includes: The sample event information is input in parallel into several preliminary information processing models to perform disaster event attribute generation operations based on the sample event information, thereby obtaining the sample event attribute information; the preliminary information processing models are mainly based on different types of large language models; Information mapping processing is performed on the sample event information and the sample event attribute information to make them correspond to each other.

5. The intelligent generation method for disaster response plans according to claim 1, characterized in that, Before clustering the aforementioned opinion information, the method further includes: The opinion information is Z-number transformed to obtain opinion information in Z-number format; The opinion information in Z-number form is subjected to normal cloud transformation to obtain opinion information in normal cloud form.

6. A disaster response plan intelligent generation device, characterized in that, include: The first module is used to obtain disaster event information; The second module is used to input the disaster event information into a pre-trained contingency plan information processing model to generate disaster event attributes based on the disaster event information and obtain corresponding disaster attribute response information; the contingency plan information processing model is a model based on a large language model, which is trained based on sample event information and corresponding sample event attribute information and sample contingency plans and corresponding sample evaluation response information. The third module is used to obtain the opinions input by multiple users regarding the disaster attribute response information; The fourth module is used to generate corresponding candidate response plans based on the opinion information; The fifth module is used to input the candidate response plan into the plan information processing model to evaluate the candidate response plan and obtain the corresponding plan evaluation response information. The sixth module is used to determine the final disaster response plan based on the evaluation response information of the plan; The generation of disaster event attributes based on the disaster event information includes: Based on a preset similarity threshold, information search is performed on the disaster event information in the corresponding domain database to obtain the corresponding disaster attribute query information; The disaster attribute query information is integrated into a preset prompt word template to construct prompt words; Based on the disaster attribute query information and the prompt words, contextual semantic association and text content generation are performed to obtain the disaster attribute response information; The step of generating corresponding candidate response plans based on the opinion information includes: The opinions described are clustered to obtain several opinion clusters; The core opinion information in the opinion cluster is determined, and the relative swing probability and self-learning probability of the user corresponding to other opinion information are calculated; the relative swing probability is the probability that the user changes their opinion information, and the self-learning probability is the probability that the user supports the core opinion information. The opinion information, the relative swing probability, and the autonomous learning probability are input into a preset contingency plan generation model to perform a contingency plan generation operation based on the opinion information, the relative swing probability, and the autonomous learning probability, thereby obtaining the candidate response contingency plan.

7. An electronic device, characterized in that, The electronic device includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the intelligent generation method for disaster response plans according to any one of claims 1 to 5.

8. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the intelligent generation method for disaster response plans as described in any one of claims 1 to 5.

Images