An intelligent generation method for a super-standard flood control dynamic plan based on a large model

Abstract

This invention discloses an intelligent method for generating dynamic flood control plans for floods exceeding standard levels based on a large-scale model. The method includes: basic data collection and preprocessing, knowledge base construction, establishment of a professional model invocation workflow, dynamic simulation and deduction of multiple scenarios, plan generation and intelligent optimization, and plan output and dynamic updating. The method described in this invention, by constructing a water resources RAG knowledge base, integrating standardized professional models, and designing an intelligent workflow engine, achieves full-process automation and intelligence from monitoring floods exceeding standard levels to generating dynamic plans, thereby improving the scientific rigor, timeliness, and foresight of flood control decisions. The rationality of this invention is verified using the control of floods exceeding standard levels in the Huai River Basin as an example.

Description

Technical Field

[0001] This invention belongs to the field of smart water conservancy and flood control and disaster reduction technology, and specifically relates to a method for intelligent generation of dynamic flood control plans for super-standard floods based on a large model. Background Technology

[0002] Super-standard floods refer to catastrophic floods that exceed the design standards of water conservancy projects or the flood control capacity of a river basin. They are extremely destructive, and traditional flood control decision-making models relying on fixed rules and static plans are no longer sufficient to cope with the challenges posed by such sudden disasters. Current flood control plan development heavily relies on expert experience, revealing several significant shortcomings in practical application: First, emergency response is delayed. Faced with rapidly evolving super-standard floods, manual analysis of water and rainfall conditions, development of multiple scheduling schemes, and simulation verification are time-consuming and cannot meet the timeliness requirements of flood control emergency response. Second, flood-related knowledge is not fully utilized. Vast amounts of historical scheduling cases, engineering scheduling procedures, and real-time water and rainfall monitoring data are scattered across different systems and documents, making comprehensive and rapid retrieval and analysis difficult during the decision-making process. Third, the ability to simulate and extrapolate schemes is limited. Manual analysis of multiple flood control scheduling schemes in a short period is insufficient, making it difficult to accurately calculate the multiple impacts of upstream reservoir scheduling on downstream river channels, flood diversion areas, and urban flood control. This results in insufficient scheme comparison and selection of the optimal decision.

[0003] In recent years, large language models have demonstrated excellent capabilities in natural language understanding and complex logical reasoning, and the application of Retrieval Augmentation (RAG) technology has effectively expanded the professional knowledge boundaries of these models. Meanwhile, the simulation accuracy of specialized models in the water conservancy field (such as one-dimensional / two-dimensional hydrodynamic models) has been continuously improving. However, a mature technical solution has yet to be developed to deeply integrate the intelligent reasoning and knowledge fusion capabilities of large models with the precise computational capabilities of specialized water conservancy models. This makes it difficult to construct an intelligent system capable of automatically sensing and understanding flood occurrence and development scenarios, rapidly retrieving flood-related professional knowledge, intelligently scheduling specialized models to conduct flood simulations, and ultimately generating high-quality dynamic flood control plans. This has become a pressing technical challenge to be solved in the fields of smart water conservancy and flood control and disaster reduction. Summary of the Invention

[0004] To overcome the problems of existing technologies, this invention proposes an intelligent generation method for dynamic flood control plans for floods exceeding standard levels, based on a large model. The method utilizes a water resources RAG knowledge base, integrates standardized professional models, and designs an intelligent workflow engine to achieve full automation and intelligence from monitoring floods exceeding standard levels to generating dynamic plans, thereby improving the scientific rigor, timeliness, and foresight of flood control decisions. The rationale for this invention is verified using the control of floods exceeding standard levels in the Huai River Basin as an example.

[0005] The objective of this invention is achieved as follows:

[0006] This invention provides an intelligent generation method for dynamic flood control plans for floods exceeding standard levels based on a large model, comprising the following steps:

[0007] Step 1: Basic Data Collection and Preprocessing

[0008] Collect basic data related to flood control and prevention in the basin, and perform preprocessing operations on the collected data, including data cleaning and format conversion.

[0009] Step 2, Knowledge Base Construction

[0010] The preprocessed data undergoes text segmentation and vectorization to construct a RAG knowledge base in the water resources field. Text segmentation, which divides long texts into smaller blocks to facilitate vector generation and retrieval, is particularly necessary for water resources data, as the text is typically long and information-rich, and direct processing would waste resources and reduce efficiency. For accuracy, the text segmentation adopts a semantic-based segmentation strategy, adding markers before and after the segmented text content. The vectorization process uses a pre-trained professional embedding model fine-tuned with water resources corpus (such as a word vector model optimized for water resources) to enhance the understanding of professional terminology, converting the segmented text into high-dimensional vectors. Cosine similarity is used to represent the semantic relationships between texts, and the generated vectors are stored in a vector database and indexed, thereby improving retrieval speed.

[0011] Step 3: Professional model call workflow setup

[0012] Establish a professional model invocation workflow to enable the large model to schedule and perform analysis calculations on the water conservancy professional model. Specific steps include:

[0013] S31, Professional Model Modeling: Develop water conservancy professional models, including water project scheduling models, hydrological forecasting models, flood evolution models, and loss assessment models;

[0014] S32, User Intent Recognition: User intent recognition is achieved through water conservancy corpus collection and scene segmentation, data annotation and enhancement;

[0015] S33, Interface Development: Integrate the aforementioned water conservancy professional model, complete the interface specification formulation, input and output parameter definition, and interface development and adaptation;

[0016] S34, Workflow Engine Development: Based on scheduling logic design and automated task arrangement, develop a workflow engine to realize the automated scheduling of large models on water conservancy professional models.

[0017] Step 4: Dynamic simulation and deduction of multiple scenarios

[0018] Based on the professional model built in step 3, the workflow is invoked to carry out dynamic simulation and deduction of multiple schemes for flood control exceeding the standard through the large model;

[0019] Step 5: Contingency Plan Generation and Intelligent Optimization

[0020] Based on the multi-scheme dynamic simulation results obtained in step 4, and combined with the RAG knowledge base described in step 2, the generation and intelligent optimization of the dynamic flood control plan for floods exceeding the standard are completed.

[0021] Step 6: Contingency Plan Output and Dynamic Updates

[0022] The optimized dynamic flood control plan is output and dynamically updated based on real-time monitoring data of the basin, forming a dynamic closed loop for flood control decision-making.

[0023] Furthermore, in step 1, the basic data includes water conservancy project data within the basin, flood control plan-related documents and materials, and water and rainfall data;

[0024] The water conservancy project data includes reservoir characteristic water levels and capacity, reservoir scheduling rules, reservoir water level-capacity-area curves, reservoir water level-flow curves, sluice gate parameters, sluice gate opening-discharge curves, sluice gate scheduling rules, flood storage and detention area water level-capacity-area curves, and flood storage and detention area activation rules.

[0025] The relevant documents and materials for the flood control plan include basin flood control plan, flood defense plan for floods exceeding standard levels, river system flood control and emergency rescue plan, urban water and drought disaster prevention plan, flood diversion gate operation plan, and flood control and emergency rescue plan for dangerous sections of the river.

[0026] The hydrological and rainfall data includes rainfall forecasts, rainfall data observed at rain gauge stations, flow and water level observation data from hydrological stations, and reservoir discharge and inflow observation data.

[0027] Furthermore, in step 1, data cleaning aims to remove noise and correct errors (such as rain gauge malfunctions or data entry errors) to improve data quality. The data cleaning operations include: correcting typos and grammatical errors in text-based data and removing redundant content (such as repeated paragraphs); filling missing values ​​in Excel spreadsheet data using interpolation or statistics, correcting, deleting, or marking outliers, and deleting and merging duplicate data through key field comparisons; to ensure effectiveness, a quality assessment mechanism is established after data cleaning, and manual spot checks are conducted for verification.

[0028] The purpose of format conversion is to unify diverse formats for easier subsequent processing. The format conversion operations include: converting text files (such as PDF and Word documents) to Markdown format and verifying the integrity of charts; converting Excel spreadsheet data to CSV or JSON format; and using tools (such as the pandas library) for batch conversion. During the conversion process, it is crucial to pay attention to the encoding method (such as UTF-8 or GBK) to avoid garbled characters.

[0029] Furthermore, in step 2, the fine-tuning process of the pre-trained professional embedding model employs a triplet loss function. Through iterative training with a large amount of water conservancy text corpus, the model parameters are optimized to enable it to more accurately capture the semantic relationships between water conservancy professional terms.

[0030]

[0031] In the formula: The embedding vector of the anchor sample; The embedding vector of a positive sample belonging to the same category as the anchor point; This is the embedding vector of a negative sample that belongs to a different category than the anchor point; Distance metrics typically employ Euclidean distance or cosine distance. The set boundary value (hyperparameter) is used to control the minimum interval between positive and negative sample pairs, typically ranging from 0.2 to 1.

[0032] The semantic relationship between texts is determined by cosine similarity. Characterization:

[0033]

[0034] In the formula: and There are two A dimensional vector representing the coordinates of two text blocks in the embedding space; and Is the vector at the th order? Components in the dimensional dimension; cosine similarity The range is between [-1, 1]. A larger value indicates that the two vectors are more consistent in direction, meaning that the semantics of the corresponding two text blocks are closer. For example, it can be used to calculate the correlation between "reservoir water level monitoring" and "river flow monitoring".

[0035] Furthermore, in step 4, the specific process of the multi-scheme dynamic simulation and deduction includes:

[0036] The large model uses the Function Call mechanism to schedule the calculation order of the drainage and water conservancy professional model. First, it calls the flood evolution model and inputs the future rainfall forecast to simulate the baseline scenario of natural flood evolution without engineering intervention.

[0037] Based on the baseline scenario and the basin flood control scheduling rules, the large model generates N different preliminary engineering scheduling plans (such as "Plan A: Reservoir X discharges floodwater at its maximum discharge capacity; Plan B: Reservoir X controls the discharge, and flood storage and detention area Y is activated at the same time").

[0038] For each preliminary engineering scheduling plan, the water engineering scheduling model is called to modify the model boundary conditions, and the flood evolution model is called again to calculate the flood process under the plan. At the same time, the loss assessment model is called to estimate the inundation range and loss corresponding to the plan.

[0039] The simulation results of all schemes (such as water level process lines at key stations, maximum inundation area, and estimated economic losses) are stored in a structured manner.

[0040] Furthermore, the generation of the preliminary engineering scheduling plan is subject to constraints. The constraints, including water balance constraints, reservoir water level not exceeding the design flood level, and downstream protected section flow not exceeding its maximum flow capacity, are formally expressed as:

[0041]

[0042]

[0043]

[0044] In the formula: This refers to the reservoir's capacity at the end of the specified period. The initial storage capacity for the period. For inbound flow, For outbound flow, Because of the long duration, For the reservoir water level, To design flood level, For the downstream protected section flow rate, To protect the maximum flow capacity of the target cross-section;

[0045] The modification of the model boundary conditions specifically includes: using the reservoir discharge flow process and flood diversion gate opening and closing rules calculated by the water engineering scheduling model as new boundary conditions to update the upstream inflow boundary or lateral inflow boundary of the flood evolution model; the calculation method of the boundary conditions follows the principle of water balance, that is:

[0046]

[0047] In the formula: This refers to the reservoir's capacity at the end of the specified period. The initial storage capacity for the period. For inbound flow, For outbound flow, The time period is long.

[0048] Furthermore, in step 5, the specific operations for generating and intelligently optimizing the contingency plan include:

[0049] First, a large model is used to compare and analyze the quantitative indicators of all simulation scenarios (such as which scenario has a lower highest water level and which scenario has a smaller economic loss). At the same time, based on the regulations retrieved from the RAG knowledge base, a qualitative judgment is made on the compliance and operability of the scenarios.

[0050] Then, the non-complex calculation part of the flood control plan is generated by combining the RAG knowledge base, and the water project scheduling part of the flood control plan is generated by calling the workflow based on the professional model. The integration forms a complete dynamic flood control plan text that includes engineering scheduling, patrol and defense, population evacuation and information sharing.

[0051] Third, the system's built-in evaluation module automatically scores the contingency plan from four dimensions: accuracy, completeness, timeliness, and consistency. If the score does not meet the preset standards, the system automatically adjusts the plan parameters and initiates a new round of simulation and evaluation until a contingency plan that meets the standards is generated.

[0052] Furthermore, the calculation formula for the automatic scoring is as follows:

[0053]

[0054] In the formula, For comprehensive scoring, , , , These are normalized scores for accuracy, completeness, timeliness, and consistency, respectively. , , , The preset weighting coefficients, and Scores for each dimension are obtained by quantitatively evaluating the content of the contingency plan against simulation results, knowledge base rules, and user needs.

[0055] Furthermore, the output in step 6 includes the optimized flood control dynamic plan, simulation report, and decision tracing report; the dynamic update is as follows: when new water and rainfall monitoring data of the basin (such as a forecast of a flood process exceeding the standard) is introduced, and a flood process exceeding the standard is forecasted, a new round of multi-scheme dynamic simulation and subsequent process as described in step 4 is automatically or manually triggered, and the flood control dynamic plan is rolled out using the latest data to achieve a dynamic closed loop of "monitoring-simulation-decision-update".

[0056] The advantages and beneficial effects of this invention are:

[0057] 1. The method described in this invention can automate the entire process of flood control plan generation, significantly shorten the time required for plan formulation, and dynamically update the plan based on real-time water and rainfall data. This solves the problem of delayed response in traditional static plans and meets the emergency needs for flood control exceeding standard levels.

[0058] 2. The method described in this invention can realize automated chain simulation and scientific comparison of multiple schemes. The large model can quickly generate and calculate multiple scheduling schemes, and select the optimal scheme by combining quantitative indicators and qualitative judgments, so as to minimize flood inundation losses and improve the scientific nature of the plan.

[0059] 3. This invention realizes the standardized collaborative scheduling of large-scale models and multi-disciplinary water conservancy models, breaks through the technical difficulty of integrating the two, expands the application boundaries of smart water conservancy in the field of flood control and disaster reduction, and provides integrated and intelligent technical support for basin flood control decision-making from data perception, simulation and deduction to plan generation. Attached Figure Description

[0060] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0061] Figure 1 This is a flowchart illustrating the intelligent generation method for dynamic flood control plans based on a large model for floods exceeding standard levels, as described in an embodiment of the present invention.

[0062] Figure 2 Example images showing the conversion of flood control plan-related documents from their original format to Markdown format;

[0063] Figure 3 This is an example diagram of a flood control professional document after semantic-based text segmentation processing.

[0064] Figure 4 This is an example image of a slice of the RAG vector knowledge base in the water conservancy field after text segmentation and vectorization.

[0065] Figure 5 This is a flowchart illustrating the workflow for calling professional models. Detailed Implementation

[0066] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments disclosed herein will be described in further detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0067] Example 1:

[0068] This embodiment takes the Huaihe River basin above the Zhengyangguan section as an example to provide a method for intelligent generation of dynamic flood control plans for floods exceeding standard levels based on a large model. Figure 1 As shown, the specific steps include:

[0069] Step 1: Basic Data Collection and Preprocessing

[0070] S11, through basin flood control analysis, has 22 important reservoirs, 6 sluice gates, and 10 flood storage and detention areas upstream of Zhengyangguan. Basic data related to flood control exceeding standard levels within the basin were collected, including data on water conservancy projects, flood control plan documents, and hydrological and rainfall data. Specifically, this includes:

[0071] ① Collect information on the characteristic water levels and storage capacities of reservoirs in these water conservancy projects, reservoir operation rules, reservoir water level-storage capacity-area curves, reservoir water level-flow curves, sluice gate parameters, sluice gate opening-discharge curves, sluice gate operation rules, flood storage and detention area water level-storage capacity-area curves, and flood storage and detention area activation rules, etc.

[0072] ② Collect relevant documents and materials on flood control plans for the Zhengyangguan control basin. These include basin flood control plans, flood defense plans for floods exceeding standard levels, river system flood control and emergency rescue plans, urban flood and drought disaster prevention plans, flood diversion gate operation plans, and flood control and emergency rescue plans for dangerous sections of dikes and embankments.

[0073] ③ Data collection interface for water and rainfall conditions, mainly including rainfall forecasts, rainfall data observed at rain gauge stations, flow and water level observation data from hydrological stations, and observation data of reservoir discharge and inflow.

[0074] S12 involves preprocessing the collected data, including data cleaning and format conversion. Specifically, for the collected text files from the Huai River basin above Zhengyangguan, typos and grammatical errors are corrected, and redundant content (such as repeated paragraphs) is removed. For Excel spreadsheets, the focus is on handling missing values ​​(filling with interpolation or statistics), outliers (identified, corrected, deleted, or marked), and duplicate data (deleted and merged after comparing key fields). To ensure effectiveness, a quality assessment mechanism needs to be established, and manual spot checks should be conducted. After the document content is checked, the text files (such as PDF and Word documents) need to be converted to Markdown format, and the integrity of special elements such as charts should be checked. An example of the converted Markdown document is attached. Figure 2As shown. Excel data can be converted to CSV format using tools (such as the pandas library) for batch conversion. After conversion, the documents should be further checked for garbled characters caused by encoding issues (such as UTF-8, GBK).

[0075] Step 2, Building the RAG Knowledge Base

[0076] For the preprocessed data, the document-type data is first processed by text chunking using a semantic-based chunking strategy. Each chunk is marked with an "@" symbol before and after it. An example document for chunking is shown below. Figure 3 As shown.

[0077] Text blocks are converted into vector representations using specialized embedding models (such as word vector models for water conservancy optimization). These models are fine-tuned on semantic corpora through a contrastive learning loss function. The embedding model maps text to a high-dimensional space, represents semantic relationships through vector distances (such as cosine similarity), and stores the generated vectors in an efficient vector database with indexes to improve retrieval speed.

[0078] Specifically, the fine-tuning process of the pre-trained professional embedding model adopts the triplet loss function. Through iterative training with a large amount of water conservancy text corpus, the model parameters are optimized so that it can more accurately capture the semantic relationships between water conservancy professional terms.

[0079]

[0080] In the formula: The embedding vector of the anchor sample; The embedding vector of a positive sample belonging to the same category as the anchor point; This is the embedding vector of a negative sample that belongs to a different category than the anchor point; Distance metrics typically employ Euclidean distance or cosine distance. The set boundary value (hyperparameter) is used to control the minimum interval between positive and negative sample pairs, typically ranging from 0.2 to 1.

[0081] The semantic relationship between texts is determined by cosine similarity. Characterization:

[0082]

[0083] In the formula: and There are two A dimensional vector representing the coordinates of two text blocks in the embedding space; and Is the vector at the th order? Components in the dimensional dimension; cosine similarity The range is between [-1, 1]. A larger value indicates that the two vectors are more consistent in direction, meaning that the semantics of the corresponding two text blocks are closer. For example, it can be used to calculate the correlation between "reservoir water level monitoring" and "river flow monitoring".

[0084] By iteratively training on the water semantic corpus dataset, the model parameters are optimized to bring semantically similar text blocks closer together in the vector space and push semantically different text blocks further apart. The training iteration parameters include, but are not limited to: the number of training epochs is set to 10-50 epochs, the learning rate is dynamically adjusted based on the cosine annealing strategy, and the initial learning rate is set to 2e-5.

[0085] An example of the knowledge base after slicing in this embodiment is as follows: Figure 4 As shown.

[0086] Step 3: Professional model call workflow setup

[0087] Establish a professional model-calling workflow to enable the large model to schedule and perform analysis calculations on the water conservancy professional model, such as... Figure 5 As shown, the specific steps include:

[0088] S31, Professional Modeling: For 38 water projects within the section above Zhengyangguan, including 22 reservoirs, 6 sluice gates and 10 flood storage and detention areas, a joint scheduling model is constructed; at the same time, a hydrological forecasting model, a flood evolution model and an inundation loss assessment model for the basin above Zhengyangguan are constructed.

[0089] The hydrological forecasting model can adopt the Xin'anjiang model based on the principles of infiltration runoff and storage runoff; the flood evolution model is constructed based on the Saint-Venant equations to simulate the evolution of floods in river channels and flood storage areas; the water engineering scheduling model calculates the engineering operation mode under different water inflow conditions according to reservoir scheduling rules (such as scheduling diagrams and discharge curves) and constraints; the loss assessment model estimates the economic losses caused by flood inundation based on water depth-loss curves and land use data.

[0090] The Xin'anjiang model is based on a three-source water division structure, which includes total runoff. Classified as surface runoff , soil in the stream and underground runoff ;

[0091] The flood evolution model is based on the Saint-Venant equations as follows:

[0092] The one-dimensional Saint-Venant equations are in the following form:

[0093] Continuity equation:

[0094]

[0095] Momentum equation:

[0096]

[0097] In the formula: The cross-sectional area of ​​the water passage (m²); Flow rate (m³ / s); Time (s); The distance (m) is along the direction of water flow. Lateral inflow (m³ / s / m); The acceleration due to gravity (m / s²) Water level (m); This refers to the friction slope.

[0098] The two-dimensional Saint-Venant equations are in the following form:

[0099] Continuity equation:

[0100]

[0101] Momentum equation:

[0102]

[0103]

[0104] In the formula: Water depth (m); , They are respectively , Flow velocity component in the direction (m / s); , This represents the shear stress component on the bed surface; The density of water is (kg / m³).

[0105] The water project scheduling model calculates the project operation mode under different water inflow conditions based on reservoir scheduling rules and constraints.

[0106] The loss assessment model estimates the economic losses caused by flooding based on the water depth-loss curve and land use data.

[0107] S32, User Intent Recognition: User intent recognition is achieved through water conservancy corpus collection and scene segmentation, data annotation and enhancement;

[0108] The user intent recognition module employs a combination of rule-based and semantic understanding methods. First, an intent classification dataset containing scenarios such as "scheduling scheme simulation" and "flood risk assessment" is constructed using water conservancy corpora. Then, a pre-trained language model (such as BERT) is fine-tuned to build an intent classifier. The determination rule is: when the confidence score of the highest probability intent category output by the model exceeds a preset threshold (e.g., 0.85), it is determined to be that intent; otherwise, a vague intent clarification mechanism is triggered to guide the user to provide more specific instructions.

[0109] S33, Interface Development: When it is recognized that a user needs to call a professional model, the professional model needs to be integrated before the call, including interface specifications, input and output parameter definitions, and interface development adaptation.

[0110] S34, Workflow Engine Development: Finally, based on the above work, a workflow engine was developed based on scheduling logic design and automated task arrangement to realize the automated scheduling of the large model on the water conservancy professional model.

[0111] Step 4: Dynamic simulation and deduction of multiple scenarios

[0112] The large-scale model automatically orchestrates the computational order of specialized models through a Function Call mechanism. First, it calls the flood evolution model, inputting future rainfall forecasts, to simulate the natural evolution of the flood without engineering intervention (baseline scenario). Then, based on the baseline scenario and the scheduling model, the large-scale model generates N different preliminary engineering scheduling schemes.

[0113] The generation of the preliminary engineering scheduling plan is subject to constraints. The constraints, including water balance constraints, reservoir water level not exceeding the design flood level, and downstream protected section flow not exceeding its maximum flow capacity, are formally expressed as:

[0114]

[0115]

[0116]

[0117] In the formula: This refers to the reservoir's capacity at the end of the specified period. The initial storage capacity for the period. For inbound flow, For outbound flow, Because of the long duration, For the reservoir water level, To design flood level, For the downstream protected section flow rate, To protect the maximum flow capacity of the target cross section.

[0118] For example, "Plan A: Foziling Reservoir and Xianghongdian Reservoir discharge floodwater at their maximum discharge capacity; Plan B: Foziling Reservoir and Xianghongdian Reservoir control the discharge, and at the same time activate Mengwa Flood Storage and Detention Area."

[0119] For each preliminary engineering scheduling plan, a loss assessment model is further invoked to estimate the inundation range and losses, and the model boundary conditions are modified. Simulation results for all plans (such as water level process lines at key stations, maximum inundation area, and estimated economic losses) are structured and stored for comparison. Specifically, modifying the model boundary conditions includes using the reservoir discharge flow process and flood diversion gate opening and closing rules calculated by the water engineering scheduling model as new boundary conditions to update the upstream inflow boundary or lateral inflow boundary of the flood evolution model. The calculation method for the boundary conditions follows the principle of water balance, i.e.:

[0120]

[0121] In the formula: This refers to the reservoir's capacity at the end of the specified period. The initial storage capacity for the period. For inbound flow, For outbound flow, The time period is long.

[0122] Step 5: Contingency Plan Generation and Intelligent Optimization

[0123] The large model performs a comprehensive analysis of the simulation results of all schemes in step 4.

[0124] First, the quantitative key indicators are compared (such as the lower highest water levels of major flood control reservoirs like Foziling Reservoir and Xianghongdian Reservoir, and the smaller economic losses after the activation of Mengwa and Jiangtanghu flood storage areas). At the same time, based on the regulations and expert experience retrieved from the constructed Huaihe River RAG knowledge base (such as "prioritizing densely populated areas" and "avoiding large-scale personnel transfer at night"), the compliance and operability of the plan are qualitatively judged, and a suitable dispatch plan is selected.

[0125] Furthermore, the large model calls upon the Huaihe River RAG knowledge base to generate the non-model calculation part of the flood control plan. Combined with the analysis results of the engineering scheduling scheme, a preliminary flood control plan is finally generated, which includes elements such as engineering scheduling, patrol and defense, population evacuation, and information sharing.

[0126] Meanwhile, the system's built-in evaluation module automatically scores the contingency plan from four dimensions: accuracy, completeness, timeliness, and consistency. The calculation formula is as follows:

[0127]

[0128] in, For comprehensive scoring, , , , The scores are normalized scores for accuracy, completeness, timeliness, and consistency (ranging from 0 to 100). , , , The preset weighting coefficients, and Scores for each dimension are obtained by quantitatively evaluating the content of the contingency plan against simulation results, knowledge base rules, and user needs.

[0129] If the score does not meet the preset standard, the system can automatically adjust the scheme parameters and initiate a new round of simulation (step 4) and evaluation (step 5) until a satisfactory plan is generated.

[0130] Step 6: Contingency Plan Output and Dynamic Updates

[0131] The system outputs the final recommended dynamic flood control plan for floods exceeding the standard in the Huai River Basin, and can provide detailed simulation reports and decision-making source tracing reports explaining why this plan was chosen. As the flood continues and new monitoring data is received, if the forecast indicates that the Zhengyangguan section exceeds the guaranteed water level by 29.3m, then a flood exceeding the standard has occurred. The system can automatically or manually trigger a new round of processes (starting from step 4), using the latest data to continuously revise the plan, achieving a dynamic closed loop of "monitoring-simulation-decision-update".

[0132] Finally, it should be noted that the above is only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention (such as the application of various formulas, the order of steps, etc.) without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for intelligently generating dynamic flood control plans for floods exceeding standard levels based on a large model, characterized in that, The method includes the following steps: Step 1: Basic Data Collection and Preprocessing Collect basic data related to the prevention and control of floods exceeding the standard in the basin, and perform preprocessing operations on the collected data, including data cleaning and format conversion. Step 2, Knowledge Base Construction The preprocessed data is subjected to text segmentation and vectorization to construct a RAG knowledge base in the field of water conservancy. The text segmentation adopts a semantic-based segmentation strategy, and markers are added before and after the segmented text content. The vectorization process uses a pre-trained professional embedding model fine-tuned with water conservancy corpus to convert the segmented text into high-dimensional vectors. The semantic relationship between the texts is represented by cosine similarity, and the generated vectors are stored in a vector database and indexed. Step 3: Professional model call workflow setup Establish a professional model invocation workflow to enable the large model to schedule and perform analysis calculations on the water conservancy professional model. Specific steps include: S31, Professional Model Modeling: Develop water conservancy professional models, including water project scheduling models, hydrological forecasting models, flood evolution models, and loss assessment models; S32, User Intent Recognition: User intent recognition is achieved through water conservancy corpus collection and scene segmentation, data annotation and enhancement; S33, Interface Development: Integrate the aforementioned water conservancy professional model, complete the interface specification formulation, input and output parameter definition, and interface development and adaptation; S34, Workflow Engine Development: Based on scheduling logic design and automated task arrangement, develop a workflow engine to realize automated scheduling of large models on water conservancy professional models; Step 4: Dynamic simulation and deduction of multiple scenarios Based on the professional model built in step 3, the workflow is invoked to carry out dynamic simulation and deduction of multiple schemes for flood control exceeding the standard through the large model; Step 5: Contingency Plan Generation and Intelligent Optimization Based on the multi-scheme dynamic simulation results obtained in step 4, and combined with the RAG knowledge base described in step 2, the generation and intelligent optimization of the dynamic flood control plan for floods exceeding the standard are completed. Step 6: Contingency Plan Output and Dynamic Updates The optimized dynamic flood control plan is output and dynamically updated based on real-time monitoring data of the basin, forming a dynamic closed loop for flood control decision-making.

2. The method according to claim 1, characterized in that, In step 1, the basic data includes water conservancy project data within the basin, flood control plan-related documents and materials, and water and rainfall data; The water conservancy project data includes reservoir characteristic water levels and capacity, reservoir scheduling rules, reservoir water level-capacity-area curves, reservoir water level-flow curves, sluice gate parameters, sluice gate opening-discharge curves, sluice gate scheduling rules, flood storage and detention area water level-capacity-area curves, and flood storage and detention area activation rules. The relevant documents and materials for the flood control plan include basin flood control plan, flood defense plan for floods exceeding standard levels, river system flood control and emergency rescue plan, urban water and drought disaster prevention plan, flood diversion gate operation plan, and flood control and emergency rescue plan for dangerous sections of the river. The hydrological and rainfall data includes rainfall forecasts, rainfall data observed at rain gauge stations, flow and water level observation data from hydrological stations, and reservoir discharge and inflow observation data.

3. The method according to claim 1, characterized in that, In step 1, the data cleaning operations include: correcting typos and grammatical errors and removing redundant content from text-based data; filling missing values ​​in tabular data using interpolation or statistics, correcting, deleting or marking outliers, deleting and merging duplicate data by comparing key fields; and establishing a quality assessment mechanism and conducting manual spot checks after data cleaning. The format conversion operations include: converting text files to Markdown format and verifying the integrity of charts, and converting Excel spreadsheet data to CSV or JSON format.

4. The method according to claim 1, characterized in that, In step 2, the fine-tuning process of the pre-trained professional embedding model uses a triplet loss function: In the formula: The embedding vector of the anchor sample; The embedding vector of a positive sample belonging to the same category as the anchor point; This is the embedding vector of a negative sample that belongs to a different category than the anchor point; Distance metric function; These are the set boundary values; The semantic relationship between texts is determined by cosine similarity. Characterization: In the formula: and There are two A dimensional vector representing the coordinates of two text blocks in the embedding space; and Is the vector at the th position? Components in the dimensional dimension; cosine similarity The range is between [-1, 1].

5. The method according to claim 1, characterized in that, Step 4, the specific process of the multi-scheme dynamic simulation and deduction includes: The large model uses the Function Call mechanism to schedule the calculation order of the drainage and water conservancy professional model. First, it calls the flood evolution model and inputs the future rainfall forecast to simulate the baseline scenario of natural flood evolution without engineering intervention. Based on the baseline scenario and the basin flood control scheduling rules, the large model generates N different preliminary engineering scheduling schemes; For each preliminary engineering scheduling plan, the water engineering scheduling model is called to modify the model boundary conditions, and the flood evolution model is called again to calculate the flood process under the plan. At the same time, the loss assessment model is called to estimate the inundation range and loss corresponding to the plan. The simulation results of all schemes are stored in a structured manner.

6. The method according to claim 5, characterized in that, The generation of the preliminary engineering scheduling plan is subject to constraints. The constraints, including water balance constraints, reservoir water level not exceeding the design flood level, and downstream protected section flow not exceeding its maximum flow capacity, are formally expressed as: In the formula: This refers to the reservoir's capacity at the end of the specified period. The initial storage capacity for the period. For inbound flow, For outbound flow, Because of the long duration, For the reservoir water level, To design flood level, For the downstream protected section flow rate, To protect the maximum flow capacity of the target cross-section; The modification of the model boundary conditions specifically includes: using the reservoir discharge flow process and flood diversion gate opening and closing rules calculated by the water engineering scheduling model as new boundary conditions to update the upstream inflow boundary or lateral inflow boundary of the flood evolution model; the calculation method of the boundary conditions follows the principle of water balance, that is: In the formula: This refers to the reservoir's capacity at the end of the specified period. The initial storage capacity for the period. For inbound flow, For outbound flow, The time period is long.

7. The method according to claim 1, characterized in that, Step 5, the specific operations of generating and intelligently optimizing the contingency plan include: First, a large model is used to compare and analyze the quantitative indicators of all simulation scenarios. At the same time, the compliance and operability of the scenarios are qualitatively judged based on the regulations retrieved from the RAG knowledge base. Then, the non-complex calculation part of the flood control plan is generated by combining the RAG knowledge base, and the water project scheduling part of the flood control plan is generated by calling the workflow based on the professional model. The integration forms a complete dynamic flood control plan text that includes engineering scheduling, patrol and defense, population evacuation and information sharing. Third, the system's built-in evaluation module automatically scores the contingency plan from four dimensions: accuracy, completeness, timeliness, and consistency. If the score does not meet the preset standards, the system automatically adjusts the plan parameters and initiates a new round of simulation and evaluation until a contingency plan that meets the standards is generated.

8. The method according to claim 7, characterized in that, The formula for calculating the automatic score is as follows: In the formula, For comprehensive scoring, , , , These are normalized scores for accuracy, completeness, timeliness, and consistency, respectively. , , , The preset weighting coefficients, and .

9. The method according to claim 1, characterized in that, The output of step 6 includes the optimized flood control dynamic plan, simulation report, and decision tracing report; the dynamic update is as follows: when new water and rainfall monitoring data of the basin are introduced and a flood process exceeding the standard is predicted, a new round of multi-scheme dynamic simulation and subsequent process as described in step 4 is automatically or manually triggered, and the flood control dynamic plan is rolled out using the latest data.

Images