Airport rainfall flood early warning method, device and medium based on large language model

Abstract

This invention relates to an airport rainfall and flood early warning method, electronic device, and medium based on a large language model, for use in sponge airports. The early warning method includes: based on a rainstorm and flood management model, using a large language model for data augmentation to generate a multi-factor flood early warning dataset containing regionally specific factor data and airport-specific factor data; based on a retrieval-enhanced generation mechanism, vectorizing and calculating feature bias on the current input data, retrieving the most relevant historical case context information from the multi-factor flood early warning dataset; adaptively fusing the multi-factor features and the most relevant historical case context information, extracting key information through multi-head self-attention and attention pooling, and outputting the flood risk level and early warning result. This invention can significantly improve the accuracy, real-time performance, and interpretability of flood risk assessment, providing more reliable, traceable, and intelligent flood early warning and decision support for sponge airports.

Description

Technical Field

[0001] This invention belongs to the field of hydrological monitoring and artificial intelligence early warning technology, specifically relating to an airport rainfall and flood early warning method, device and medium based on a large language model. Background Technology

[0002] With accelerating urbanization and intensifying climate change, major transportation hubs such as airports face increasingly severe stormwater risks. In recent years, the application of Storm Water Management Models (SWMMs) and sponge airport design concepts in flood warning systems marks significant progress in rainfall data acquisition, simulation analysis, and integrated management. Existing warning methods primarily rely on multi-source stormwater management data, including historical rainfall records, real-time runoff monitoring data, and various environmental parameters unique to sponge airport systems. This multi-factor data fusion system provides comprehensive information support for flood prediction, reflecting a shift from single-data-driven decision-making to comprehensive environmental awareness. However, while traditional machine learning methods can handle such heterogeneous data and achieve input-output mapping, their models are typically trained from scratch, lacking an understanding of prior knowledge such as hydrological mechanisms and environmental evolution patterns. Therefore, they have limitations in identifying key environmental factors and discerning abnormal hydrological conditions. Summary of the Invention

[0003] In view of the above-mentioned defects or deficiencies in the prior art, the present invention proposes an airport rainfall and flood early warning method, system and storage medium based on a large language model.

[0004] This invention provides an airport rainfall and flood early warning method based on a large language model, for use in sponge airports. The early warning method includes: based on a rainstorm and flood management model, using a large language model for data augmentation to generate a multi-factor flood early warning dataset containing regional specific factor data and airport specific factor data; based on a retrieval-enhanced generation mechanism, performing vectorized encoding and feature bias calculation on the current input data, and retrieving the most relevant historical case context information from the multi-factor flood early warning dataset; adaptively fusing the multi-factor features and the most relevant historical case context information, extracting key information through multi-head self-attention and attention pooling, and outputting the flood risk level and early warning result.

[0005] In one alternative implementation, the region-specific factors include precipitation, date, regional climate type, historical weather conditions, permeable pavement area, rainwater harvesting system capacity, vegetation coverage, reservoir capacity, and groundwater recharge rate; the airport-specific factors include storm event simulation, flood and waterlogging analysis, drainage system processes, real-time hydrological monitoring, identification of key waterlogging points, and early warning thresholds.

[0006] In one optional implementation, based on a retrieval enhancement generation mechanism, the current input data is vectorized and feature deviations are calculated to retrieve the most relevant historical case context information from the multi-factor flood warning dataset. Specifically, this includes: constructing a vectorized retrieval model for the current input data; calculating key feature deviations based on the vectorized retrieval model to output the feature dimension most relevant to the current scenario; and performing similarity retrieval based on the feature dimension most relevant to the current scenario to output the most relevant historical case context information.

[0007] In one alternative implementation, the vectorized retrieval model is encoded using a multilayer perceptron and combined with a sinusoidal transform-based position embedding mechanism to generate a unified multi-factor embedding representation.

[0008] In one optional implementation, the key feature deviation calculation uses the Z-score method to standardize the query features and historical data of the knowledge base, so as to eliminate redundant feature interference and improve retrieval accuracy.

[0009] In one optional implementation, the similarity retrieval uses Euclidean distance to calculate vector similarity, and selects the k historical records with the smallest distance as the most relevant historical case context information.

[0010] In one optional implementation, the multi-factor features and the context information of the most relevant historical cases are adaptively fused, and key information is extracted through multi-head self-attention and attention pooling to output the flood risk level and early warning result. Specifically, this includes: concatenating and compressing the multi-factor features to obtain an embedding model; fusing the context information of the most relevant historical cases and the embedding model based on the multi-head self-attention mechanism to obtain fused features; and extracting key information through an attention pooling layer based on the fused features to output the flood risk level and early warning result.

[0011] In one alternative implementation, residual connections and batch normalization are introduced during the multi-head self-attention fusion process. The final fused representation is generated through attention pooling, thereby achieving adaptive fusion of local details and global semantics and improving the accuracy and robustness of flood warning.

[0012] The present invention also proposes an electronic device, comprising: at least one processor; a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to execute any of the airport rainfall and flood early warning methods based on large language models.

[0013] The present invention also proposes a storage medium storing a computer program, which, when executed by a processor, implements any of the airport rainfall and flood early warning methods based on a large language model as described in the present invention.

[0014] The beneficial technical effects of this invention are as follows: by constructing a multi-factor airport flood early warning dataset through large language model data augmentation, combining retrieval enhancement to generate accurate matching historical cases, and using multi-head self-attention and attention pooling to achieve feature adaptive fusion, it effectively integrates multi-dimensional influencing factors of region and airport, significantly improving the accuracy, real-time performance and interpretability of flood risk assessment, and providing more reliable, traceable and intelligent flood early warning and decision support for sponge airports. Attached Figure Description

[0015] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0016] Figure 1 A flowchart of an airport rainfall and flood early warning method based on a large language model provided in an embodiment of the present invention;

[0017] Figure 2 This is a diagram of the large language model input structure for an airport rainfall and flood early warning method based on a large language model, provided in another embodiment of the present invention.

[0018] Figure 3 The structure diagram of the enhanced RAG framework for airport rainfall and flood early warning method based on a large language model provided in an embodiment of the present invention;

[0019] Figure 4 The diagram shows the multi-factor fusion structure of the airport rainfall and flood early warning method based on a large language model, as provided in an embodiment of the present invention. Detailed Implementation

[0020] In recent years, some studies have attempted to introduce external knowledge bases to enhance the environmental perception and reasoning capabilities of models, improving interpretability and the completeness of decision-making basis by integrating structured environmental information. However, in the context of SWMM and sponge airport design, existing knowledge fusion methods mostly focus on extracting hierarchical and structured knowledge from specific knowledge sources and embedding it into a fixed framework, resulting in limited improvement in actual prediction tasks. Furthermore, these methods typically favor structured data such as rainfall, drainage capacity, and soil moisture, while neglecting unstructured information such as text reports, maintenance records, and real-time sensor logs, which are often more prevalent in real-world environments and contain important early warning clues.

[0021] The emergence of Large Language Models (LLMs) has offered new possibilities for integrating complex environmental knowledge with multi-source rainfall data. Their powerful semantic understanding and knowledge association capabilities demonstrated in various prediction tasks give them the potential to serve as a generalized environmental knowledge base. However, directly applying them to flood early warning still faces significant challenges, especially the possibility of models producing "illusionary" outputs that contradict reality, affecting the reliability and security of the early warning system. Therefore, how to effectively utilize the massive knowledge understanding and reasoning capabilities of LLMs, combined with SWMM simulation mechanisms and multi-source monitoring data from sponge airports, to construct an intelligent early warning method and system that can integrate structured and unstructured environmental information and possess hydrological mechanism cognition and reliable inference capabilities has become an urgent technical problem to be solved.

[0022] To address the aforementioned issues, this invention designs a multi-factor data structure for flood early warning at sponge airports and constructs a high-quality dataset covering different sponge airports and different areas within the same airport. This dataset serves as a knowledge base, providing prior knowledge for LLM (Large Language Modeling) and leveraging the large language model (MLM) to enhance the input context, enabling the model to produce more effective outputs. This expands the application of LLM in vertical fields, particularly in flood early warning. This invention proposes REMFLOW (a multi-factor rainfall and flood early warning method based on retrieval-enhanced generation (RAG), an LLM method for flood data analysis and early warning. This method improves prediction accuracy and enhances the intelligence of sponge airport early warning.

[0023] The following is in conjunction with the appendix Figures 1 to 4 The present application will be further described in detail with reference to the embodiments. It is understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the invention. It should also be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0024] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0025] Please refer to Figure 1 This invention provides an airport rainfall and flood early warning method based on a large language model, applicable to sponge airports. The early warning method includes the following steps:

[0026] Step S101: Based on the rainstorm and flood management model, data augmentation is performed using a large language model to generate a multi-factor flood early warning dataset that includes regional specific factor data and airport specific factor data.

[0027] Step S103: Based on the retrieval enhancement generation mechanism, the current input data is vectorized and feature deviation is calculated to retrieve the most relevant historical case context information from the multi-factor flood warning dataset.

[0028] Step S105: Adaptively fuse multi-factor features with the contextual information of the most relevant historical cases, extract key information through multi-head self-attention and attention pooling, and output flood risk level and early warning results.

[0029] In this embodiment, a multi-factor flood early warning dataset for sponge airports based on SWMM is constructed, covering 9 region-specific factors (such as precipitation and climate type) and 5 airport-specific factors (such as drainage system and real-time hydrological monitoring). GPT-4o is used for data augmentation and manual validation to provide a high-quality domain knowledge foundation for the model, addressing the problems of traditional data being singular and lacking professional background. Using GPT-4o for data augmentation and combined with manual validation, 1600 high-quality data entries are generated. The region-specific factor dataset contains 1000 data entries from different climate zones, and the airport-specific factor dataset contains 600 data entries from the flight area, terminal area, and ground transportation area. The data output includes risk levels ([low], [medium], [high]) and corresponding recommended measures, ensuring that the data covers high-risk scenarios and meets actual early warning needs.

[0030] like Figure 3 As shown, the enhanced large language model based on RAG is as follows:

[0031] Step 1: Generate by referring to the example.

[0032] Step 2: Relevant content search.

[0033] Step 3: Risk Prediction and Recommendations.

[0034] Multiple factors:

[0035] f1 f2 … f n

[0036] Dataset Sample

[0037] {

[0038] "enter":{

[0039] Rainfall: "50",

[0040] Regional climate type: "Subtropical monsoon climate"

[0041] Permeable pavement area: "40000"

[0042] Rainwater harvesting system capacity: "500,000", ...

[0043] },

[0044] Output:{

[0045] Current risk level: "Medium"

[0046] }

[0047] }

[0048] Context

[0049] Regional specific factors, precipitation, defined as the total rainfall within a specific region, are crucial for assessing flood risk and the effectiveness of water resource management systems.

[0050] Unit of measurement: usually millimeters (mm).

[0051] Definition: The SWMM model is used to simulate water accumulation and analyze the flood depth and duration in different areas of the airport.

[0052] Units of measurement: water depth is measured in centimeters (cm) or meters (m), and duration is measured in hours (h)...

[0053] Note: Based on the context and search paragraphs, potential risks related to the input data are identified, and corresponding reasonable suggestions are provided as output.

[0054] Large language model output

[0055] Current risk level: Medium

[0056] Recommended measures: [

[0057] Monitor the water level in the rainwater harvesting system and increase storage capacity as needed.

[0058] ……]"

[0059] This design employs an enhanced large language model based on RAG (Radar Encoding and Data Type). Through vectorized retrieval and Z-score bias calculation, it matches the most relevant historical cases from the knowledge base. Combined with the semantic understanding capabilities of LLM (Limited Language Model), it generates accurate warning levels and response suggestions, addressing the knowledge gaps and illusions inherent in LLM within vertical domains, thus improving prediction accuracy and reliability. This step achieves rapid knowledge matching through vectorized retrieval and eliminates low-relevance and highly interfering redundant information through feature bias calculation. By fully combining the high reliability of the retrieval-enhanced generation mechanism with the strong semantic understanding capabilities of the large language model, it effectively alleviates the knowledge boundary limitations, real-time insufficiency, and content illusions inherent in large language models in the airport hydrological vertical domain, ensuring the authenticity and professionalism of the warning basis.

[0060] This invention introduces a retrieval enhancement generation mechanism, which combines vectorized retrieval with feature deviation calculation to achieve accurate matching of historical cases. Relying on an external knowledge base, it provides real and traceable professional evidence for large language models, effectively overcoming the technical defects of large language models in vertical domain applications, such as knowledge lag, reasoning illusion, and semantic bias, and significantly improving the authenticity, reliability, and interpretability of flood warning results.

[0061] like Figure 4 As shown, a multi-factor adaptive fusion module is proposed, employing a multi-head self-attention mechanism and attention pooling to collaboratively fuse retrieved contextual information with multi-factor embeddings. This enables the complementarity of local details and global semantics, solving the feature loss problem in multi-factor information integration and enhancing the robustness and accuracy of flood early warning. In the figure, q (query) represents the query vector, k (key) represents the key vector, and v (value) represents the value vector.

[0062] By leveraging the synergistic effect of multi-head self-attention and attention pooling, efficient integration of multi-source heterogeneous data can be achieved, avoiding problems such as feature loss, information redundancy, and weakening of key factors that exist in traditional fusion methods. This significantly improves the adaptability and judgment accuracy of flood early warning models under complex and variable rainfall conditions.

[0063] This invention constructs a multi-dimensional early warning dataset covering regional geographical factors and airport-specific operational factors by combining the SWMM model with a large language model for data augmentation. This solves the problems of traditional flood early warning methods, such as single data sources, lack of feature dimensions, and lack of adaptability to airport-specific scenarios. It significantly improves the completeness, richness, and domain professionalism of the dataset, laying a data foundation for high-precision early warning.

[0064] like Figure 2 As shown in the context note, rainfall: rainfall in a specific area, in millimeters (mm), helps in assessing flood risk and the effectiveness of water resource management.

[0065] Rainwater harvesting system capacity: The maximum volume of rainwater that can be collected, measured in liters, used to characterize the system's ability to handle excess rainwater.

[0066] ...

[0067] Related elements

[0068] Rainfall date...

[0069] Nodes retrieved from the knowledge base

[0070] Regional climate type, historical weather conditions, and permeable pavement area

[0071] knowledge base

[0072] Retrieved knowledge

[0073] [Context]

[0074] Rainfall: ...

[0075] Rainwater harvesting system capacity: ...

[0076] ...

[0077]

Example

[0078] {

[0079] Input: {

[0080] Rainfall: 50 mm

[0081] Regional climate type: Subtropical monsoon climate

[0082] Permeable pavement area: 40,000 square meters

[0083] ...

[0084] }

[0085] }

[0086] This invention employs a multi-factor adaptive fusion module that combines multi-head self-attention with attention pooling. This module can dynamically weight and deeply aggregate multi-source heterogeneous features, achieving complementary enhancement of local hydrological details and global scene semantics. It solves the problems of key feature loss, unreasonable weight allocation, and low information coupling in the process of multi-factor information integration, and significantly improves the robustness and generalization ability of the early warning model under special conditions such as extreme rainfall and complex terrain.

[0087] This invention is specifically designed for the facility characteristics and operational needs of sponge airports. The early warning results can directly output the risk level and response plan, enabling real-time perception, rapid assessment and accurate early warning of flood risks in sponge airports. This effectively reduces the risk of airport flooding, ensures the safe operation of key facilities such as airport runways, terminals and drainage systems, and improves the overall efficiency of airport flood prevention and emergency response and the level of intelligent management.

[0088] The method of this invention is modular, scalable and portable, and can be adapted to sponge airport scenarios of different scales and layouts. It can complete data adaptation and model iteration without large-scale reconstruction, reduce engineering application costs, and has good industrialization and promotion value.

[0089] Furthermore, region-specific factors include precipitation, date, regional climate type, historical weather conditions, permeable pavement area, rainwater harvesting system capacity, vegetation coverage, reservoir capacity, and groundwater recharge rate; airport-specific factors include storm event simulation, flood and waterlogging analysis, drainage system processes, real-time hydrological monitoring, identification of key waterlogging points, and early warning thresholds.

[0090] By finely dividing region-specific factors into airport-specific factors, a multi-dimensional feature system is constructed that is comprehensive, hierarchical, and scenario-specific. Region-specific factors include precipitation, date, regional climate type, historical weather conditions, permeable pavement area, rainwater harvesting system capacity, vegetation coverage, reservoir capacity, and groundwater recharge rate. Airport-specific factors include storm event simulation, flood and urban waterlogging analysis, drainage system processes, real-time hydrological monitoring, identification of key urban waterlogging points, and early warning thresholds. Based on this dual-dimensional factor system, data augmentation, retrieval matching, and feature fusion are performed, resulting in the following significant and beneficial technical effects:

[0091] This invention employs nine regionally specific factors to construct a complete dimension of natural hydrology and sponge city facilities characteristics, comprehensively covering key indicators such as meteorological conditions, surface infiltration, water storage capacity, ecological regulation and storage, and groundwater circulation. It can fully reflect the natural runoff patterns and the actual regulation and storage capacity of sponge city facilities in the area where the airport is located, making the early warning model fully consistent with the regional hydrological background conditions, greatly improving the authenticity and environmental adaptability of data-driven approaches, and avoiding early warning deviations caused by the lack of regional characteristics.

[0092] By setting five airport-specific factors, focusing on airport-specific storm simulation, flooding analysis, drainage network operation status, real-time hydrological monitoring data, and key point early warning rules, this system deeply aligns with the flood prevention needs of key facilities such as airport runways, aprons, terminals, and underground utility tunnels. This enables the early warning model to transform from general hydrological forecasting to a professional model that is airport-specific, operation-oriented, and precise in its location, solving the problem that traditional flood early warning methods cannot adapt to the complex operating conditions and operational safety requirements of airports.

[0093] The multi-dimensional factors covering static facility parameters, dynamic monitoring data, historical patterns, and simulation results enable the early warning method to simultaneously possess long-term pattern learning capabilities and real-time operational condition response capabilities. It can adapt to complex scenarios such as continuous rainfall, sudden rainstorms, tidal backwater, and pipeline congestion, significantly improving the generalization ability, robustness, and long-term operational reliability of the sponge airport flood early warning system. The dual-dimensional factor system achieves deep coupling between natural environmental conditions and the operational status of artificial facilities, organically combining regional macro-hydrological patterns with the micro-level characteristics of airport waterlogging points. This allows the early warning model to not only grasp the overall rainfall and flood trends but also accurately locate key risk points at the airport, significantly improving the pertinence, practicality, and enforceability of the early warning results. It provides direct and reliable decision-making basis for airport flood control scheduling and emergency response.

[0094] Data augmentation and retrieval matching are carried out based on clear and detailed regional and airport-specific factors, enabling large language models to perform inference calculations within a highly structured, domain-specific feature space. This effectively reduces interference from irrelevant factors, enhances the model's understanding of flood formation mechanisms, urban flooding transmission paths, and risk diffusion range, further suppresses model illusions, and improves the consistency and stability of early warning level judgments.

[0095] By identifying key flood points and designing customized early warning thresholds, the system can accurately output early warning results from macro-level risk levels to point-specific, threshold-specific, and visualized data. It can directly connect to the airport's intelligent management and control platform and flood control emergency system, shortening emergency response time, improving the efficiency of drainage facility scheduling, and minimizing the impact of flooding on airport flight operations, facility safety, and personnel safety.

[0096] Further, in step S103, based on the retrieval enhancement generation mechanism, the current input data is vectorized and feature bias is calculated to retrieve the most relevant historical case context information from the multi-factor flood warning dataset. This specifically includes the following steps:

[0097] Step S1031: Construct a vectorized retrieval model for the current input data.

[0098] Step S1033: Calculate the key feature deviation based on the vectorized retrieval model and output the feature dimension most relevant to the current scene.

[0099] Step S1035: Perform similarity retrieval based on the feature dimension most relevant to the current scene, and output the most relevant historical case context information.

[0100] By using dedicated vectorized encoding for the current input data, we can avoid feature ambiguity and information loss caused by general encoding, which better matches the multi-dimensional and strongly coupled business characteristics of flood early warning, and significantly improve retrieval relevance and hit rate.

[0101] First, the feature dimensions most relevant to the current scenario are selected by calculating the key feature deviation. Then, similarity retrieval is performed to achieve dimensionality reduction and retrieval focus, greatly reducing invalid calculations, accelerating the recall speed of historical case context, and meeting the real-time requirements of flood warning.

[0102] Directly outputting historical case context information most relevant to the current scenario can provide traceable and verifiable historical references for flood warning levels and response strategies, thereby enhancing the credibility of warning conclusions and the persuasiveness of decision-making.

[0103] For multi-factor flood early warning datasets, key influencing factors (such as rainfall, water level, topography, underlying surface, etc.) are dynamically focused through feature bias, effectively handling multi-factor coupling interference and maintaining stable and reliable retrieval results under complex working conditions.

[0104] By leveraging real historical cases for enhanced retrieval, and avoiding reliance on a single model for reasoning, the system effectively mitigates prediction biases in small sample sizes and extreme flood scenarios, thereby enhancing its adaptability to different river basins and varying rainfall and water conditions.

[0105] The vectorized retrieval model uses a multilayer perceptron for encoding and combines it with a sinusoidal transform-based position embedding mechanism to generate a unified multi-factor embedding representation.

[0106] For input data f, including regional factors Date factors Other continuous variables are encoded using a multilayer perceptron (MLP) and combined with a sinusoidal transform-based positional embedding mechanism to generate a unified multi-factor embedding model. This method can not only capture the complex relationships between different regions, but also represent the temporal characteristics of dates and other continuous variables, through multi-factor embedding models. It can be represented as:

[0107] (1)

[0108] For query text Large-Scale Embedding Model (LEM) is used to generate text embedding models suitable for capturing rich contextual information in text records. :

[0109] (2)

[0110] This invention uses MLP to encode continuous variables such as regions and dates, combined with LEM to generate text embeddings, to achieve a unified vector expression of structured data and unstructured text. This can solve the technical problem of heterogeneous data types and difficulty in fusion retrieval in multi-factor flood early warning.

[0111] By introducing sinusoidal transform position embedding, the temporal patterns of continuous variables such as dates and their spatial relationships can be effectively modeled, capturing complex spatiotemporal dependencies and significantly improving the adaptability of retrieval to time- and space-sensitive scenarios.

[0112] MLP and LEM are adapted to numerical and textual modalities respectively. The generated embeddings align semantics with numerical features and have unified dimensions, improving the accuracy of cross-modal similarity calculation and enhancing the reliability of historical case retrieval.

[0113] This invention unifies multiple factors such as region, date, and text into an embedding vector, providing high-quality, low-noise input for subsequent feature deviation calculation and similarity retrieval, thereby improving the accuracy of early warning context matching and the interpretability of decisions. The MLP encoding is lightweight, the sinusoidal position embedding is parameterless, and the LEM handles text semantics. The overall architecture has controllable computational overhead and high parallelism, meeting the real-time retrieval requirements of flood early warnings.

[0114] Among them, the Z-score method is used to calculate the key feature deviation, and the query features and historical data of the knowledge base are standardized to eliminate redundant feature interference and improve retrieval accuracy.

[0115] The embedded query data Q is used to find relevant known data from the knowledge base, including data on region, weather, and sponge airport features. This retrieval process matches the current query with past records that have similar multi-factor structures. For each feature in the query data... To identify the differences between the current query and historical data, each feature is calculated. Z-score Si:

[0116] (3)

[0117] in, This represents historical data corresponding to a feature in the knowledge base; mean( K i )express K i The mean (average); std( K i )express K i The standard deviation; the Z-score is used to filter the feature dimensions most relevant to the current scene, improving retrieval accuracy.

[0118] This invention standardizes query features and historical database features to the same dimension using Z-scores, eliminating differences in dimensions and numerical ranges, and enabling fair and comparable retrieval of features based on multiple factors such as region, weather, and airport. It quantifies the deviation of current features from historical distributions using standard deviation as a scale, automatically highlighting anomalous dimensions that play a decisive role in flood warnings, and achieving adaptive filtering of key features. Irrelevant features with low deviation are weakened and filtered, reducing redundant dimensional interference and significantly improving the purity and accuracy of subsequent similarity retrieval.

[0119] Z-scores are simple to calculate, have clear statistical significance, and have low computational overhead, making them suitable for the timeliness requirements of real-time flood early warning retrieval and online decision-making.

[0120] The similarity retrieval uses Euclidean distance to calculate vector similarity, and selects the k historical records with the smallest distance as the most relevant historical case context information.

[0121] Euclidean distance is used to measure query embedding. With knowledge base embedding Similarity, Euclidean distance It can be represented as:

[0122] (4)

[0123] in, q This represents the query sample that needs to be predicted at present. n Indicates the first in the knowledge base n The system has 10 known historical records; the k historical records with the smallest distance are selected as the most relevant contexts to provide key reference information for LLM, such as similar rainfall patterns or historical drainage performance, effectively alleviating the illusion problem of LLM in vertical fields.

[0124] Based on step S105, the multi-factor features and the contextual information of the most relevant historical cases are adaptively fused. Key information is extracted through multi-head self-attention and attention pooling, and the flood risk level and early warning results are output. The specific steps include the following:

[0125] Step S1051: Concatenate and compress the multi-factor features to obtain the embedding model.

[0126] Step S1053: Based on the multi-head self-attention mechanism, fuse the contextual information of the most relevant historical cases with the embedding model to obtain the fused features.

[0127] Step S1055: Based on the fusion features, extract key information through the attention pooling layer and output the flood risk level and early warning results.

[0128] Multi-factor feature concatenation and compression. Extending the RAG framework, the key factor text data of multiple retrieved weather and scenario information that meet the criteria are represented as... These factors are then connected together and mapped to a unified embedding model using an MLP. Embedded Model It can be represented as:

[0129] (5)

[0130] By constructing a unified embedding model through feature splicing and compression, current monitoring features can be organically combined with historical case contexts, making full use of historical early warning experience, which can improve the completeness and reliability of the basis for flood risk assessment.

[0131] By leveraging a multi-head self-attention mechanism to accurately capture associated features, it can adaptively focus on key factors and historical case fragments that significantly impact flood risk, effectively model the complex dependencies between multiple factors, and improve the accuracy and discriminativeness of feature fusion.

[0132] By efficiently extracting core decision-making information through attention pooling, the high-dimensional features after fusion are reduced in dimensionality and key information is aggregated. Redundant information interference is eliminated, core risk features are highlighted, and the accuracy and stability of risk level judgment are greatly improved.

[0133] Furthermore, residual connections and batch normalization are introduced in the multi-head self-attention fusion process, and the final fusion representation is generated through attention pooling, so as to achieve adaptive fusion of local details and global semantics and improve the accuracy and robustness of flood warning.

[0134] To capture the dependencies between key factors, a fusion network based on self-attention is proposed. LLM is used to compress contextual information to obtain context embeddings. Then, a multi-head self-attention (MHSA) mechanism is applied to fuse... and To capture complex dependencies between factors:

[0135] (6)

[0136] Residual connections and batch normalization (BatchNorm) are introduced to stabilize the training process.

[0137] Through the attention pooling layer (AttnPool) from the augmented representation Extract the most critical information to generate the final fused representation Z:

[0138] (7)

[0139] This representation is used to output risk levels and specific recommendations, enabling adaptive fusion of local details and global semantics, significantly improving the accuracy and robustness of flood warnings.

[0140] This invention proposes a multi-factor data structure for flood early warning at sponge airports and constructs a high-quality dataset covering different sponge airports and different areas within the same airport. This dataset serves as a knowledge base, providing prior knowledge for LLM (Large Language Modeling), and enhances the input context using a large language model (LLM), enabling the model to produce more effective outputs. This expands the application of LLM in vertical fields, particularly in flood early warning.

[0141] This invention proposes REMFLOW, an LLM method based on retrieval-enhanced generation (RAG), for flood data analysis and early warning. This method improves forecast accuracy and enhances the intelligence of sponge airport early warning systems.

[0142] A multi-factor flood early warning dataset based on SWMM was constructed, employing a progressive strategy of "factor definition - data augmentation - knowledge base construction," as detailed below:

[0143] This invention first defines a multi-factor early warning framework based on the SWMM model and expert knowledge, covering nine region-specific factors (including precipitation, date, regional climate type, historical weather conditions, permeable pavement area, rainwater harvesting system capacity, vegetation coverage, reservoir capacity, and groundwater recharge rate) and five airport-specific factors (including storm event simulation, flood and waterlogging analysis, drainage system processes, real-time hydrological monitoring, identification of key waterlogging points, and early warning thresholds), providing a structured foundation for data generation. Then, GPT-4o is used for data augmentation and supplemented by manual verification to generate 1600 high-quality data entries. The region-specific factor dataset contains 1000 data entries from different climate zones, and the airport-specific factor dataset contains 600 data entries from the flight area, terminal area, and ground transportation area. Data output includes risk levels (low, medium, and high) and corresponding recommended measures, ensuring data coverage of high-risk scenarios and meeting actual early warning needs. Finally, the generated vectorized data is stored in the FAISS vector database to construct a searchable knowledge base, providing a high-quality domain knowledge foundation for the model and solving the problems of traditional data being singular and lacking professional background.

[0144] The cascaded process of "multi-factor embedding - bias calculation - similarity retrieval" is adopted, and the specific process is as follows:

[0145] This invention first uses a multilayer perceptron (MLP) combined with sinusoidal position embedding to generate a unified multi-factor embedding for the input data f (including continuous variables such as region and date). At the same time, the query text model Generating text embeddings using a large-scale embedding model (LEM) Then, Si is calculated for each feature Q in the query data, and finally, Euclidean distance is used to measure the query embedding. With knowledge embedding By using similarity to select the k historical records with the smallest distance as the most relevant context, key reference information is provided for LLM, effectively alleviating the knowledge gap and illusion problem of LLM in vertical fields, and improving the accuracy and reliability of prediction.

[0146] A multi-factor adaptive fusion module is constructed, employing a collaborative mechanism of "feature concatenation—self-attention fusion—attention pooling," as detailed below:

[0147] This invention concatenates multiple key factor texts retrieved and maps them to a unified embedding model using an MLP. Meanwhile, LLM is used to compress context information to obtain a context embedding model. Then, a multi-head self-attention (MHSA) mechanism is applied for fusion. and Residual connections and batch normalization are introduced to stabilize the training process; finally, attention pooling layers are used to extract data from the unified embedding model. Key information is extracted to generate the final fusion model Z. This enables the sponge airport stormwater forecasting system to output risk levels and specific recommendations, achieving adaptive fusion of local details and global semantics, significantly improving the accuracy and robustness of flood warnings.

[0148] The present invention also proposes an electronic device comprising: at least one processor; a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform any one of the airport rainfall and flood early warning methods based on a large language model.

[0149] The present invention also proposes a storage medium storing a computer program, which, when executed by a processor, implements an airport rainfall and flood early warning method based on a large language model.

[0150] In the above embodiments, the storage medium may include, but is not limited to, various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0151] Optionally, the processor executes the method steps described in the above embodiments based on the program code stored in the storage medium.

[0152] Specific examples in this embodiment can be found in the examples described in the above embodiments and optional implementations, and will not be repeated here.

[0153] Obviously, those skilled in the art should understand that the modules or steps of the present disclosure described above can be implemented using general-purpose computing devices, which can be centralized on a single computing device or distributed on a network of multiple computing devices.

[0154] Alternatively, they can be implemented using computer-executable program code, thereby allowing them to be stored in a storage device for execution by a computing device. Furthermore, in some cases, the steps shown or described can be performed in a different order than those described herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, this disclosure is not limited to any particular combination of hardware and software.

[0155] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the inventive concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A method for early warning of rainfall and flooding at airports based on a large language model, characterized in that, For sponge airports, the early warning method includes: Based on the rainstorm and flood management model, data augmentation is performed using a large language model to generate a multi-factor flood early warning dataset that includes regional specific factor data and airport specific factor data. Based on the retrieval enhancement generation mechanism, the current input data is vectorized and feature deviation is calculated to retrieve the most relevant historical case context information from the multi-factor flood warning dataset; Adaptive fusion of multi-factor features and the contextual information of the most relevant historical cases is performed, and key information is extracted through multi-head self-attention and attention pooling, outputting flood risk level and early warning results; The region-specific factors include precipitation, date, regional climate type, historical weather conditions, permeable pavement area, rainwater harvesting system capacity, vegetation coverage, reservoir capacity, and groundwater recharge rate; the airport-specific factors include storm event simulation, flood and waterlogging analysis, drainage system processes, real-time hydrological monitoring, identification of key waterlogging points, and early warning thresholds.

2. The airport rainfall and flood early warning method based on a large language model according to claim 1, characterized in that, Based on the retrieval enhancement generation mechanism, the current input data is vectorized and feature bias is calculated. The most relevant historical case context information is retrieved from the multi-factor flood warning dataset, specifically including: A vectorized retrieval model is constructed based on the current input data; Based on the vectorized retrieval model, key feature deviations are calculated, and the feature dimension most relevant to the current scene is output. Similarity retrieval is performed based on the feature dimension most relevant to the current scene, and the most relevant historical case context information is output.

3. The airport rainfall and flood early warning method based on a large language model according to claim 2, characterized in that, The vectorized retrieval model uses a multilayer perceptron for encoding and combines it with a sinusoidal transform-based position embedding mechanism to generate a unified multi-factor embedding representation.

4. The method according to claim 2, characterized in that, The key feature deviation calculation adopts the Z-score method to standardize the query features and historical data of the knowledge base, so as to eliminate redundant feature interference and improve retrieval accuracy.

5. The airport rainfall and flood early warning method based on a large language model according to claim 2, characterized in that, The similarity retrieval uses Euclidean distance to calculate vector similarity, and selects the k historical records with the smallest distance as the most relevant historical case context information.

6. The airport rainfall and flood early warning method based on a large language model according to any one of claims 1 to 5, characterized in that, Adaptive fusion of multi-factor features and the contextual information of the most relevant historical cases is performed. Key information is extracted through multi-head self-attention and attention pooling to output flood risk level and early warning results, specifically including: The multi-factor features are concatenated and compressed to obtain an embedding model; Based on the multi-head self-attention mechanism, the most relevant historical case context information is fused with the embedding model to obtain fused features; Based on the fusion features, key information is extracted through an attention pooling layer to output the flood risk level and early warning results.

7. The airport rainfall and flood early warning method based on a large language model according to claim 6, characterized in that, By introducing residual connections and batch normalization in the multi-head self-attention fusion process, and generating the final fusion representation through attention pooling, adaptive fusion of local details and global semantics is achieved, thereby improving the accuracy and robustness of flood early warning.

8. An electronic device, characterized in that, include: At least one processor; A memory communicatively connected to at least one processor; wherein the memory stores instructions executable by at least one processor, the instructions being executed by at least one processor to enable at least one processor to perform the airport rainfall and flood early warning method based on a large language model as described in any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that, The device contains a computer program that, when executed by a processor, implements the airport rainfall and flood early warning method based on a large language model as described in any one of claims 1 to 7.

Images