A Method for Modeling and Blocking Urban Flooding Disaster Chains Based on Large Language Models and Complex Network Analysis

Abstract

This invention relates to a method for modeling and blocking urban flooding disaster chains based on large language models and complex network analysis, belonging to the field of disaster prevention and mitigation. Using multi-source media text as input data, a large language model is used to automatically extract disaster events and their causal relationships. Through association rule analysis, support, confidence, and lift between events are calculated to construct a weighted directed disaster chain network, which characterizes the cascading propagation relationships of urban flooding events. Based on complex network theory, indicators such as the centrality, connectivity, and network efficiency of nodes and edges are calculated to identify key nodes and vulnerable propagation paths. A progressive node and edge removal experiment is designed to evaluate various intervention strategies and determine the optimal disaster chain blocking scheme. This invention enables automated modeling and quantitative analysis of disaster chain networks from unstructured text data, rapidly identifying key propagation links and system vulnerabilities in urban flooding, significantly improving the accuracy of urban flooding disaster chain identification, analysis efficiency, and the scientific nature of prevention and control decisions.

Description

Technical Field

[0001] This invention belongs to the field of disaster prevention and mitigation, and involves large language model information processing technology and complex network theory. It relates to a method for modeling and assessing the disruption of urban flooding disaster chains based on large language models and complex network analysis. Background Technology

[0002] Urban flooding has a significant chain reaction of destructive effects on urban systems, particularly critical infrastructure such as subways, underground commercial spaces, and power and communication hubs. A single flooding event often triggers a series of secondary and cascading disasters, including power outages, traffic disruptions, and people being trapped, posing a significant threat to urban safety. Therefore, clarifying the key nodes and propagation paths of the disaster chain is a core prerequisite for achieving precise prevention and control and effectively preventing the spread of disasters.

[0003] Currently, the identification and analysis of urban flooding disaster chains mostly employ complex network-based modeling methods. These techniques typically abstract disaster events as nodes and the causal or influence relationships between events as directed edges, thereby constructing a disaster propagation network for identifying key nodes, analyzing system vulnerabilities, or predicting disaster propagation paths. For example, in their paper "JIAO L, LUO Q, LU H, et al. Research on the urban rail transit disaster chain: Critical nodes, edge vulnerability and breaking strategy[J]. International Journal of Disaster Risk Reduction, 2024, 102: 104258.", Jiao et al., based on 562 urban rail transit flooding cases, applied complex network theory to identify key nodes and vulnerable propagation paths such as subway stations and underground transportation hubs. Similarly, in their paper "GUO H, HUANG C, ZHANG C, et al. A novel comprehensive system for analyzing and evaluating storm surge disaster chains based on complex networks[J]. Frontiers in Marine Science, 2024, 11: 1510791.", Guo et al. analyzed 87 urban rainstorm disaster cases, revealing the key triggering points for secondary events such as subway flooding, traffic congestion, and power outages. These published papers all establish disaster chain network models using existing structured data or manually extracted information, and conduct risk identification and assessment based on topological indicators.

[0004] However, existing research generally relies on manual extraction or rule-based analysis of disaster chain event information, making it difficult to process large-scale, multi-source, and unstructured disaster text data, resulting in low efficiency and insufficient coverage in disaster chain construction. Furthermore, existing research largely focuses on risk identification and topology analysis, lacking a systematic assessment mechanism for disaster chain "blocking strategies," thus failing to provide actionable intervention basis for pre-disaster decision-making. Summary of the Invention

[0005] To address the aforementioned issues, this invention proposes a method for modeling urban flooding disaster chains and evaluating blocking strategies that combines large language models with complex network analysis. This method automatically extracts disaster events and their cascading relationships from multi-source text data, constructs a weighted directed disaster chain network, and identifies key nodes, critical paths, and optimal blocking strategies through network indicators and simulation analysis. This improves the automation, accuracy, and prevention guidance of disaster chain analysis.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A method for modeling and blocking urban flooding disaster chains based on large language models and complex network analysis is presented. This method takes multi-source media text as input and sequentially completes the automatic extraction of disaster event types, identification of causal relationships between disaster event types, construction of disaster chain networks, identification of key nodes and edges, and simulation of blocking strategies, realizing a complete technical path from data acquisition to model analysis. The specific steps are as follows:

[0008] The first step involves collecting multi-source media text data, using the large language model GPT-4o to evaluate the relevance of the multi-source media texts, and extracting information to obtain the disaster event types and causal relationships between disaster event types in the multi-source media texts; specifically:

[0009] Step 1.1: Based on the research objectives, determine the keywords for data retrieval and collect text data from a multi-source media information database. Each piece of text data consists of the title and body of the multi-source media text. After collection, firstly, by designing prompt words, the large language model GPT-4o is asked to analyze the content of each multi-source media text title to determine whether the corresponding multi-source media text body is likely to contain a description of urban flooding disaster events, thereby filtering out relevant multi-source media texts that are highly related to the flooding events.

[0010] Step 1.2: For the relevant multi-source media texts selected in Step 1.1, their main text is divided into different text blocks. The purpose of text segmentation is to process text content at a finer granularity, allowing each text block to focus on describing one aspect of the event (such as time, location, impact, etc.). Then, word embedding is performed on each text block to convert the text data into a vector representation so that the system can understand and analyze its semantics.

[0011] Step 1.3: For the text blocks and their vector representations segmented in Step 1.2, a semantic similarity-based method is used for precise information matching. Specifically, firstly, each text block is systematically compared with a set of predefined prompt words containing information about disaster event types and causal relationships between disaster event types extracted from the text content. Subsequently, the text block with the highest semantic similarity is identified based on the semantic similarity ranking between the text block and the prompt words.

[0012] Step 1.4: The text blocks identified in Step 1.3 are used as contextual information and input together with prompt words containing information on the types of disaster events and the causal relationships between disaster event types into the large language model GPT-4o to guide it to perform the information extraction task and generate the types of disaster events and the causal relationships between disaster event types involved in each text block.

[0013] Step 1.5, to address the inconsistency in terminology across different data sources, post-processing is required after obtaining the disaster event types and causal relationships between them in Step 1.4, to standardize the descriptive content. Specifically, the post-processing involves:

[0014] First, using an unsupervised clustering algorithm based on semantic similarity, disaster event types with similar expressions but different wording are automatically grouped into the same group, generating initial clustering results. These initial clustering results are then manually verified and adjusted to ultimately form a stable disaster event type classification system.

[0015] Secondly, the causal relationships between disaster event types were also manually verified to correct any unreasonable causal links, thereby obtaining a stable set of causal relationships between disaster event types.

[0016] The second step involves constructing a disaster chain network based on the disaster event types and causal relationships obtained in the first step, and calculating edge weights using association rules to obtain a weighted directed disaster chain network. Specifically:

[0017] Step 2.1: Based on the disaster event types and causal relationships between disaster event types output in Step 1.5, establish a disaster chain network. According to complex network theory, each disaster event type is defined as a node, and the causal relationships between disaster event types are defined as directed edges, pointing from the primary disaster event to the secondary event, forming a directed disaster chain network.

[0018] Step 2.2 introduces an association rule analysis method to quantify the edge association strength in the disaster chain network formed in Step 2.1 and ultimately determine the edge weights. This method is used to discover frequent co-occurrence relationships and directional dependencies between different disaster event types. Quantification is achieved by calculating three typical indicators: Support, Confidence, and Lift. Specifically:

[0019] First, all historical disaster cases from multi-source media texts are compiled, with each independent case considered a transaction, and the specific disaster event type appearing in each case considered an item within that transaction. A "disaster event type-case co-occurrence matrix" is constructed to record the frequency of different disaster event types co-occurring in the same case. Based on this, the support, confidence, and lift between any two disaster event types are calculated using the following formulas:

[0020] (1)

[0021] (2)

[0022] (3)

[0023] Among them, support reflects the co-occurrence frequency (Formula (1)); confidence measures the probability of disaster event type B occurring under the condition that disaster event type A occurs (Formula (2)); lift is used to quantify the relative dependence between two disaster event types (Formula (3)).

[0024] Finally, the edge weights of the three indicators are calculated to form a weighted directed disaster chain network. To unify the dimensions and improve the comparability of the indicators, the three indicators are each Min-Max normalized to the [0,1] interval, resulting in three normalized values. For each directed edge in the disaster chain network, the arithmetic mean of the three normalized values ​​is taken as the comprehensive weight value, which comprehensively considers co-occurrence intensity, directionality, and dependency.

[0025] The third step involves performing complex network index analysis based on the weighted directed disaster chain network formed in the second step to identify key nodes and edges in the disaster chain network; specifically:

[0026] After constructing the weighted directed disaster chain network in step 2.2, a multi-dimensional network characteristic analysis is performed. This analysis includes two parts: node index calculation and edge index calculation.

[0027] Step 3.1, Calculate node indicators;

[0028] To assess the importance of each node in a weighted directed disaster chain network, four typical indicators need to be calculated: degree centrality, intermediate centrality, connectivity, and the change in network efficiency after node removal.

[0029] Degree centrality reflects the degree of direct connection between a node and other nodes, and its calculation formula is as follows:

[0030] (4)

[0031] (5)

[0032] (6)

[0033] in, Represents a node The out-degree; N is the total number of nodes; Represents a node Is it possible to reach the node directly? If possible, the value is 1; otherwise, it is 0. Represents a node in-degree; This indicates the degree centrality of a node.

[0034] The mediator centrality measures the frequency with which a node appears in the shortest paths of all reachable node pairs, reflecting its mediating role in disaster propagation. Its calculation formula is:

[0035] (7)

[0036] in, Represents a node The median centrality, This represents the distance from node i to j. , This represents the number of paths that pass through node k.

[0037] The connectivity represents the breadth of influence an event can propagate within the system, and is calculated using the following formula:

[0038] (8)

[0039] in, Indicates connectivity, From the starting node The number of nodes that can be connected.

[0040] The change in network efficiency after node removal is defined as the difference between the network efficiency after removing a single node and the original network efficiency, and is calculated using the following formula:

[0041] (9)

[0042] (10)

[0043] in, It is a node and nodes The shortest path distance between them, where E represents network efficiency; This represents the change in network efficiency after node i is removed. This represents the network efficiency after node i is removed.

[0044] Step 3.2, calculate the edge index;

[0045] Further evaluation of the structural role and vulnerability of each edge in a weighted directed disaster chain network requires calculation of five types of indicators: edge betweenness, change in average path length, edge connectivity, edge vulnerability, and change in network efficiency after edge removal.

[0046] The formula for calculating the edge betweenness number is as follows:

[0047] (11)

[0048] in, Representing an edge The edge betweenness number, where V represents the set of nodes. Represents a node and nodes Between the edges The number of shortest paths.

[0049] The change in average path length is used to measure the difference between the average path length after removing a single edge and the original average path length. The calculation formula is as follows:

[0050] (12)

[0051] (13)

[0052] in, It is the total number of nodes. Represents a node and nodes The distance between them Indicates the average path length; This represents the change in average path length after edge e is removed. This represents the average path length after edge e is removed.

[0053] The edge connectivity measures the change in the number of reachable nodes in the network after removing an edge, and the formula is as follows:

[0054] (14)

[0055] in, It is the network connectivity after removing edge e. Indicates the starting node after removing edge e. The number of nodes that can be connected, and It represents the total number of nodes in the new network.

[0056] Edge vulnerability is a composite index used to comprehensively measure edge structure control, transmission efficiency, and network robustness. The calculation formula is:

[0057] (15)

[0058] in, Denotes the edge betweenness number of edge e; This represents the average path length after removing edge e; This represents the network connectivity after removing edge e.

[0059] The change in network efficiency after edge removal is defined as the difference between the network efficiency after removing a single edge and the original network efficiency, and the calculation formula is as follows:

[0060] (16)

[0061] (17)

[0062] in, It is a node and nodes The shortest path distance between them, where E represents network efficiency; This represents the change in network efficiency after edge e is removed. This represents the network efficiency after edge e is removed.

[0063] The fourth step involves constructing a node and edge removal sequence based on the metrics calculated in the third step, simulating the successive removal process, and using global network metrics to quantitatively evaluate the destructive effects of different sorting strategies on the weighted directed disaster chain network. Specifically:

[0064] Step 4.1 establishes an evaluation framework for the optimal blocking strategy of the weighted directed disaster chain network. An experiment is designed to sequentially remove nodes and edges based on the node and edge indices obtained in Steps 3.1 and 3.2. The removal order is determined according to different index values, and nodes or edges in the weighted directed disaster chain network are removed sequentially to simulate the disaster chain disruption process under different intervention logics, thereby evaluating the relative effectiveness of each node / edge index in blocking disaster propagation. During this process, the degree of damage to the weighted directed disaster chain network structure is monitored using two global network indices: the maximum connected subgraph ratio, which reflects the degree to which the backbone structure of the weighted directed disaster chain network is preserved after the removal operation; and the number of connected subgraphs, which measures the fragmentation level of the weighted directed disaster chain network. A decrease in the former indicates a break in the main disaster chain, while an increase in the latter indicates that the system has decomposed into multiple isolated units.

[0065] Step 4.2: Based on the global network metrics determined in Step 4.1, perform node removal simulations and edge removal simulations respectively. Specifically:

[0066] First, a node removal simulation scheme is set up based on the indicators in step 3.1. These indicators include degree centrality, between-centrality, connectivity, the change in network efficiency after node removal, and a comprehensive index obtained by the arithmetic mean of the first three, totaling five categories. Node removal sequences are constructed according to the numerical values ​​of these five categories of indicators in descending order, and corresponding nodes in the weighted directed disaster chain network are removed sequentially. After each removal step, the two global network indicators mentioned in step 4.1 are calculated: the maximum connected subgraph ratio and the number of connected subgraphs. The changes in the global network indicators under different sequences are compared. The removal sequence with the faster decrease rate of the maximum connected subgraph ratio and the faster increase rate of the number of connected subgraphs represents a strategy capable of disrupting the weighted directed disaster chain network.

[0067] Secondly, an edge removal simulation scheme is set based on the indicators in step 3.2. These indicators include five categories: edge betweenness, average path length change, edge connectivity, edge vulnerability, and network efficiency change after edge removal. An edge removal sequence is constructed according to the numerical values ​​of these five categories of indicators, from largest to smallest, and corresponding edges in the weighted directed disaster chain network are removed sequentially. After each removal step, two global system indicators are calculated. The removal order in which the maximum connected subgraph ratio decreases faster and the number of connected subgraphs increases faster represents the strategy capable of disrupting the weighted directed disaster chain network.

[0068] Through the above steps, this invention realizes the entire process of automatically identifying disaster event types and their causal relationships from multi-source text data, constructing a weighted directed disaster chain network, and evaluating blocking strategies, which can significantly improve the identification accuracy, analysis efficiency, and scientific nature of prevention and control decisions for urban flooding disaster chains.

[0069] The beneficial effects of this invention are as follows:

[0070] (1) Intelligent and automated information extraction: This invention utilizes the large language model GPT-4o to automatically identify disaster event types and extract causal relationships between disaster event types from multi-source media texts, replacing the traditional manual annotation and rule-based filtering process. This technical feature significantly improves the accuracy and processing efficiency of disaster event identification, enabling the processing of large-scale datasets in a short time and achieving rapid generation and updating of disaster chain data.

[0071] (2) Quantification and weighting of disaster chain modeling: By introducing association rule analysis method, the support, confidence and lift indices are calculated and normalized, and the co-occurrence relationship between disaster event types is transformed into quantifiable edge weights, which improves the accuracy of disaster chain network structure expression and the ability to characterize the intensity of disaster propagation.

[0072] (3) Quantitative and comparable evaluation of blocking strategies: Based on various indicators of nodes and edges, experiments were designed to remove nodes and edges sequentially. Two global network indicators, the maximum connected subgraph ratio and the number of connected subgraphs, were used for dynamic monitoring. This enabled quantitative evaluation of the impact of different intervention strategies on network connectivity and propagation efficiency. Attached Figure Description

[0073] Figure 1 This is a flowchart of the method for modeling and blocking urban flooding disaster chains based on the large language model GPT-4o, as described in this invention;

[0074] Figure 2 This is a network diagram of urban underground space flooding disaster chain constructed in Example 1;

[0075] Figure 3 This is a diagram showing the experimental results of node removal designed in Example 1;

[0076] Figure 4 This is an experimental result diagram of edge removal designed in Example 1. Detailed Implementation

[0077] The present invention will be further described below with reference to specific implementation examples.

[0078] Example 1: Construction and Analysis of Urban Underground Space Flooding Disaster Chain Network

[0079] This embodiment provides a specific implementation of the technical solution of the present invention to illustrate its operability and technical effects, but the scope of the present invention is by no means limited to this example. This embodiment selects "urban underground space" as the research object, and uses the method described in this invention to construct and analyze the urban underground space flooding disaster chain. The overall process is attached. Figure 1 As shown, the specific implementation process is as follows:

[0080] The first step involves collecting multi-source media text data, using the large language model GPT-4o to evaluate the relevance of the multi-source media texts, and extracting information to obtain the disaster event types and causal relationships between them. The specific process is as follows:

[0081] Step 1.1: Determine the keywords for data retrieval based on the research objectives and collect text data from multi-source media information databases. In this embodiment, historical news texts from WiseNews, CNKI (China National Knowledge Infrastructure), and WeChat public account platforms are selected as data sources. Search keyword combinations include "waterlogging" or "flood," as well as "underground space," "subway," "underground parking lot," and "tunnel." The time range is set from 2000 to 2024. After initial collection, a total of 8766 original multi-source media texts were obtained. After data collection, relevance filtering prompts were constructed and input into the large language model GPT-4o. The relevance filtering prompts contain specific instructions for the large language model GPT-4o, requiring the model to perform semantic analysis on the title of each multi-source media text, inferring whether the corresponding body text contains a description of urban waterlogging disasters based on the title content, thereby filtering out multi-source media texts highly related to waterlogging events. After filtering, 1570 texts highly related to underground space waterlogging were retained.

[0082] Step 1.2: For the relevant multi-source media texts selected in Step 1.1, their main text is divided into different text blocks. In this embodiment, the above 1570 texts are segmented in a fine-grained manner, so that each text block focuses on describing a single event such as "the time of flooding in the subway station" or "the situation of people being trapped". Then, word embedding processing is performed on each text block to convert the text data into a vector representation.

[0083] Step 1.3: For the text blocks and their vector representations segmented in Step 1.2, a semantic similarity-based method is used for precise information matching. The system compares each text block with predefined prompts for extracting disaster event types and causal relationships between disaster event types, and sorts them according to semantic similarity, identifying the text block with the highest similarity ranking to ensure that the information input into the model has high-quality contextual relevance. The prompts for extracting disaster event types and causal relationships between disaster event types contain specific instructions for the GPT-4o large language model, used to guide the GPT-4o large language model to analyze the main text content of multi-source media and output the following information: basic information of a specific flood disaster (including time and location), information on the types of disaster events affected by the flood (including event name and specific event description), and causal logical relationships between different disaster event types.

[0084] Step 1.4: The text blocks identified in Step 1.3 are used as contextual information and input together with prompt words containing information about the types of disaster events and the causal relationships between these types into the large language model GPT-4o to guide it in performing the information extraction task. In this embodiment, after manual sampling verification, the model achieved a comprehensive accuracy of 98% in identifying the types of urban underground space flooding disasters and extracting their causal relationships. After deduplication, 630 independent cases of urban underground space flooding disasters were finally compiled.

[0085] Step 1.5: To address the inconsistency in terminology across different data sources, post-processing is required after obtaining the disaster event types and causal relationships between them in Step 1.4 to standardize the descriptive content. First, an unsupervised clustering algorithm based on semantic similarity is used to generate initial clustering results, which are then manually verified. In this embodiment, after standardization, 31 typical urban underground space flooding disaster event types were identified (as shown in Table 1), including "subway station flooding," "tunnel water accumulation," "drainage system overload," and "electric shock accidents." Second, the causal relationships between disaster event types are verified, unreasonable causal links are corrected, and a stable set of causal relationships is formed, as shown in Table 2.

[0086] Table 1 Event types identified by GPT-4o

[0087]

[0088] The second step involves constructing a disaster chain network based on the disaster event types and causal relationships obtained in the first step, and calculating edge weights using association rules to obtain a weighted directed disaster chain network. The specific process is as follows:

[0089] Step 2.1: Based on the disaster event types and causal relationships between disaster event types output in Step 1.5, a directed disaster chain network is established. In this embodiment, the 31 disaster event types in Table 1 are defined as nodes, and the causal relationships between events in Table 2 are defined as directed edges, thus initially constructing an urban underground space flooding disaster chain network structure containing 31 nodes and 54 directed edges.

[0090] Step 2.2, to quantify the correlation strength of edges in the disaster chain network formed in Step 2.1 and ultimately form the edge weights in the disaster chain network, an association rule analysis method is introduced. By constructing a "disaster event type-case co-occurrence matrix", using the aforementioned 630 historical case data, the support, confidence, and lift between any two disaster event types are calculated (the calculation formulas are as described in formulas (1)-(3) of the invention content). Finally, the edge weights are calculated by combining the above three indicators to form a weighted directed disaster chain network. The three indicators are respectively Min-Max normalized to the [0,1] interval, and the arithmetic mean is taken as the comprehensive weight value. In this embodiment, the weighted directed disaster chain network is as follows: Figure 2 As shown.

[0091] Table 2 Event-Triggered Chains Identified by LLM

[0092]

[0093] The third step involves performing complex network index analysis based on the weighted directed disaster chain network formed in the second step to identify key nodes and edges in the disaster chain network. The specific process is as follows:

[0094] Step 3.1, Node Indicator Calculation. To assess the importance of each node in the weighted directed disaster chain network, four typical indicators are calculated: degree centrality, intermediate centrality, connectivity, and the change in network efficiency after node removal (the calculation formulas are as described in formulas (4)-(10) of the invention). In the calculation results of this embodiment, the nodes “Subway Station Flooding (ID:1)”, “Tunnel Water Accumulation (ID:2)”, and “Drainage System Overload (ID:23)” all show high values ​​in degree centrality and intermediate centrality, indicating that they are core hubs in the underground space flooding network. In addition, the change in network efficiency after the node “Floodgate / Barrier Failure (ID:26)” is significant, indicating that the failure of this facility node has a huge impact on the overall network transmission efficiency.

[0095] Step 3.2, edge index calculation. Further assess the structural role and vulnerability of each edge in the weighted directed disaster chain network, and calculate five categories of indicators: edge betweenness, average path length change, edge connectivity, edge vulnerability, and network efficiency change after edge removal (calculation formulas are as described in formulas (11)-(17) of the invention). The calculation results identified several key propagation paths, such as "overloaded drainage system → flooded subway station," "flooded subway station → trapped personnel," and "network failure → difficult evacuation." These edges have high edge betweenness and vulnerability, meaning that cutting off these paths will effectively curb the spread of disaster from facility damage to personnel injury.

[0096] The fourth step involves constructing a node and edge removal sequence based on the metrics calculated in the third step, simulating the successive removal process, and using global network metrics to quantitatively evaluate the destructive effects of different sorting strategies on the weighted directed disaster chain network. The specific process is as follows:

[0097] Step 4.1 establishes an evaluation framework for the optimal blocking strategy of the weighted directed disaster chain network. An experiment is designed to sequentially remove nodes and edges based on the node and edge indices obtained in Steps 3.1 and 3.2. The removal order is determined according to different index values, and nodes or edges in the weighted directed disaster chain network are removed sequentially to simulate the disaster chain disruption process under different intervention logics, thereby evaluating the relative effectiveness of each node / edge index in blocking disaster propagation. During this process, the degree of damage to the weighted directed disaster chain network structure is monitored using two global network indices: the maximum connected subgraph ratio, which reflects the degree to which the backbone structure of the weighted directed disaster chain network is preserved after the removal operation; and the number of connected subgraphs, which measures the fragmentation level of the weighted directed disaster chain network. A decrease in the former indicates a break in the main disaster chain, while an increase in the latter indicates that the system has decomposed into multiple isolated units.

[0098] Step 4.2: Based on the global network metrics determined in Step 4.1, perform node removal simulations and edge removal simulations respectively. Specifically:

[0099] First, a node removal simulation scheme is set up based on the indicators in step 3.1. A node removal sequence is constructed according to five categories of indicators, ranked from largest to smallest: degree centrality, between centrality, connectivity, change in network efficiency after node removal, and comprehensive index. In this embodiment, the node removal experimental results are as follows: Figure 3 As shown. Among them, Figure 3 (a) shows the decreasing trend of the proportion of the largest connected subgraph in the weighted directed disaster chain network as the number of nodes removed increases; Figure 3 (b) The changes in the number of connected subgraphs formed by the splitting of the weighted directed disaster chain network during node removal were statistically analyzed. The results show that the removal strategy based on the change in network efficiency after node removal is the most effective. Under this strategy, removing the top 10 critical nodes in terms of numerical value (such as flooded subway stations, overloaded drainage systems, etc.) rapidly reduced the maximum connected subgraph ratio to below 0.2, and the number of connected subgraphs increased sharply to 14. This indicates that identifying and prioritizing the blocking of critical nodes according to the "change in network efficiency after node removal" index can most quickly break up the underground space flooding disaster chain network, fragmenting it and thus preventing the cascading propagation of the disaster.

[0100] Secondly, an edge removal simulation scheme is set based on the indicators in step 3.2. An edge removal sequence is constructed according to five categories of indicators, ranked from largest to smallest: edge betweenness, average path length change, edge connectivity, edge vulnerability, and change in network efficiency after edge removal. In this embodiment, the edge removal experimental results are as follows: Figure 4 As shown. Among them, Figure 4 (a) Plots the decreasing curves of the maximum connected subgraph proportion as the number of removed edges increases under different edge removal strategies; Figure 4 (b) The growth trend of the number of connected subgraphs with the number of removed edges was recorded. Compared with node removal, the edge removal strategy causes relatively slower damage to the network structure. Among them, the strategy based on "the change in network efficiency after edge removal" is relatively better and can effectively identify critical connections such as "network failure → evacuation difficulties". Through comparison, it was found that in the prevention and control of urban underground space flooding, prioritizing the protection or blocking of key nodes (such as strengthening flood control at subway entrances and upgrading drainage systems) has a higher intervention efficiency than simply cutting off the propagation path.

[0101] This embodiment verifies that the present invention can achieve automatic modeling from multi-source media text to weighted directed disaster chain networks, effectively identify key nodes and edges, and provide technical support for urban flood disaster risk assessment and prevention and control decisions for different key protection objects.

[0102] Finally, it should be noted that the above examples are only used to illustrate the implementation of the present invention. It should be understood that the examples are only used to illustrate the present invention and are not intended to limit the scope of the present invention. After reading the present invention, those skilled in the art can make several modifications and improvements without departing from the concept of the present invention. All modifications to the present invention in various equivalent forms fall within the scope defined by the appended claims and are within the protection scope of the present invention.

Claims

1. A method for modeling and blocking urban flooding disaster chains based on large language models and complex network analysis, characterized in that, The urban flooding disaster chain modeling and blocking method uses multi-source media text as input and sequentially completes the automatic extraction of disaster event types, identification of causal relationships between disaster event types, construction of disaster chain networks, identification of key nodes and edges, and simulation of blocking strategies, realizing the transformation from data acquisition to model analysis; including the following steps: The first step is to collect multi-source media text data, use the large language model GPT-4o to evaluate the relevance of multi-source media texts, and extract information to obtain the disaster event types and causal relationships between disaster event types in multi-source media texts. The second step is to construct a disaster chain network based on the disaster event types and causal relationships between disaster event types obtained in the first step, and to calculate edge weights using association rules to obtain a weighted directed disaster chain network. The second step is specifically as follows: Step 2.1: Based on the types of disaster events and the causal relationships between them, establish a disaster chain network. According to complex network theory, define each type of disaster event as a node and define the causal relationships between disaster event types as directed edges, pointing from the primary disaster event to the secondary event, thus forming a directed disaster chain network. Step 2.2 introduces an association rule analysis method to quantify the association strength of edges in the disaster chain network formed in Step 2.1 and ultimately determine the weights of edges in the disaster chain network. This association rule analysis method is used to discover frequent co-occurrence relationships and directional dependencies between different disaster event types. It is quantified by calculating three typical indicators: Support, Confidence, and Lift. In step 2.2: First, historical disaster cases involved in all multi-source media texts are summarized, with each independent case treated as a transaction, and the specific disaster event type appearing in each case treated as an item in that transaction. A "disaster event type-case co-occurrence matrix" is constructed to record the frequency of different disaster event types co-occurring in the same case. Based on this, the support, confidence, and lift between any two disaster event types are calculated. Finally, the edge weights are calculated by combining the three typical indicators to form a weighted directed disaster chain network; the three typical indicators are respectively Min-Max normalized to the [0,1] interval to obtain three normalized values; for each directed edge in the disaster chain network, the arithmetic mean of the three normalized values ​​is taken as the comprehensive weight value. The third step is to perform complex network index analysis based on the weighted directed disaster chain network formed in the second step, and identify the key nodes and edges in the disaster chain network. After constructing the weighted directed disaster chain network in step 2.2, a multi-dimensional network feature analysis is performed; this analysis includes two parts: node index calculation and edge index calculation. The fourth step involves constructing a node and edge removal sequence based on the indicators calculated in the third step, simulating the successive removal process, and using global network indicators to quantitatively evaluate the destructive effects of different sorting strategies on the weighted directed disaster chain network.

2. The method for modeling and blocking urban flooding disaster chains based on large language models and complex network analysis according to claim 1, characterized in that, The first step is specifically as follows: Step 1.1: Determine the keywords for data retrieval based on the research objectives, collect text data from the multi-source media information database, and each piece of text data consists of the title and body of the multi-source media text; by designing prompt words, the large language model GPT-4o is required to analyze the content of each multi-source media text title and filter out relevant multi-source media texts that are highly related to the flooding event; Step 1.2: For the relevant multi-source media texts selected in Step 1.1, divide their main text into different text blocks; perform word embedding processing on each text block to convert the text data into vector representation; Step 1.3: For the text blocks and their vector representations segmented in Step 1.2, a semantic similarity-based method is used for precise information matching. First, each text block is systematically compared with a set of predefined prompt words containing information about the types of disaster events and the causal relationships between disaster event types in the extracted text content. Subsequently, the text block with the highest semantic similarity to the prompt word is identified. Step 1.4: The text blocks identified in Step 1.3 are used as context information and input together with prompt words containing information on the types of disaster events and the causal relationships between disaster event types in the text content into the large language model GPT-4o to guide it to perform the information extraction task and generate the types of disaster events and the causal relationships between disaster event types involved in each text block. Step 1.5: After obtaining the disaster event type and the causal relationship between disaster event types in step 1.4, post-processing is performed to standardize the description content.

3. The method for modeling and blocking urban flooding disaster chains based on large language models and complex network analysis according to claim 2, characterized in that, The post-processing in step 1.5 specifically includes: First, using an unsupervised clustering algorithm based on semantic similarity, disaster event types with similar expressions but different wording are automatically grouped into the same group to generate initial clustering results. The initial clustering results are then manually verified and adjusted to ultimately form a stable disaster event type classification system. Secondly, the causal relationships between disaster event types are also manually verified to correct any unreasonable causal links and obtain a stable set of causal relationships between disaster event types.

4. The method for modeling and blocking urban flooding disaster chains based on large language models and complex network analysis according to claim 1, characterized in that, The formulas for calculating support, confidence, and lift are as follows: (1) (2) (3) Among them, support reflects the co-occurrence frequency; confidence measures the probability that disaster event type B will occur under the condition that disaster event type A occurs; lift is used to quantify the relative dependency between two disaster event types.

5. The method for modeling and blocking urban flooding disaster chains based on large language models and complex network analysis according to claim 4, characterized in that, The third step is specifically as follows: Step 3.1, Calculate node indicators; Four key metrics—degree centrality, between centrality, connectivity, and the change in network efficiency after node removal—are used to assess the importance of each node in a weighted directed disaster chain network. Step 3.2, calculate the edge index; Five types of indicators are calculated: edge betweenness, average path length change, edge connectivity, edge vulnerability, and network efficiency change after edge removal, to further evaluate the structural role and vulnerability of each edge in a weighted directed disaster chain network.

6. The method for modeling and blocking urban flooding disaster chains based on large language models and complex network analysis according to claim 5, characterized in that, In step 3.1: Degree centrality reflects the degree of direct connection between a node and other nodes, and its calculation formula is as follows: (4) (5) (6) in, Represents a node The out-degree; N is the total number of nodes; Represents a node Is it possible to reach the node directly? If possible, the value is 1; otherwise, it is 0. Represents a node in-degree; Indicates the degree centrality of a node; The medianity is used to measure the frequency with which a node appears in the shortest paths of all reachable node pairs, reflecting its mediating role in the disaster propagation process; its calculation formula is: (7) in, Represents a node The median centrality, This represents the distance from node i to j. , This represents the number of paths that pass through node k. The connectivity represents the breadth of influence an event can propagate within the system, and is calculated using the following formula: (8) in, Indicates connectivity, From the starting node The number of nodes that can be connected; The change in network efficiency after node removal is defined as the difference between the network efficiency after removing a single node and the original network efficiency, and is calculated using the following formula: (9) (10) in, It is a node and nodes The shortest path distance between them E Indicates network efficiency; Represents a node i Change in network efficiency after removal. Represents a node i Network efficiency after removal.

7. The method for modeling and blocking urban flooding disaster chains based on large language models and complex network analysis according to claim 6, characterized in that, In step 3.2: The formula for calculating the edge betweenness number is as follows: (11) in, Representing an edge The edge betweenness number, where V represents the set of nodes. Represents a node and nodes Between the edges The number of shortest paths; The change in average path length is used to measure the difference between the average path length after removing a single edge and the original average path length. The calculation formula is as follows: (12) (13) in, It is the total number of nodes. Represents a node and nodes The distance between them Indicates the average path length; Representing an edge e The change in average path length after removal. Representing an edge e Average path length after removal; The edge connectivity measures the change in the number of reachable nodes in the network after removing an edge, and the formula is as follows: (14) in, It is the network connectivity after removing edge e. Indicates the starting node after removing edge e. The number of nodes that can be connected, and It is the total number of nodes in the new network; The edge vulnerability is a composite index used to comprehensively measure edge structure control, transmission efficiency, and network robustness; the calculation formula is: (15) in, Representing an edge e The number of edge betweennesses; Indicates removing an edge e Average path length after; Indicates removing an edge e Network connectivity afterward; The change in network efficiency after edge removal is defined as the difference between the network efficiency after removing a single edge and the original network efficiency, and the calculation formula is as follows: (16) (17) in, It is a node and nodes The shortest path distance between them E Indicates network efficiency; Representing an edge e Change in network efficiency after removal. Representing an edge e Network efficiency after removal.

8. The method for modeling and blocking urban flooding disaster chains based on large language models and complex network analysis according to claim 7, characterized in that, The fourth step is specifically as follows: Step 4.1: Establish an evaluation framework for the optimal blocking strategy in a weighted directed disaster chain network; The design incorporates a sequential removal experiment based on the node and edge indices obtained in steps 3.1 and 3.

2. The removal order is determined according to different index values, and nodes or edges in the weighted directed disaster chain network are removed sequentially to simulate the disaster chain disruption process under different intervention logics, evaluating the relative effectiveness of each node and edge index in blocking disaster propagation. During this process, the degree of damage to the weighted directed disaster chain network structure is monitored using two global network indices: the maximum connected subgraph ratio, reflecting the degree to which the backbone structure of the weighted directed disaster chain network is preserved after the removal operation; and the number of connected subgraphs, measuring the fragmentation level of the weighted directed disaster chain network. A decrease in the former indicates a break in the main disaster chain, while an increase in the latter indicates that the system decomposes into multiple isolated units. Step 4.2: Based on the global network metrics determined in Step 4.1, perform node removal simulation and edge removal simulation respectively; specifically: First, a node removal simulation scheme is set up based on the indicators in step 3.

1. The indicators include degree centrality, between centrality, connectivity, the change in network efficiency after node removal, and a comprehensive index obtained by the arithmetic mean of the first three, totaling 5 types of indicators. Node removal sequences are constructed according to the numerical values ​​of these 5 types of indicators from largest to smallest, and corresponding nodes in the weighted directed disaster chain network are removed in sequence. After each removal step, the two global network indicators mentioned in step 4.1 are calculated, namely the maximum connected subgraph ratio and the number of connected subgraphs. The changes in the global network indicators under different sequences are compared. The removal sequence with a faster rate of decrease in the maximum connected subgraph ratio and a faster rate of increase in the number of connected subgraphs represents a strategy that can destroy the weighted directed disaster chain network. Secondly, an edge removal simulation scheme is set based on the indicators in step 3.

2. The indicators include five categories: edge betweenness, average path length change, edge connectivity, edge vulnerability, and network efficiency change after edge removal. An edge removal sequence is constructed according to the numerical values ​​of these five categories of indicators from largest to smallest, and the corresponding edges in the weighted directed disaster chain network are removed in turn. After each removal step, two global network indicators are calculated. The removal order in which the maximum connected subgraph ratio decreases faster and the number of connected subgraphs increases faster represents the strategy that can destroy the weighted directed disaster chain network. Ultimately, the entire process of automatically identifying disaster event types and their causal relationships from multi-source text data, constructing a weighted directed disaster chain network, and evaluating blocking strategies will be realized.

Images