A method for evaluating safety of groundwater under extreme weather events

Abstract

This invention relates to a method for groundwater safety assessment under extreme climate events, comprising: collecting heterogeneous data from multiple sources such as meteorology, geology, hydrology, and remote sensing; standardizing, cleaning, semantically aligning, and completing missing data to construct a unified groundwater safety knowledge graph; monitoring extreme climate events in real time; updating node relationship weights and sparsifying transmission paths based on the dynamic evolution of the knowledge graph and a temporal attention mechanism; combining an improved graph neural network to infer risk propagation paths from the dynamic knowledge graph, identifying key pollution nodes, and outputting structured risk levels and response suggestions; the system continuously optimizes the graph and model parameters based on evolutionary feedback to achieve high adaptability and inference accuracy in emergency response. This method improves the intelligence, real-time performance, and accuracy of groundwater risk assessment, and solves the problems of difficulty in dynamically identifying groundwater pollution propagation paths and insufficient adaptability in decision-making under extreme climate events.

Description

Technical Field

[0001] This invention relates to the field of groundwater safety knowledge graph modeling and dynamic graph evolution reasoning analysis technology, and in particular to a groundwater safety assessment method under extreme climate events. Background Technology

[0002] Currently, with the frequent occurrence of extreme weather events, such as torrential rains, typhoons, prolonged high temperatures, or droughts, the safety of global groundwater systems is facing unprecedented and complex threats. Dynamic monitoring of groundwater safety and scientific analysis of risk propagation pathways have become research hotspots in water resource management, urban and rural disaster prevention and mitigation, and ecological restoration. The main problems and shortcomings of existing technologies are as follows:

[0003] (1) Insufficient dynamic adaptability: In the face of rapid and real-time changes in hydrogeological conditions during extreme climate events, existing prediction models are difficult to adjust in a timely manner, and the identification of risk propagation paths is significantly delayed, which cannot support the real-time decision-making needs in high-frequency dynamic scenarios.

[0004] (2) The system relies heavily on full reconstruction or manual intervention, and incremental (local) dynamic updates are inflexible. When extreme weather events occur, the real-time response capability of the prediction model is significantly insufficient.

[0005] (3) Under extreme climate conditions, the driving force of the temporal changes of different meteorological factors on the groundwater risk transmission path is difficult to be incorporated into the inference model in a timely manner. Traditional inference paths are often based on preset weights and fixed relationships, while ignoring the real-time and dynamic evolution process of the "meteorology-geology-hydrology" ternary linkage, which affects the timeliness and reliability of the inference results. Summary of the Invention

[0006] This application provides a groundwater safety assessment method under extreme climate events, aiming to solve one of the problems or issues of the prior art mentioned in the background section.

[0007] This application provides a groundwater safety assessment method under extreme climate events, specifically including:

[0008] S1: Collect multi-source heterogeneous data from meteorology, geology, hydrology, and remote sensing as the initial input data source for the groundwater safety assessment system.

[0009] S2: Standardize, clean, and semantically align the collected multi-source data to eliminate data format differences and redundant information interference, so as to generate a dataset with unified semantic representation.

[0010] S3: Construct an initial groundwater safety knowledge graph based on the preprocessed dataset, define the semantic structure of entity nodes and relation edges, and establish a graph ontology model.

[0011] S4: Deploy an extreme climate event monitoring module to receive meteorological early warning data and hydrological anomaly indicators in real time, and determine whether the conditions for dynamic evolution of the map are met.

[0012] S5: If it is determined that an extreme climate event has occurred or is about to occur, the graph dynamic evolution module is activated to perform local structural updates on key nodes and relation edges in the knowledge graph based on real-time incremental data.

[0013] S6: In the dynamic evolution of the map, a time-series attention mechanism is introduced to calculate the time-varying impact weight of climate change factors on the risk propagation path between nodes.

[0014] S7: Input the updated dynamic knowledge graph into the graph neural network model to perform reasoning analysis on risk propagation paths and identify potential groundwater pollution diffusion paths and key impact nodes.

[0015] S8: Output the risk propagation path results obtained from the reasoning to the risk decision-making module, and generate a structured risk level and response recommendation report.

[0016] S9: After each graph evolution and inference is completed, record the differences in graph structure and changes in inference results before and after the evolution, which will be used for subsequent model iteration optimization and semantic consistency verification.

[0017] S10: Based on evolutionary history data and feedback results, dynamically adjust the map update strategy and inference model parameters to improve the adaptability and accuracy of groundwater risk propagation path identification under extreme climate events.

[0018] This application provides a groundwater safety assessment method under extreme climate events, which has the following beneficial effects:

[0019] (1) This invention introduces for the first time a dynamic knowledge graph evolution mechanism triggered by extreme climate events. Through incremental data acquisition, standardized fusion of heterogeneous multi-source data, and dynamic ontology structure adjustment, it achieves rapid perception and response of the knowledge graph to extreme climate change. The system can activate the graph evolution and risk propagation path identification process within minutes based on meteorological warnings and hydrological anomaly signals, effectively solving the shortcomings of traditional methods such as the need for periodic graph reconstruction and response lag, and ensuring the real-time and timely nature of risk path reasoning.

[0020] (2) This invention achieves groundbreaking spatiotemporal alignment and multimodal feature fusion of heterogeneous data from various types, including meteorology, geology, hydrology, and remote sensing. Through principal component analysis, automatic detection and completion of missing and outlier values, and semantic alignment, the completeness, standardization, and consistency of the input data are ensured. In numerous real-world scenarios, the spatiotemporal consistency index of the data is significantly improved, and the accuracy of data integrity and risk propagation path identification is also significantly enhanced.

[0021] (3) The innovative introduction of incremental triplet structured update and temporal attention mechanism not only tracks the impact of extreme climate on nodes and relation edges in real time, but also drives the dynamic adjustment and sparsification of relation edge weights through attention weights, thereby improving the model’s sensitivity and discrimination ability to sudden environmental changes.

[0022] In summary, this invention, with dynamic graph evolution, temporal attention weight allocation, incremental data fusion, and adaptive optimization as its core, achieves highly timely, accurate, and robust identification and response to groundwater risk propagation paths under extreme climate events. It completely breaks through the bottlenecks of existing static reasoning and slow updates, demonstrating significant technological leadership and broad practical application value, and is of great significance for improving national and regional water security under extreme environments. Attached Figure Description

[0023] Figure 1 This is the main flowchart of a groundwater safety assessment method under extreme climate events proposed in this application.

[0024] Figure 2 This is a sub-flowchart of a groundwater safety assessment method under extreme climate events proposed in this application.

[0025] Figure 3 This is another sub-flowchart of a groundwater safety assessment method under extreme climate events proposed in this application. Detailed Implementation

[0026] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0027] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0028] like Figure 1 As shown, this application provides a groundwater safety assessment method under extreme climate events, specifically including:

[0029] S1: Collect multi-source heterogeneous data from meteorology, geology, hydrology, and remote sensing as the initial input data source for the groundwater safety assessment system.

[0030] S2: Standardize, clean, and semantically align the collected multi-source data to eliminate data format differences and redundant information interference, so as to generate a dataset with unified semantic representation.

[0031] S3: Construct an initial groundwater safety knowledge graph based on the preprocessed dataset, define the semantic structure of entity nodes and relation edges, and establish a graph ontology model.

[0032] S4: Deploy an extreme climate event monitoring module to receive meteorological early warning data and hydrological anomaly indicators in real time, and determine whether the conditions for dynamic evolution of the map are met.

[0033] S5: If it is determined that an extreme climate event has occurred or is about to occur, the graph dynamic evolution module is activated to perform local structural updates on key nodes and relation edges in the knowledge graph based on real-time incremental data.

[0034] S6: In the dynamic evolution of the map, a time-series attention mechanism is introduced to calculate the time-varying impact weight of climate change factors on the risk propagation path between nodes.

[0035] S7: Input the updated dynamic knowledge graph into the graph neural network model to perform reasoning analysis on risk propagation paths and identify potential groundwater pollution diffusion paths and key impact nodes.

[0036] S8: Output the risk propagation path results obtained from the reasoning to the risk decision-making module, and generate a structured risk level and response recommendation report.

[0037] S9: After each graph evolution and inference is completed, record the differences in graph structure and changes in inference results before and after the evolution, which will be used for subsequent model iteration optimization and semantic consistency verification.

[0038] S10: Based on evolutionary history data and feedback results, dynamically adjust the map update strategy and inference model parameters to improve the adaptability and accuracy of groundwater risk propagation path identification under extreme climate events.

[0039] Step S1: Collect multi-source heterogeneous data from meteorology, geology, hydrology, and remote sensing as the initial input data source for the groundwater safety assessment system. Specifically, this includes:

[0040] S1.1: Obtain extreme climate event prediction data from meteorological departments, including time-series meteorological parameters such as precipitation intensity, wind speed, and temperature change rate, to construct the initial input dataset for extreme climate events.

[0041] For the real-time data interface of the meteorological department, a distributed data acquisition method is adopted (parameters: interface refresh cycle of 5 minutes, data transmission protocol is HTTPS) to realize the batch capture and caching of prediction data of extreme climate events such as precipitation intensity, wind speed, and temperature change rate.

[0042] Furthermore, a time series denoising algorithm (parameters: wavelet threshold selection based on the soft threshold principle, decomposition layer 4) is used to filter out noise from the original meteorological parameter time series curves and obtain smoothed precipitation intensity, wind speed and temperature change rate data sequences.

[0043] Furthermore, a multivariate normalization method is adopted (parameter: Min-Max normalization range is set to 0-1) to achieve dimensionless processing of meteorological parameters with different dimensions and generate time series feature vectors of extreme climate events with a unified scale.

[0044] Furthermore, an event labeling method is adopted to combine the weights of various meteorological elements with the prediction parameters to generate an initial input dataset label matrix for extreme climate events, including precipitation intensity, wind speed, and temperature change rate, for use by subsequent multi-source data fusion and knowledge graph construction modules.

[0045] Through the aforementioned time-series data acquisition, denoising, normalization, weight calculation, and label construction, meteorological forecast data is transformed into a structured and semantically consistent initial input feature set for extreme climate events, thereby establishing the data foundation for subsequent groundwater risk propagation path analysis.

[0046] For example, in a typhoon-triggered monitoring scenario, the meteorological department's interface refresh cycle is 5 minutes, and the typhoon forecast time series includes precipitation intensity (unit: mm / h), wind speed (unit: m / s), and temperature change rate (unit: ℃ / h).

[0047] The Daubechies-6 wavelet basis was used for four-level decomposition and denoising, reducing the noise variance of the precipitation intensity sequence to 15% of the original, the peak error of the wind speed sequence to 10%, and the root mean square error of the temperature change rate curve to 8%. Min-Max normalization (range 0-1) was performed on the three types of parameters. The feature vectors within the 8-hour prediction window were input into the weight calculation model, resulting in precipitation intensity weight w1 = 0.45, wind speed weight w2 = 0.35, and temperature change rate weight w3 = 0.20, forming a label matrix for subsequent use by the atlas module. In this scenario, after combining meteorological features with hydrological anomaly indicators, the accuracy of extreme weather event risk prediction improved to 92% on the historical validation set, meeting the technical requirements for real-time response to extreme weather events.

[0048] S1.2: Collect geological structure maps, lithological distribution data and soil permeability coefficients provided by geological survey institutions, perform spatial coordinate alignment and stratigraphic attribute coding based on the Geographic Information System (GIS) platform, and generate a standardized geological feature dataset.

[0049] The system receives geological structure maps, lithological distribution data, and soil permeability coefficients provided by geological survey institutions as input conditions, and loads them into the raw data management module of the Geographic Information System (GIS) platform to achieve unified management of geological data and spatial reference systems.

[0050] A coordinate projection transformation method (parameters: source coordinate system, target coordinate system, ellipsoidal datum) is used to align the geological structure map with the unified spatial reference of the system and output the vector dataset after coordinate transformation.

[0051] Furthermore, by using a spatial topology matching algorithm (parameter: tolerance distance 0.5 meters), the lithological distribution data and the geological structure map layer boundaries are matched, and a spatially consistent lithological layer is obtained.

[0052] Furthermore, an attribute encoding mapping method (parameters: field standardization dictionary, ontology attribute library) is adopted to achieve standardized encoding of soil permeability coefficient data and GIS platform layer attribute fields, and to generate an attribute matrix that can directly drive knowledge graph modeling.

[0053] Furthermore, by using spatial interpolation (Kriging variogram parameter range: exponential model, semivariogram threshold 0.05), the soil permeability coefficient in the unsampled area was estimated, resulting in raster data of soil permeability coefficient covering the entire study area.

[0054] By using data fusion and structuring algorithms, the coordinate-aligned vector data, boundary-matched lithology layers, and standardized soil permeability coefficient matrices are integrated into a standardized geological feature dataset in a unified format, enabling the direct usability of basic geological data in subsequent knowledge graph construction modules.

[0055] S1.3: Receive real-time monitoring data such as groundwater level, flow velocity, and water quality indicators from hydrological monitoring stations, and extract hydrological time series features using the sliding window method to form dynamic hydrological state characterization data.

[0056] S1.4: Acquire multispectral satellite imagery and synthetic aperture radar (SAR) imagery data through the remote sensing data interface, perform image preprocessing and land cover type classification, and generate remote sensing assessment layers for surface water distribution and surface infiltration capacity.

[0057] S1.5: Based on the data fusion engine, the four types of heterogeneous data, namely meteorological, geological, hydrological and remote sensing data, are spatiotemporally aligned and multi-source data are fused to generate a multi-dimensional groundwater safety assessment data cube under a unified spatiotemporal benchmark, which serves as the input data source for the knowledge graph construction module.

[0058] Based on the data fusion engine, spatiotemporal alignment processing is performed on four types of heterogeneous data: meteorological, geological, hydrological and remote sensing data. A multi-channel spatiotemporal registration algorithm (parameters: spatial resolution Δr = 30m, time step Δt = 1h) is used to map the spatial coordinate system of each type of data to the global geographic reference system WGS84, so as to achieve consistency of location reference.

[0059] Furthermore, a distributed time interpolation algorithm (parameters: interpolation kernel function is Gaussian weight kernel, time window length is 6h) is used to synchronize the data collected at different times, aligning the collection time of each data source to a unified timestamp t0, thus obtaining a synchronized time series data matrix.

[0060] Furthermore, a multimodal feature fusion method is adopted (parameters: the fusion strategy is weighted feature splicing, the weight coefficients αmeteorology = 0.25, αgeology = 0.25, αhydrology = 0.3, αremote sensing = 0.2) to achieve cross-source fusion of meteorological, geological, hydrological and remote sensing features, and generate a high-dimensional multimodal feature vector set F with a vector dimension of n = 1024.

[0061] Furthermore, the high-dimensional multimodal feature vector set F is reduced in dimensionality using principal component analysis (PCA) algorithm (parameter: retain cumulative variance contribution rate ≥ 95%) to obtain a low-dimensional feature matrix F′. After dimensionality reduction, the dimension m is 256, and the main information is not lost.

[0062] Furthermore, a multidimensional groundwater safety assessment data cube D (parameters: spatial grid size of 30m×30m, time slice interval of 1h, attribute dimension of 256) is constructed using a three-dimensional raster reconstruction method. Each cell of D contains a fusion feature vector F′ under a unified spatiotemporal benchmark, thus completing the generation of the input data carrier for the knowledge graph construction module.

[0063] By using a multi-channel spatiotemporal registration algorithm and a multi-modal feature fusion method, the heterogeneous data collected in the previous step is transformed into a data cube with a unified spatiotemporal benchmark and semantic consistency, thereby achieving high-quality data support for knowledge graph construction.

[0064] For example, in an extreme rainstorm scenario, meteorological data is sampled at a frequency of 1 hour with a spatial resolution of approximately 5 km; geological data has a spatial resolution of 30 m and is collected intermittently; hydrological data is sampled at a frequency of 30 minutes, located at riverside stations; remote sensing data has a spatial resolution of 10 m; multispectral imagery is acquired every 12 hours; and SAR imagery is acquired every 6 hours. The data fusion engine first interpolates the meteorological data to a 30 m spatial resolution and then uses a Gaussian weight kernel to interpolate to a unified timestamp t0 within a 6-hour window. Geological data is converted to WGS84 using a GIS platform and stratigraphic attribute encoding is performed. Hydrological data is aggregated to 1-hour intervals using a sliding window method and spatially resampled to a unified grid. Remote sensing data undergoes image preprocessing and classification to generate land cover and permeability layers, which are then spatially aligned to a unified grid. The four types of data were weighted and concatenated according to set weight coefficients to generate a 1024-dimensional fused feature vector. This vector was then reduced to 256 dimensions using PCA and rasterized into a 3D data cube with a 30m×30m grid and a time step of 1 hour, filling each grid cell with 256-dimensional features. Validation results show that the fused cube achieved a spatiotemporal consistency of over 99% and data integrity of 98% during the rainstorm event, providing high-quality input for the knowledge graph construction module.

[0065] Step S2: Standardize, clean, and semantically align the collected multi-source data to eliminate data format differences and redundant information interference, thereby generating a dataset with unified semantic representation. Specifically, this includes:

[0066] S2.1: Standardize the format of heterogeneous data from multiple sources, such as meteorology, geology, hydrology, and remote sensing, to unify the storage structure and field naming rules of the original data.

[0067] For multi-source heterogeneous data such as meteorological, geological, hydrological and remote sensing data, a multi-format parsing engine (parameters: supports CSV, NetCDF, Shapefile, GeoTIFF, HDF5) is used to realize the content parsing and structured extraction of raw data files.

[0068] Furthermore, through the field mapping rule base (parameters: field alias, unit system, symbol standardization), a unified mapping of various data field names is achieved, and intermediate data with a unified field naming system is obtained.

[0069] Furthermore, a data pattern alignment algorithm (parameters: column order template, data type inference rules) is adopted to unify the table structure of meteorological, geological, hydrological and remote sensing data, and generate a cross-source consistent column structure pattern description file.

[0070] Furthermore, by utilizing the unit conversion and dimensional normalization module (parameters: supports conversion between temperature units ℃ and K, length units m and km, and time units s and h), a unified dimensional representation of numerical fields can be achieved.

[0071] Furthermore, based on the coding specification generator (parameters: geographic coordinate system EPSG code, timestamp format YYYY-MM-DD hh:mm:ss), the spatial location and time information fields are converted into the system's standard coding form to support spatiotemporal retrieval and fusion.

[0072] Through the above-mentioned standardized transformation chain, the data parsing results of the previous step are transformed into a standardized multi-source dataset under a unified storage structure and field naming rules, thereby achieving the expected technical effect of eliminating format differences.

[0073] S2.2: Based on the data quality assessment model, missing value detection and outlier identification are performed on the standardized data to generate a groundwater-related dataset with preliminary integrity.

[0074] For datasets that have undergone format standardization, a data quality assessment model (parameters: the set of indicators includes completeness rate, consistency rate, and anomaly frequency) is used to quantitatively analyze the missing status of different types of groundwater-related data records.

[0075] Furthermore, by using a missing value detection algorithm (parameters: based on σ multiple standard deviation threshold to determine missing anomalies, time series window length 7 days), missing value marking is achieved for meteorological time series, geological attribute sets, hydrological dynamic state values ​​and remote sensing raster data, and a missing matrix dataset is obtained.

[0076] Furthermore, a multidimensional outlier identification algorithm (parameters: Local Outlier Factor (LOF), number of neighborhood points 15, outlier threshold 1.5) is adopted to detect anomalies in spatial location, attribute values, and time series trends of multi-source data, and generate an anomaly sample index vector.

[0077] Furthermore, based on the statistical consistency assessment method (parameters: Pearson correlation coefficient and standard score deviation joint criterion), cross-consistency tests of similar variables from different data sources are performed, and a consistency deviation score table is generated.

[0078] By using the missing matrix, outlier index, and consistency deviation score output by the data quality assessment model, the standardized data from the previous step is transformed into a preliminary complete dataset related to groundwater with missing labeling, outlier identification results, and consistency score, thus ensuring the quality of subsequent completion and semantic alignment processing.

[0079] For example, in the data quality assessment before and after an extreme rainfall event, the meteorological dataset contained three main variables: precipitation, wind speed, and temperature. Precipitation was missing for two consecutive hours, wind speed had a single point exceeding historical extremes, and the temperature curve showed a sudden drop of -12°C within a day. Geological data had five missing records in the lithology coding field, hydrological data showed an anomaly of a 2-meter instantaneous rise in groundwater level at observation well G3, and remote sensing data showed reflectance values ​​below the normal range in some raster areas. The missing value detection algorithm, using a 7-day time window and multiple standard deviations, gave a missing value matrix labeling rate of 2.7% for meteorological variables. The LOF anomaly detection model identified 38 anomalous sample indices in wind speed, groundwater level, and remote sensing reflectance with 15 neighborhood points. In the consistency test, variables with a Pearson correlation coefficient less than 0.3 were marked as low consistency, and records with a standard score offset greater than 2 were marked as high-biased samples. The resulting preliminary complete dataset provides definite missing locations, anomaly localization, and consistency scores for subsequent interpolation completion and semantic mapping, thus achieving quality controllability in the data processing chain.

[0080] S2.3: A data completion method based on statistical analysis and interpolation algorithms is used to repair missing data records in order to improve the integrity and usability of the dataset.

[0081] S2.4: Perform semantic alignment processing on the cleaned multi-source data, and perform normalized mapping on entity names, attribute descriptions and relation expressions based on the domain ontology library to eliminate semantic ambiguity.

[0082] S2.5: Based on knowledge embedding technology, the semantically aligned data is vectorized to generate entity and relation embedding vectors in a unified semantic space to support the subsequent construction and reasoning analysis of knowledge graphs.

[0083] Step S3: Construct an initial groundwater safety knowledge graph based on the preprocessed dataset, define the semantic structure of entity nodes and relation edges, and establish a graph ontology model. For example... Figure 2 As shown, it specifically includes:

[0084] S3.1: Perform entity recognition and classification processing on the structured fields in the preprocessed multi-source heterogeneous dataset to extract geological entities, hydrological entities, meteorological entities and remote sensing entities involved in groundwater safety assessment, and generate an entity type set.

[0085] For the structured fields of preprocessed meteorological, geological, hydrological and remote sensing multi-source heterogeneous datasets, a conditional random field (CRF) sequence labeling model (parameters: feature template set, window size of 3, regularization coefficient of 0.1) is used to achieve preliminary automatic identification of various groundwater safety-related entities.

[0086] Furthermore, by using a domain keyword dictionary and rule matching algorithm (parameters: 5000 entity words in the groundwater safety domain, matching rules covering prefixes, suffixes, and unit identification), the category labeling of identified entities is achieved, resulting in four types of labeled output results: geological entities (aquifers, lithology, faults, etc.), hydrological entities (groundwater level, flow velocity, water quality index, etc.), meteorological entities (precipitation, temperature gradient, wind speed, etc.), and remote sensing entities (land cover type, spectral reflectance, SAR interferometric parameters, etc.).

[0087] Furthermore, by using the entity attribute feature vector calculation method (parameters: entity feature dimension 64, feature extraction based on statistics and time series patterns), the quantitative encoding of differentiated attributes within the same type of entity is realized, and an entity feature set that is convenient for subsequent relationship extraction is generated.

[0088] Furthermore, by using dimensionality reduction based on principal component analysis (PCA) (parameters: cumulative contribution rate threshold of 95%, number of features retained not exceeding 20), the entity feature set is compressed to improve the computational efficiency of subsequent map construction, while retaining the main information components.

[0089] By using entity classification and feature encoding, the above identification and encoding results are transformed into a set of entity types, thereby forming the initial semantic structure of the entity nodes in the groundwater safety knowledge graph.

[0090] For example, on a mixed dataset containing meteorological (time series of precipitation records from hydrological stations), geological (fault locations and soil permeability coefficients), hydrological (groundwater level change curves), and remote sensing (multispectral and SAR imagery) data, a conditional random field model was used for sequence labeling. The window size was set to 3, the regularization coefficient was set to 0.1, and the domain entity lexicon contained 5000 geological, hydrological, meteorological, and remote sensing-specific terms. The model output identified 120 geological entities (e.g., sandstone layer, fault F1), 95 hydrological entities (e.g., water level in monitoring well M12), 80 meteorological entities (e.g., rainstorm unit P5), and 110 remote sensing entities (e.g., NDVI value region R34). For these entities, an attribute feature vector encoding method was used to represent each entity as a 64-dimensional vector, containing statistical features and temporal pattern features. The PCA method was used to reduce the dimensionality of the encoded vectors, with a cumulative contribution rate threshold set to 95%, ultimately compressing each entity into a 15-dimensional feature space. During the verification process, the entity recognition and classification process achieved an accuracy of 96% and a recall rate of 94%, providing a high-quality set of entity types for subsequent relation extraction and graph ontology modeling.

[0091] S3.2: Based on the entity type set, the named entity disambiguation algorithm is used to perform semantic deduplication and normalization on entities of the same type, so as to eliminate the name differences of the same entity in multi-source data and generate a standardized entity name library.

[0092] S3.3: Perform relation extraction on entities in the standardized entity name library, use dependency parsing and rule matching methods to identify semantic relationships between entities, and generate a preliminary set of relation triples.

[0093] S3.4: Based on the preliminary set of relational triples, combined with the knowledge base of experts in the field of groundwater safety, perform relational semantic verification and classification processing to establish a relational semantic model that conforms to the logic of groundwater risk propagation.

[0094] S3.5: Perform ontology modeling on the entity and relation semantic model, use the OWL ontology language to define the classes, attributes and constraint rules of the graph, and construct the ontology structure framework of the groundwater safety knowledge graph.

[0095] The expansion of S3.5 is as follows:

[0096] Based on the established entity type set and relational semantic model, the OWL ontology language modeling method (parameters: class hierarchy rule set, attribute domain range constraint, relational semantic consistency rule) is used to realize the ontology structure framework definition of the groundwater safety knowledge graph.

[0097] By defining classes and using hierarchical organization methods (parameters: parent-child relationships of geological, hydrological, meteorological, and remote sensing classes), a domain class tree is constructed, and a set of unique identifier URIs for each class is obtained.

[0098] Furthermore, through data attribute and object attribute modeling methods (parameters: attribute data type constraints, object attribute association direction rules), a formal expression of entity node feature attributes and relationships between nodes is realized, and attribute constraint description documents are generated.

[0099] Furthermore, through the semantic constraint encoding method (parameters: consistency rule expression, constraint priority setting), the core semantic constraints in the field of groundwater safety are realized, including the association conditions between geological entities and hydrological entities, the influence relationship between meteorological entities and geological conditions, etc., and a set of rules that can be called by the reasoning module is generated.

[0100] Furthermore, a logical reasoning-supported structure generation algorithm (parameters: class inheritance relationship, attribute passing rules) is adopted to form an ontology framework structure with semantic inference capabilities, which explicitly expresses domain knowledge and embeds it into the graph reasoning engine.

[0101] By using the OWL ontology language definition and semantic constraint processing, the entity and relation semantic model from the previous step is transformed into an ontology structure framework that conforms to the logic of groundwater safety risk propagation, thus achieving the uniformity and reasonability of graph semantic modeling.

[0102] For example, in the practical application of constructing a knowledge graph ontology framework for groundwater security, the following categories are selected: Geology (including "lithology type" and "strata structure"), Hydrology (including "groundwater level" and "flow velocity"), Meteorology (including "precipitation intensity" and "temperature change"), and Remote Sensing (including "land cover type"). Each category is assigned a unique URI, such as http: / / kg / geology#Lithology for the Geology category and http: / / kg / hydro#WaterLevel for the Hydrology category. The object attribute "hasImpactOn" is defined to represent the impact of meteorological objects on hydrological objects, with its domain being the meteorological category and its scope being the hydrological category. The data attribute "permeabilityCoeff" represents the permeability coefficient of the lithology type, with a data type of double. Semantic constraint rules are set: when the precipitation intensity of the meteorological category is >50 mm / h and the temperature change is <5 degrees / h, the risk of triggering the "groundwater level" attribute of the hydrological category is marked as high. By using a semantic reasoning engine to load and verify the above rules, the consistency of object attributes and data attributes within the ontology framework is ensured, thereby enabling the structured representation and logical inference capability of groundwater risk under extreme climate conditions.

[0103] S3.6: Based on the ontology structure framework, standardized entities and semantic relationships are mapped to the graph database model, and graph instantiation processing is performed to generate the storage structure of the initial groundwater safety knowledge graph.

[0104] S3.7: Perform semantic consistency verification on the initial groundwater safety knowledge graph. Use a graph traversal algorithm to detect semantic conflicts and redundant paths of entities and relationships in the graph, so as to improve the logical integrity and reasoning usability of the graph.

[0105] Step S4: Deploy an extreme climate event monitoring module to receive meteorological early warning data and hydrological anomaly indicators in real time, and determine whether the conditions for dynamic evolution of the weather map are met. Figure 3 As shown, it specifically includes:

[0106] S4.1: Obtain extreme weather event warning data issued by meteorological departments, and perform semantic analysis on the warning information based on preset event classification standards to identify the current event type and intensity level.

[0107] S4.2: Collect abnormal index data such as groundwater level, flow velocity, and water quality parameters provided by real-time hydrological monitoring stations, and use the sliding window statistical method to calculate the fluctuation range of the deviation from the baseline value in order to assess the degree of abnormality of the current hydrological state.

[0108] S4.3: Based on the types and intensity levels of extreme climate events after semantic parsing, and combined with the amplitude of hydrological anomalies, construct a multi-dimensional event feature vector to quantify the comprehensive risk index of the current environmental state.

[0109] Based on the extreme climate event type and intensity level data obtained from semantic parsing, a feature encoding method (parameters: event type encoding table, intensity level numerical mapping rules) is used to convert meteorological event information into a computable numerical representation.

[0110] Furthermore, by using a fluctuation amplitude calculation algorithm (parameters: sliding window size 24 hours, baseline hydrological state vector), the fluctuation amplitude of hydrological anomalies is transformed into a standardized anomaly index, and the quantitative characteristics of hydrological state are obtained.

[0111] Furthermore, a multi-dimensional vector concatenation method (parameters: meteorological event type coding value, hydrological anomaly index, and geological permeability sensitivity coefficient) is adopted to construct an initial event feature vector set and generate a comprehensive feature matrix containing meteorological, hydrological, and geological multi-source elements.

[0112] Furthermore, by using a normalization algorithm (parameters: minimum-maximum normalization range 0,1), the differences in feature dimensions across dimensions are eliminated, and a multi-dimensional event feature vector with a consistent range is generated.

[0113] Furthermore, a weighted comprehensive calculation method (parameters: meteorological weight 0.5, hydrological weight 0.3, geological weight 0.2) is adopted to quantify the comprehensive risk index R of the current environmental state. The calculation formula is as follows:

[0114] R = 0.5E + 0.3H + 0.2G

[0115] Where E is the numerical code for meteorological events, H is the hydrological anomaly index, and G is the geological permeability sensitivity coefficient.

[0116] By fusing multidimensional features and weighting calculations, the coding and fluctuation amplitude analysis results from the previous step are transformed into a single comprehensive risk index, enabling a quantitative assessment of the current environmental state.

[0117] For example, for an extreme rainstorm event with a precipitation intensity level of 6, the meteorological event type code is set to 6, the hydrological monitoring fluctuation range is 150% of the baseline value, and after standardization, the hydrological anomaly index is 0.85. The geological permeability sensitivity coefficient is set to 0.65 based on the regional lithology of sandy loam. Using a given weight configuration, the comprehensive risk index is calculated as follows:

[0118] R = 0.56 + 0.3 + 0.85 + 0.2 + 0.65

[0119] The result is 3.0, and the risk index is higher than the preset trigger threshold of 2.5. Therefore, the dynamic evolution trigger process of the spectrum is initiated, which effectively improves the system's response accuracy and speed under extreme rainstorm conditions.

[0120] S4.4: Input the constructed multidimensional event feature vector into the pre-trained event triggering classification model. This model is trained based on the support vector machine algorithm and outputs a binary judgment result on whether the current dynamic evolution triggering condition of the knowledge graph is met.

[0121] Based on the constructed multidimensional event feature vector, a support vector machine algorithm (kernel function type: radial basis function kernel RBF, penalty coefficient C ranging from 0.1 to 100, kernel width parameter γ adaptively optimized according to the distribution characteristics of the training set) is used to realize the function of judging whether the current environmental state meets the triggering conditions of dynamic evolution of the map.

[0122] Furthermore, by using a feature vector normalization method (scaling the range to the [0,1] interval), the scale of event features in each dimension is unified in the input space, and a normalized event feature matrix is ​​obtained.

[0123] The execution of subsequent map evolution modules is driven by binary label output, enabling automatic detection and signal generation of extreme climate event triggering conditions.

[0124] S4.5: If the judgment result is that the triggering condition is met, a dynamic evolution activation signal for the graph is generated, and the event feature vector and timestamp information are encapsulated as input parameters for the evolution module to start the subsequent graph update process.

[0125] Step S5: If it is determined that an extreme climate event has occurred or is about to occur, the knowledge graph dynamic evolution module is activated to perform local structural updates on key nodes and relation edges in the knowledge graph based on real-time incremental data. Specifically, this includes:

[0126] S5.1: Based on the event trigger signal output by the extreme climate event monitoring module, initiate the dynamic evolution process of the knowledge graph to determine whether an incremental update operation needs to be performed on the knowledge graph.

[0127] S5.2: Perform format parsing and semantic mapping on real-time incremental data from meteorological early warning, hydrological monitoring and remote sensing to generate structured triplet data representations compatible with knowledge graph ontology models.

[0128] For real-time incremental data from meteorological early warning, hydrological monitoring and remote sensing, a multi-source data parsing engine is used to achieve unified parsing and field mapping of heterogeneous data formats.

[0129] Furthermore, through a field extraction algorithm based on regular expression matching and data descriptors (parameter: matching pattern library = pattern set based on domain ontology), the semantic association binding between the original data fields and the attributes of the knowledge graph ontology model is realized, and a preliminary semantically labeled data vector is obtained.

[0130] Furthermore, a multi-dimensional spatiotemporal mapping algorithm (parameters: time window = real-time event trigger window, spatial reference system = GIS standard coordinate system) is adopted to realize the synchronization of timestamps and the standardization of geographic coordinates of various types of data, and to generate a candidate set of entities with a unified spatiotemporal reference.

[0131] Furthermore, the entity relation pattern generator (parameter: relation extraction rule set = dependency syntax rules + domain-specific pattern) is used to transform the elements in the entity candidate set into structured triples <entity1, relation, entity2> according to semantic relations, where the triple relation type strictly follows the ontology structure constraints constructed in the previous stage.

[0132] Furthermore, a semantic consistency verification algorithm (parameters: verification rule = ontology constraint clause + semantic conflict detection rule) is adopted to detect the compatibility of the triple set with the existing knowledge graph ontology model and remove triples that do not conform to the semantic or relational patterns.

[0133] Through the above processing method, the parsed real-time incremental data is transformed into a set of structured triples that are fully compatible with the knowledge graph ontology model, thereby achieving consistency of input data and structural stability of the dynamic evolution module.

[0134] For example, under the triggering of an extreme rainstorm event, the meteorological early warning system outputs JSON data showing a precipitation intensity of 85 mm / h, a wind speed of 20 m / s, and a temperature drop rate of 0.5 degrees Celsius / h; the hydrological monitoring station sends XML data showing a groundwater level rise of 1.2 meters, a flow velocity increase of 0.8 m / s, and a water turbidity increase of 15 NTU; and multispectral imagery transmitted by remote sensing satellites, after preprocessing, identifies a newly added surface runoff area of ​​2.5 square kilometers. The parsing engine maps the meteorological data field "precipitation intensity" to the ontology attribute "Precipitation_Intensity," the timestamp to a unified and coordinated time, and the spatial coordinates are aligned to a latitude and longitude accuracy of 0.0001 degrees using the GIS reference system. The relational schema generator generates a relational triple <rainstorm event, leading to, groundwater level rise> for "rainstorm event" and "groundwater level rise," and a triple <increased surface runoff, increased, pollutant diffusion risk> for "increased surface runoff." Finally, the semantic verification algorithm removes conflicting triples such as <precipitation anomaly, equal to, wind speed anomaly>, and outputs a set of structured triples as input to the dynamic evolution module of the graph, ensuring the semantic consistency and accuracy of subsequent inference analysis.

[0135] S5.3: Based on incremental triplet data, perform graph entity recognition and relation extraction tasks. Utilize graph embedding technology to calculate the semantic similarity between new entities and existing knowledge graph nodes to determine whether they belong to existing entities or need to be added as entity nodes.

[0136] S5.4: Perform incremental graph structure update operations on the newly matched triples, including adding nodes, expanding edges, or updating node attributes, to maintain the semantic consistency and topological stability of the knowledge graph.

[0137] S5.5: After the graph structure is updated, perform graph consistency verification and conflict resolution. Based on rule-based reasoning and semantic constraint mechanisms, detect and correct semantic conflicts or logical errors introduced by incremental updates to ensure the semantic integrity and reasoning effectiveness of the knowledge graph.

[0138] Step S6: During the dynamic evolution of the map, a temporal attention mechanism is introduced to calculate the time-varying impact weights of climate change factors on the risk propagation paths between nodes. Specifically, this includes:

[0139] S6.1: Normalize the real-time climate change factor data collected during the map evolution process to eliminate the dimensional differences between meteorological data of different dimensions and obtain standardized climate change time series input.

[0140] S6.2: Based on the normalized climate change time series input, construct a multivariate time series feature vector to capture the dynamic evolution characteristics of extreme climate events such as sudden temperature drops, abnormal precipitation, and evaporation fluctuations.

[0141] S6.3: Input the multivariate time-series feature vector into the designed time-series attention mechanism model, calculate the attention weight distribution of the impact of each time node on the risk propagation path, so as to identify the time-varying driving effect of key climate change factors on the groundwater risk propagation path.

[0142] S6.4: Based on the attention weight distribution, the relationship edges between nodes in the knowledge graph are dynamically weighted to generate a time-sensitive dynamic risk propagation weight graph structure.

[0143] Based on the attention weight distribution calculated by the temporal attention mechanism, a dynamic weighting processing algorithm for relation edge weights (parameters: node pair identifier set, initial relation weight matrix, time weight vector) is adopted to realize the dynamic weighting processing of relation edges between nodes in the knowledge graph.

[0144] Furthermore, by using a weighted coefficient fusion method (parameters: original edge weights, attention weight normalization coefficients), the edge weights are proportionally adjusted according to time sensitivity to obtain an updated edge weight matrix that includes the time-varying characteristics of risk propagation.

[0145] Furthermore, a temporal weighted mapping function is used to map the updated edge weight matrix to the relation edge attribute field in the graph storage structure, and to generate a dynamic risk propagation weight graph structure that conforms to the ontology model constraints.

[0146] Furthermore, the instantaneous volatility of edge weights is reduced by using a weight smoothing algorithm (parameters: time window length, smoothing coefficient), thus maintaining the structural stability and inference consistency of the graph over continuous time scales.

[0147] Furthermore, a structural integrity verification mechanism (parameters: edge set, consistency constraint rules) is adopted to detect and repair potential disconnected situations or broken risk propagation paths introduced by edge weight adjustment, ensuring that the dynamic weight graph can be used for subsequent inference and analysis.

[0148] By using the above dynamic weighting processing method, the output of the attention model in the previous step is transformed into time-sensitive relation edge weight data, thereby realizing the ability to dynamically represent the risk propagation path under extreme climate events.

[0149] For example, in an extreme precipitation event, the initial edge weight of the input node between "groundwater source-A" and "pollution source-X" is 0.65, and the attention weight at the corresponding time point is 0.82. The weighted coefficient fusion formula is used: w t =w0×αt , where w t Let w0 be the updated weight at time t, and α be the initial weight. t These are the corresponding attention weights. Substituting the values ​​into the table yields:

[0150] w t =0.65 × 0.82 = 0.533. Subsequently, after smoothing with a time window length of 3, a weighted moving average algorithm is used to eliminate instantaneous peaks, resulting in a final smoothed edge weight of 0.521. This result, mapped to the graph relation attribute field, is used in the dynamic risk propagation weight graph, significantly improving the timeliness and accuracy of path identification in subsequent graph neural network inference.

[0151] S6.5: Perform graph sparsification and normalization operations on the dynamic risk propagation weight graph structure to remove redundant weak connections and improve the efficiency of subsequent graph neural network inference, thereby obtaining an optimized time-weighted risk propagation path model.

[0152] Step S7: Input the updated dynamic knowledge graph into the graph neural network model to perform inference analysis of risk propagation paths and identify potential groundwater pollution diffusion paths and key impact nodes. Specifically, this includes:

[0153] S7.1: Perform graph structure feature extraction processing on the updated dynamic knowledge graph to obtain the topological connection features of entity nodes and the semantic weight distribution of relation edges in the graph.

[0154] Using the relation edge weight matrix of the dynamic risk propagation weight graph structure as input data, a graph structure parsing algorithm (parameters: entity node set, relation edge set, edge weight matrix) is adopted to realize the parsing and extraction of the topological connection features of each entity node in the knowledge graph.

[0155] Furthermore, by using the adjacency matrix construction method (parameters: node index mapping table, weighted relation edge set), the graph weighted adjacency matrix is ​​generated, and the weighted connectivity data between nodes is obtained.

[0156] Furthermore, a node degree calculation algorithm (parameter: weighted adjacency matrix) is used to calculate the in-degree and out-degree of each node and generate a node connectivity feature vector.

[0157] Furthermore, by using a weighted relation aggregation method (parameters: edge weight matrix, semantic label set), the semantic weight distribution of each relation edge is extracted, and a relation weight feature matrix covering the entire graph is obtained.

[0158] Furthermore, a graph structure feature normalization algorithm (parameters: node connectivity feature vector, relation weight feature matrix, normalization coefficient) is adopted to normalize the topological features and semantic weights of different dimensions, so as to obtain a graph structure feature representation under a unified scale.

[0159] The above algorithm transforms the dynamic risk propagation weight graph structure into standardized feature data containing node topological connection features and relation edge semantic weight distribution, thus enabling high-quality input for subsequent node embedding representation learning.

[0160] S7.2: Based on the Graph Attention Network (GAT) model, node embedding representation learning is performed on the extracted graph structure features to generate node vector representations with context awareness.

[0161] For the updated dynamic knowledge graph nodes and their topological connection features, the Graph Attention Network (GAT) node embedding representation learning method (parameters: 4 attention heads, 128 feature dimensions, LeakyReLU negative slope 0.2) is adopted to realize the weighted aggregation of node features in local neighborhood and the construction of context-aware capabilities.

[0162] Furthermore, a multi-attention head mechanism (parameters: 4 parallel heads, fusion method: average pooling) is used to aggregate the attention coefficients from multiple perspectives, resulting in a highly robust node embedding representation matrix H. out .

[0163] Furthermore, the embedding representation matrix is ​​normalized by a node feature normalization algorithm to maintain the balance of feature scales of each node and improve the numerical stability of subsequent multi-hop path inference.

[0164] Furthermore, the nonlinear fitting capability of the embedded feature space is enhanced by using a nonlinear activation function (parameter: ELU), thereby generating a final set of node vector representations with context sensitivity and temporal risk response capabilities.

[0165] By using graph attention network embedding representation learning, the topological connection features and dynamic weight distribution extracted in the previous step are transformed into high-dimensional, time-sensitive node embedding data, thereby optimizing the feature representation of the updated dynamic knowledge graph in risk propagation path reasoning.

[0166] S7.3: The node vector representation is fused with the weights of time-sensitive climate change factors to obtain a dynamic representation of the node state.

[0167] A multi-dimensional feature fusion method is used to combine the node vector representation after node embedding via Graph Attention Network (GAT) with the weights of time-sensitive climate change factors output by the temporal attention mechanism (parameter: node embedding matrix H). nThe system uses a time weight vector α and a fusion weight coefficient matrix W to achieve joint modeling of node semantics and temporal dynamic features.

[0168] Furthermore, using a weighted linear combination algorithm (parameters: node vector dimension d, number of climate factors m), the embedded vector elements of each node are weighted dimension-wise with their corresponding time weights to calculate the fusion representation h. tfusion The formula is defined as:

[0169] h tfusion =H n ×W+α, where H n Let W be the node embedding matrix, W be the fusion weight coefficient matrix, and α be the time-sensitive weight vector.

[0170] Furthermore, a temporal convolutional CNN feature enhancement method (parameters: kernel size k, stride s, number of channels c) is employed to perform one-dimensional convolution on the fused vector sequence to capture the local temporal patterns of node states driven by climate factors, and to generate an enhanced node state matrix S. enh .

[0171] Furthermore, by normalizing (parameters: mean μ, variance σ), S enh Perform the standardization transformation, the formula is:

[0172] This ensures that the numerical values ​​of different node states are distributed on the same scale, thereby improving the consistency of subsequent inference.

[0173] Furthermore, through the node state orthogonal mapping algorithm (parameter: orthogonal transformation matrix Q), the normalized node states are mapped to the temporal topological feature space, obtaining the final dynamic node state vector set V with both temporal and topological representations. dyn .

[0174] By employing the aforementioned fusion modeling, feature enhancement, and orthogonal mapping processing methods, the node vector representation and climate time weights from the previous step are transformed into a comprehensive node state representation that can be directly used for dynamic path reasoning, thereby improving the timeliness and adaptability of groundwater risk propagation analysis under extreme climate events.

[0175] S7.4: Perform multi-hop path reasoning on dynamic node state representations based on the Graph Attention Network (GAT) model to identify potential risk propagation paths and their impact intensity in the groundwater system.

[0176] Based on dynamic node state representation data input, a multi-hop path reasoning algorithm (parameters: node embedding matrix, relation weight matrix, maximum number of hops, attenuation coefficient) is adopted to realize recursive traversal and information aggregation along the risk propagation weight edge in the knowledge graph topology to simulate the multi-stage transmission process of potential pollution diffusion in groundwater systems.

[0177] Furthermore, through path search and weight accumulation methods (parameters: initial node set, edge weights, hop count constraints), the sum of weights for all possible propagation paths from the pollution source node to the groundwater source node is calculated, and the cumulative risk impact value of each path is obtained. The cumulative calculation formula is:

[0178]

[0179] Among them, R p Let P be the cumulative risk impact value of path p. p Let w be the set of edges on path p. e Let e ​​be the risk propagation weight.

[0180] Furthermore, a path normalization method (parameters: cumulative risk vector, normalization coefficient) is employed to map the cumulative risk impact values ​​of paths of different lengths to a uniform scale, thereby eliminating the impact of differences in path length. The normalization formula is:

[0181]

[0182] Among them, R p′ L is the normalized path risk value. p Let p be the number of edges of path p.

[0183] Furthermore, by combining the node time-sensitive state vector and the path risk value, a quantitative indicator of the impact intensity of the risk propagation path is calculated using a path importance evaluation function (parameters: node state vector, path risk value, importance weight vector). The calculation formula is as follows:

[0184] I p =R p′ ×β p

[0185] Among them, I p β represents the influence intensity of path p. p The importance coefficient is obtained by combining the node state and path characteristics.

[0186] Through the above-mentioned multi-hop path reasoning and quantitative evaluation processing, the dynamic node state representation and time-series weighted edge weights of the previous step are transformed into groundwater risk propagation paths and their impact intensity indicators with length normalization and time sensitivity, so as to achieve high timeliness and high accuracy in identifying pollution diffusion paths under extreme climate events.

[0187] S7.5: Assess the importance of key nodes in the identified risk propagation pathways to determine the set of nodes that play a dominant role in the spread of groundwater pollution under extreme climate events.

[0188] S7.6: Generate a structured risk propagation path map based on path strength and node importance indicators, which serves as the input for subsequent risk decision-making modules.

[0189] Step S8: Output the risk propagation path results obtained through reasoning to the risk decision-making module, and generate a structured risk level and response recommendation report. Specifically, this includes:

[0190] S8.1: Perform topological analysis on the groundwater risk propagation path output by the inference module, extract the attribute features of each node in the path and the semantic relationship of the edges, in order to identify key pollution sources, propagation media and potential impact areas.

[0191] S8.2: Based on the extracted key node information of the path, combined with the preset groundwater risk assessment index system, the pollution diffusion intensity, propagation rate and impact range of each propagation path are quantitatively scored.

[0192] S8.3: Input the quantitative scoring results into the risk level classification model, classify the overall risk status of the groundwater system based on the fuzzy comprehensive evaluation algorithm, and output three types of structured risk level labels: low, medium and high.

[0193] Based on the quantitative scoring data of pollution diffusion intensity, propagation rate and impact range output from step S8.2, a risk level classification model (parameters: scoring vector, weight coefficient, membership function) is adopted to realize the fuzzy comprehensive evaluation and classification of the overall risk status of the groundwater system.

[0194] A standardization transformation method (parameter: maximum and minimum value method) is used to normalize each indicator in the scoring vector to ensure the comparability of indicators with different dimensions in the fuzzy evaluation model and to generate a normalized scoring matrix.

[0195] Furthermore, by using the fuzzy membership degree calculation method (parameters: triangular membership function, trapezoidal membership function), the membership degree mapping of each normalized indicator to low, medium and high risk levels is realized, and the risk membership degree matrix μ is obtained, where the matrix elements are the membership degree values ​​of the indicators to each risk level.

[0196] Furthermore, a weighted fuzzy synthesis operation (parameters: weight vector w, weighted average method of synthesis operator) is adopted to realize the comprehensive calculation of the membership degree of each indicator and generate a comprehensive risk membership degree vector U; the corresponding formula is:

[0197] U = w × μ, where w is the weight column vector and μ is the membership matrix.

[0198] Furthermore, the final risk classification is achieved through the maximum membership degree discrimination principle (parameter: membership degree vector U); the classification result is three structured risk level labels: low risk, medium risk, or high risk.

[0199] By using a fuzzy comprehensive evaluation algorithm, the quantitative scoring results from the previous step are transformed into risk level labels, thereby achieving a quantifiable classification of the overall risk status of the groundwater system under extreme climate events.

[0200] S8.4: Based on the risk level classification results and path propagation characteristics, match the corresponding response suggestion template from the predefined emergency response strategy knowledge base, and optimize the parameters and content in combination with the current climate and hydrological conditions.

[0201] S8.5: Integrate the structured risk levels and optimized response recommendations into a standardized report format, transmit it to the risk decision-making module through an interface, and simultaneously display the correlation mapping relationship between critical paths and response strategies in the visualization interface.

[0202] Step S9: After each graph evolution and inference is completed, record the differences in graph structure and changes in inference results before and after the evolution, for use in subsequent model iteration optimization and semantic consistency verification. Specifically, this includes:

[0203] S9.1: Compare the differences between the entity node sets before and after the graph evolution, and calculate the node addition and deletion rate based on the graph structure hash algorithm to quantify the degree of structural change of the entity layer during the graph evolution process.

[0204] S9.2: Perform semantic consistency detection on the relation edge set before and after evolution, and use the relation embedding space alignment algorithm to evaluate the semantic offset of the edge type in order to identify potential semantic conflict node pairs during the evolution process.

[0205] S9.3: Generate path coverage index based on inference results, and conduct a comparative analysis of the coverage of risk propagation paths identified before and after evolution to assess the impact of dynamic map updates on the output of the inference module.

[0206] S9.4: Record the feature snapshots of the intermediate states of the graph neural network inference during the evolution process, and extract the hidden state vector sequences of key nodes to construct an interpretability analysis dataset of the graph evolution trajectory.

[0207] S9.5: Integrate structural difference indicators, semantic offsets, path coverage, and inference intermediate features into an evolutionary evaluation report, which serves as feedback input for model parameter tuning and graph update strategy optimization.

[0208] To address the differences in graph structure and inference results before and after evolution, the input objects are the node addition / deletion change rate, relation edge semantic offset, path coverage index, and inference intermediate feature snapshot sequence data generated from S9.1 to S9.4.

[0209] A unified encoding method with multiple indicators (parameters: node change rate, edge semantic offset, path coverage, node hidden state vector) is adopted to realize the index mapping of cross-dimensional data and transform it into a structured description matrix for subsequent processing.

[0210] Furthermore, a multi-factor weighted normalization algorithm (parameters: weight coefficients are calculated according to three indicators: node importance, severity of relationship conflict, and path coverage difference) is used to normalize data of different dimensions and obtain a comparable set of dimensionless indicators.

[0211] Furthermore, by using principal component analysis (parameter: retain components with a cumulative contribution rate greater than 0.9), dimensionality reduction of multi-indicator data is achieved, and feature vectors expressing the core change trend of the spectrum evolution are generated.

[0212] Furthermore, a multimodal feature fusion method (parameters: cross-attention weight matrix of structural feature vector and inference hidden state vector) is adopted to realize the correlation modeling of information from different modalities and generate a comprehensive evolution evaluation feature vector.

[0213] The report generation engine (parameter: output format follows system evaluation report standards) transforms the comprehensive evolution evaluation feature vector into a complete evolution evaluation report containing structural difference indicators, semantic offsets, path coverage analysis conclusions, and key features of the reasoning process, providing quantifiable feedback for model parameter tuning and graph update strategy optimization.

[0214] Step S10: Based on evolutionary history data and feedback results, dynamically adjust the map update strategy and inference model parameters to improve the adaptability and accuracy of groundwater risk propagation path identification under extreme climate events. Specifically, this includes:

[0215] S10.1: Extract features from the structural change data during the historical map evolution process, and use principal component analysis algorithm to identify key evolutionary factors that affect map stability and inference performance in order to generate structural evolution feature vectors.

[0216] S10.2: Quantify the deviation between the risk propagation path results output by each inference analysis and the actual monitoring data, and generate an inference error index sequence based on the mean square error function to evaluate the model prediction accuracy.

[0217] S10.3: Based on the structural evolution feature vector and the inference error index sequence, a multi-objective optimization algorithm is used to construct a graph update strategy optimization model to obtain an update rule set that achieves the optimal balance between structural stability and inference accuracy.

[0218] Based on the structural evolution feature vector and the inference error index sequence, a multi-objective optimization algorithm (parameters: structural stability objective function, inference accuracy objective function, weight allocation coefficient) is used to construct the map update strategy optimization model.

[0219] Furthermore, by using a normalization method (parameters: standard deviation normalization coefficient, minimum-maximum scaling range), the data scale of the structural evolution feature vector and the inference error index sequence is unified, and a comparable normalized feature matrix is ​​obtained.

[0220] Furthermore, a weighted comprehensive scoring function is adopted to combine the structural stability index S(v,p) and the inference accuracy index A(v,p) into a multi-objective performance function F=αS+βA, where α is the structural stability weight and β is the inference accuracy weight.

[0221] Furthermore, the performance function F is globally maximized by using a genetic algorithm optimization method (parameters: population size, crossover probability, mutation probability), and a corresponding set of strategy parameters is generated, including node update ratio, relation edge adjustment priority, and update cycle.

[0222] Furthermore, based on the Pareto optimal solution set selection strategy (parameters: non-dominated solution sorting, crowding distance calculation), an update rule set that balances structural stability and inference accuracy is generated and used as the output of the optimization model.

[0223] By using a multi-objective optimization model, the feature vectors and error indicators from the previous step are transformed into an executable set of graph update rules, thereby achieving a synergistic improvement in the stability of the graph structure and the accuracy of inference under extreme climate event scenarios.

[0224] S10.4: Dynamically adjust the parameter configuration of the graph neural network inference model and fine-tune the model weights based on the reinforcement learning mechanism to improve the adaptability of groundwater risk propagation path identification under extreme climate events.

[0225] S10.5: Write the optimized graph update strategy and inference model parameter configuration into the system knowledge base, and set up a version control mechanism to support adaptive model loading and strategy execution in subsequent extreme weather event scenarios.

[0226] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention. Unless otherwise defined, the technical or scientific terms used herein should have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms "first," "second," "third," and similar terms used in the specification and claims of this patent application do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms "an" or "a" and similar terms do not indicate a quantity limitation, but rather indicate the presence of at least one. The terms "comprising" or "including" and similar terms mean that the element or object preceding "comprising" or "including" covers the element or object listed following "comprising" or "including" and its equivalents, and does not exclude other elements or objects. The multiple involved in the embodiments of this application refers to two or more. A and / or B indicate the presence of three cases: A; B; and A and B.

[0227] The above description is merely an exemplary embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and such modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A groundwater safety assessment method under extreme climate events, specifically including: S1: Collect multi-source heterogeneous data from meteorological, geological, hydrological and remote sensing sources; S2: Preprocess the collected multi-source heterogeneous data to generate a dataset with unified semantic representation; S3: Construct an initial groundwater safety knowledge graph based on the preprocessed dataset, define the semantic structure of entity nodes and relation edges, and establish a graph ontology model; S4: Deploy an extreme climate event monitoring module to receive meteorological early warning data and hydrological anomaly indicators in real time, and determine whether the dynamic evolution triggering conditions of the map are met. S5: If it is determined that an extreme climate event has occurred or is about to occur, the graph dynamic evolution module is activated to update the local structure of key nodes and relation edges in the knowledge graph based on real-time incremental data; S6: In the dynamic evolution of the knowledge graph, a temporal attention mechanism is introduced to calculate the time-varying impact weight of climate change factors on the risk propagation path between nodes, thereby realizing the dynamic weighted processing of the relationship edges between nodes in the knowledge graph. S7: Input the updated dynamic knowledge graph into the graph neural network model to perform reasoning analysis of risk propagation paths and identify potential groundwater pollution diffusion paths and key impact nodes; S8: Output the risk propagation path results obtained from the reasoning to the risk decision-making module, and generate a structured risk level and response recommendation report; S6 includes: S6.1: Normalize the real-time climate change factor data collected during the map evolution process to obtain standardized climate change time series input; S6.2: Construct a multivariate time series feature vector based on the normalized and standardized climate change time series input; S6.3: Input the multivariate temporal feature vector into the designed temporal attention mechanism model, and calculate the attention weight distribution of the impact of each time node on the risk propagation path; S6.4: Based on the attention weight distribution, the relationship edges between nodes in the knowledge graph are dynamically weighted to generate a time-sensitive dynamic risk propagation weight graph structure; S6.5: Perform graph sparsification and normalization operations on the dynamic risk propagation weight graph structure to obtain the optimized time-weighted risk propagation path model.

2. The groundwater safety assessment method under extreme climate events according to claim 1, characterized in that, Step S8 is followed by: S9: After each graph evolution and inference is completed, record the differences in graph structure and changes in inference results before and after the evolution, which will be used for subsequent model iteration optimization and semantic consistency verification. S10: Based on evolutionary history data and feedback results, dynamically adjust the map update strategy and inference model parameters to improve the adaptability and accuracy of groundwater risk propagation path identification under extreme climate events.

3. The groundwater safety assessment method under extreme climate events according to claim 1, characterized in that, The preprocessing of the collected multi-source heterogeneous data includes standardization cleaning and semantic alignment to eliminate data format differences and redundant information interference.

4. The groundwater safety assessment method under extreme climate events according to claim 1, characterized in that, Step S1 involves collecting data, including obtaining precipitation intensity, wind speed, and temperature change rate from the meteorological department.

5. The groundwater safety assessment method under extreme climate events according to claim 1, characterized in that, The meteorological data collection in step S1 includes: collecting geological structure maps, lithological distribution data and soil permeability coefficients provided by geological survey institutions, performing spatial coordinate alignment and stratigraphic attribute encoding based on the Geographic Information System (GIS) platform, and generating a standardized geological feature dataset.

6. The groundwater safety assessment method under extreme climate events according to claim 1, characterized in that, The hydrological data collection in step S1 includes: receiving groundwater level, flow velocity, and water quality indicators from hydrological monitoring stations, and extracting hydrological time series features using the sliding window method to form dynamic hydrological state characterization data.

7. The groundwater safety assessment method under extreme climate events according to claim 1, characterized in that, Step S1 involves acquiring remote sensing data, including: obtaining multispectral satellite imagery and synthetic aperture radar imagery through a remote sensing data interface, performing image preprocessing and land cover type classification, and generating remote sensing assessment layers for surface water distribution and surface infiltration capacity.

8. The groundwater safety assessment method under extreme climate events according to claim 3, characterized in that, Semantic alignment processing is performed on the cleaned multi-source data, and entity names, attribute descriptions and relation expressions are normalized and mapped based on the domain ontology library; the semantically aligned data is vectorized based on knowledge embedding technology to generate entity and relation embedding vectors in a unified semantic space.

9. A groundwater safety assessment method under extreme climate events according to claim 1, characterized in that, Step S4 specifically includes: Acquire extreme weather event warning data issued by meteorological departments, and perform semantic analysis on the warning information based on preset event classification standards; collect groundwater level, flow velocity, and water quality parameters provided by real-time hydrological monitoring stations, and calculate the fluctuation range of their deviation from the baseline value using a sliding window statistical method; Based on the semantic analysis of extreme climate event types and intensity levels, combined with the amplitude of hydrological anomalies, a multi-dimensional event feature vector is constructed to quantify the comprehensive risk index of the current environmental state. The constructed multidimensional event feature vector is input into the pre-trained event triggering classification model, and the output is a binary judgment result indicating whether the current triggering condition of the knowledge graph dynamic evolution is met. If the judgment result indicates that the triggering condition is met, a knowledge graph dynamic evolution activation signal is generated, and the event feature vector and timestamp information are encapsulated as input parameters of the evolution module.

10. A groundwater safety assessment method under extreme climate events according to claim 1, characterized in that, Step S5 specifically includes: Based on the event trigger signal output by the extreme climate event monitoring module, the dynamic evolution process of the knowledge graph is initiated to determine whether an incremental update operation needs to be performed on the knowledge graph. Real-time incremental data from meteorological early warning, hydrological monitoring and remote sensing are parsed and semantically mapped to generate a structured triplet data representation compatible with the knowledge graph ontology model; Based on incremental triplet data, we perform graph entity recognition and relation extraction tasks. We use graph embedding technology to calculate the semantic similarity between new entities and existing knowledge graph nodes to determine whether they belong to existing entities or need to be added as entity nodes. For newly matched triples, perform incremental graph structure update operations, including adding nodes, expanding edges, or updating node attributes; after the graph structure update is completed, perform graph consistency verification and conflict resolution.

Images