A meteorological service information intelligent pushing method based on a knowledge graph

Abstract

The application discloses a kind of meteorological service information intelligent push method based on knowledge graph, it is related to meteorological information service technical field, comprising the following steps: S1, generating meteorological service knowledge graph and meteorological grid entity;S2, generating user portrait node;S3, establish the service association edge of early warning event and early warning target grid;S4, establish the disaster form fingerprint quantization extraction mechanism based on persistent homology group, constructs regional topological complex and tracks topological evolution track, extracts disaster form fingerprint, generates early warning push characteristic vector by fusing meteorological statistical characteristics;S5, screening target push user set;S6, generating rich media meteorological push content;S7, update meteorological service knowledge graph.The application effectively overcomes meteorological data interpolation structure distortion, disaster form quantization difficulty and the problem that push content weight distribution is unreasonable, significantly improves the precision and intelligent level of meteorological service information.

Description

Technical Field

[0001] This invention relates to the field of meteorological information service technology, and in particular to an intelligent method for pushing meteorological service information based on knowledge graphs. Background Technology

[0002] With the explosive growth of meteorological monitoring data, traditional meteorological services face severe challenges in data processing and delivery. Existing spatial interpolation methods mainly rely on geometric distance, ignoring the complex topological structure implicit in meteorological fields, resulting in structural distortions and a lack of confidence assessment in the interpolation results. Meanwhile, traditional delivery mechanisms are mostly based on administrative divisions or simple buffer matching, lacking quantitative extraction of disaster area morphological fingerprints, making it difficult to accurately match user preferences with disaster characteristics. Furthermore, existing content generation methods rely on fixed templates, failing to dynamically evaluate the contribution weight of multi-source science popularization content, leading to severe homogenization of delivered information and making it difficult to meet the needs of refined and personalized meteorological services.

[0003] Therefore, how to provide a knowledge graph-based intelligent method for pushing meteorological service information is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] This invention proposes an intelligent meteorological service information push method based on knowledge graphs. It constructs a meteorological monitoring field potential function and extracts the topological skeleton as the constraint boundary through a topological analysis mechanism based on Morse theory's meteorological feature skeleton constraints and persistent homology groups. Interpolation is then performed using a spatiotemporal separation variogram model to generate meteorological grid entities with confidence intervals. Dempster-Shafer evidence theory is used to synthesize location evidence in the superpower set space, and service confidence measures are calculated to bind user preferences, constructing user profile nodes containing a set of meteorological preference features. Early warning target grids are selected based on the shortest graph distance of the early warning topological skeleton. A disaster morphology fingerprint quantification extraction mechanism based on persistent homology groups is established, constructing a regional topological complex and tracking its evolution trajectory. Betti number sequences and persistent barcodes are extracted as disaster morphology fingerprints, and bottleneck distance is combined to measure preference matching differences to select target push users. A cooperative game model of push content features is constructed, using Shapley values ​​to quantify the contribution weight of node content, fusing and generating rich media meteorological push content. Based on the persistent homology groups of the user feedback topological complex, feedback anomalies are identified, and the meteorological service knowledge graph is updated. This invention effectively overcomes the problems of distorted meteorological data interpolation structure, difficulty in quantifying disaster morphology, and unreasonable weight allocation of pushed content. It realizes full-process topology optimization from data quality control to personalized service feedback, significantly improving the accuracy and intelligence level of meteorological service information.

[0005] According to an embodiment of the present invention, a method for intelligent push of meteorological service information based on knowledge graphs specifically includes: S1. Construct the meteorological monitoring field potential function, use Morse theory to extract meteorological feature skeleton constraints for spatiotemporal interpolation, construct confidence intervals based on topological invariants, and generate a meteorological service knowledge graph and meteorological grid entities. S2. Analyze user subscription requests to construct a demand probability allocation function, synthesize location evidence in the superpower set space, calculate the service confidence measure of meteorological grid entities and bind user preferences to generate user profile nodes; S3. Analyze meteorological early warning information to construct the early warning impact area, use the Morse function to extract the early warning topology skeleton, calculate the shortest graph distance of meteorological grid entities to filter early warning target grids, and establish service association edges between early warning events and early warning target grids; S4. Aggregate the early warning target grid to generate entities of the disaster area, establish a disaster morphology fingerprint quantification extraction mechanism based on continuous homology groups, construct regional topological complex and track the topological evolution trajectory, extract disaster morphology fingerprints, and integrate meteorological statistical features to generate early warning push feature vectors. S5. Traverse user profile nodes with disaster-stricken area entities as the source point, construct topological feature complexes and extract topological fingerprints respectively, use bottleneck distance to measure preference matching differences, and filter target push user sets. S6. Retrieve associated meteorological science popularization guidance nodes by disaster type code, construct a cooperative game model of push content features, calculate the content contribution Shapley value, and generate rich media meteorological push content by fusing feature vectors based on contribution weight; S7. Distribute rich media weather push content based on the target user set, construct the user feedback topology complex and calculate the continuous homology group to identify feedback anomalies and holes, and update the weather service knowledge graph.

[0006] Optionally, S1 specifically includes: S11. Obtain discrete observation data from meteorological monitoring stations, construct an initial meteorological monitoring field using inverse distance weighted interpolation, and perform smoothing and noise reduction processing and Morse regularization processing on the initial meteorological monitoring field in sequence to obtain the meteorological monitoring field potential function. S12. Calculate the gradient vector field and Hessian matrix of the meteorological monitoring field potential function. Determine the critical point type based on the eigenvalue sign of the Hessian matrix. Construct an integral curve along the gradient vector field that starts at the saddle point and ends at the extreme point to form the meteorological feature skeleton. S13. Using meteorological feature skeleton as constraint boundary conditions, construct a spatiotemporal separation variability function model, perform interpolation calculations on the time series of discrete observation data, and generate spatial attribute estimates of meteorological grid entities. S14. Extract the Betti number and Euler characteristic number of the meteorological feature skeleton as topological invariants, construct a spatial distribution model of the estimation error based on the neighborhood constraint relationship of the meteorological feature skeleton, and calculate the meteorological data confidence interval of the spatial attribute estimation of the meteorological grid entity. S15. Construct a meteorological service knowledge graph by taking meteorological grid entities as node objects, spatial attribute estimates and meteorological data confidence intervals as node attributes, and meteorological feature skeletons as topological relationship edges between entities.

[0007] Optionally, S2 specifically includes: S21. Analyze the keywords and geographic location information in the user's subscription request, identify the types and spatial range of meteorological elements that the user is interested in, construct an identification framework and determine the set of propositions, and define the demand probability allocation function based on the set of propositions. S22. Obtain the real-time GPS coordinate sequence uploaded by the user terminal, calculate the projection distribution probability of the coordinate sequence on the meteorological grid entity, and map the projection distribution probability to the basic probability number of the location evidence. S23. In the superpower set space, the demand probability assignment function and the basic probability number are orthogonally summed according to the Dempster-Shafer composition rule to generate the service confidence measure of the meteorological grid entity. S24. Compare the service confidence measure values ​​of each meteorological grid entity, select the meteorological grid entity with the largest service confidence measure as the user's home unit, and bind the user's preference attribute to the user's home unit. S25. Extract the spatiotemporal features of the user's home unit and the key tags in the user's preference attributes, construct a set of meteorological preference features, and integrate the spatial location of the user's home unit to generate user profile nodes.

[0008] Optionally, S3 specifically includes: S31. Analyze the latitude and longitude coordinates and influence radius parameters in the meteorological warning information, construct the buffer polygon, rasterize the buffer polygon, and generate the warning influence area. S32. Use the geographic elevation or warning level of the area affected by the warning as the input scalar value of the Morse function, calculate the contour topology of the scalar field, and construct the warning topology skeleton based on the contour topology. S33. Construct a spatial adjacency graph within the area affected by the early warning, and use Dijkstra's algorithm to calculate the shortest graph distance from meteorological grid entities to early warning topology skeleton nodes; S34. Mark meteorological grid entities whose shortest graph distance is less than the preset proximity threshold as early warning target grids, establish a relationship between the unique identifier of the early warning event and the early warning target grid, and generate service association edges.

[0009] Optionally, the disaster morphology fingerprint quantification extraction mechanism based on persistent homology groups specifically includes: Aggregate the early warning target grid to generate entities in the disaster area, calculate the spatial adjacency relationship of the entities in the disaster area, and construct the Vietoris-Rips complex based on the spatial adjacency relationship; Persistent homology calculations were performed on the Vietoris-Rips complex to generate persistent homology groups by tracking topological invariants under different scale parameters. Based on the statistical analysis of the generation and disappearance of topological holes in various dimensions using persistent homology groups, Betti number sequences and persistent barcodes are extracted as disaster morphological fingerprints. Statistical features of disaster morphology fingerprints are extracted, and attribute information corresponding to standard disaster codes and meteorological statistical features are integrated to construct early warning push feature vectors.

[0010] Optionally, S5 specifically includes: S51. Using the disaster-stricken area entity as the retrieval source point, traverse and retrieve the associated user profile nodes in the meteorological service knowledge graph to obtain the set of meteorological preference features stored in the user profile nodes. S52. Using the spatial extent of the disaster-stricken area entity and the vector space of the meteorological preference feature set as objects respectively, construct the corresponding Vietoris-Rips complex to generate the disaster-stricken area topological feature complex and the user preference topological feature complex. S53. Perform continuous cohomology calculation on the topological feature complex of the disaster-stricken area and the topological feature complex of user preferences, and extract the continuous barcode topological fingerprint; S54. Calculate the bottleneck distance between the topological feature complex of the disaster-stricken area and the persistent barcode topological fingerprint of the topological feature complex of the user preference, and quantify the difference in preference matching based on the bottleneck distance value. S55. Filter user profile nodes whose preference matching difference is less than the preset matching threshold, and summarize the user profile nodes to form a target push user set.

[0011] Optionally, S6 specifically includes: S61. Using the disaster type code as the index key, traverse the meteorological service knowledge graph to retrieve the associated meteorological science popularization guidance nodes, and extract the text and multimedia resources of the meteorological science popularization guidance nodes as a set of game participants. S62. Construct a cooperative game model for push content features, and define the feature fusion function of the early warning push feature vector and the set of game participants as the payoff function. S63. Calculate the marginal contribution of each node in the set of game participants in the feature fusion function, and calculate the content contribution Shapley value of each node based on the marginal contribution. S64. Normalize the Shapley value of the content contribution to generate contribution weights. Based on the contribution weights, perform a weighted summation of the feature vector of the early warning push and the feature vector of the game participant set to generate rich media weather push content.

[0012] Optionally, S7 specifically includes: S71. Analyze the communication address information of the user profile nodes in the target push user set, determine the user channel attributes, and distribute the rich media weather push content to the corresponding user terminals according to the user channel attributes. S72. Collect reading status and interaction behavior data returned by user terminals, construct user feedback topological complex, perform continuous homology calculation on user feedback topological complex, and generate feedback continuous homology group. S73. Analyze the dimension and duration of topological holes based on feedback persistent homology groups, and identify feedback anomalous holes based on the dimension and duration. S74. Abstract the user feedback data corresponding to the feedback anomaly holes into feedback nodes, and abstract the relationship between feedback nodes and early warning target grids into service feedback association edges, and update the meteorological service knowledge graph.

[0013] The beneficial effects of this invention are: (1) This invention establishes a topological skeleton constraint and confidence assessment mechanism for meteorological data fields by introducing Morse theory and continuous cohomology analysis. The meteorological feature skeleton is extracted using gradient vector fields and Hessian matrices as interpolation constraint boundaries, and data confidence intervals are constructed based on topological invariants. This mechanism breaks through the limitation of traditional geometric distance interpolation ignoring the potential structure, realizes the accurate capture of the implicit topological structure of meteorological data, effectively suppresses structural distortion in the interpolation process, and significantly improves the geometric fidelity and data credibility of the basic entities of meteorological knowledge graphs.

[0014] (2) This invention establishes a disaster morphology fingerprint quantification extraction mechanism based on persistent homology groups by constructing regional topological complexes and calculating persistent homology groups. The Vietoris-Rips complex is used to aggregate early warning target grids, track the topological evolution trajectory under different scale parameters, and extract Betti number sequences and persistent barcodes as disaster morphology fingerprints. This mechanism utilizes the invariance of topological features to geometric deformation to achieve a mathematical representation of the macroscopic morphology and microscopic connectivity structure of disasters, accurately identify connected components and void features, solve the problem of difficulty in quantifying complex disaster morphologies, and provide a core matching basis with topological invariance for accurate disaster delivery.

[0015] (3) This invention achieves scientific allocation of content weights and closed-loop optimization of service quality by constructing a cooperative game model of push content features and feedback topology analysis. It uses Shapley values ​​to quantify the content contribution weights of multi-source science popularization nodes, solving the weight allocation problem in the integration of early warning information and science popularization resources; and identifies abnormal holes in the user behavior topology through continuous feedback homology groups, accurately locating service blind spots. This mechanism realizes intelligent control of the entire process from content generation to feedback adjustment, ensuring the personalized optimal configuration of push content and the dynamic evolution of service strategies. Attached Figure Description

[0016] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of a knowledge graph-based intelligent push method for meteorological service information proposed in this invention. Figure 2 This is a flowchart illustrating the working principle of the disaster morphology fingerprint quantification extraction mechanism based on continuous homology groups in the intelligent push method for meteorological service information based on knowledge graphs proposed in this invention. Detailed Implementation

[0017] The invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0018] refer to Figure 1 and Figure 2 A knowledge graph-based intelligent method for pushing meteorological service information specifically includes: S1. Construct the meteorological monitoring field potential function, use Morse theory to extract meteorological feature skeleton constraints for spatiotemporal interpolation, construct confidence intervals based on topological invariants, and generate a meteorological service knowledge graph and meteorological grid entities. S2. Analyze user subscription requests to construct a demand probability allocation function, synthesize location evidence in the superpower set space, calculate the service confidence measure of meteorological grid entities and bind user preferences to generate user profile nodes; S3. Analyze meteorological early warning information to construct the early warning impact area, use the Morse function to extract the early warning topology skeleton, calculate the shortest graph distance of meteorological grid entities to filter early warning target grids, and establish service association edges between early warning events and early warning target grids; S4. Aggregate the early warning target grid to generate entities of the disaster area, establish a disaster morphology fingerprint quantification extraction mechanism based on continuous homology groups, construct regional topological complex and track the topological evolution trajectory, extract disaster morphology fingerprints, and integrate meteorological statistical features to generate early warning push feature vectors. S5. Traverse user profile nodes with disaster-stricken area entities as the source point, construct topological feature complexes and extract topological fingerprints respectively, use bottleneck distance to measure preference matching differences, and filter target push user sets. S6. Retrieve associated meteorological science popularization guidance nodes by disaster type code, construct a cooperative game model of push content features, calculate the content contribution Shapley value, and generate rich media meteorological push content by fusing feature vectors based on contribution weight; S7. Distribute rich media weather push content based on the target user set, construct the user feedback topology complex and calculate the continuous homology group to identify feedback anomalies and holes, and update the weather service knowledge graph.

[0019] In this embodiment, S1 specifically includes: S11. Obtain discrete observation data from meteorological monitoring stations, calculate the reciprocal of the Euclidean distance from the monitoring station to the grid point as the weight, accumulate the product of the observation value of each station with the corresponding weight to generate an initial meteorological monitoring field, perform Gaussian low-pass filtering on the initial meteorological monitoring field to eliminate high-frequency noise interference, calculate the first-order difference value of the grid point value along the longitudinal and latitudinal directions, when the difference value in both directions is less than 0.001, the point is determined to be a critical point, calculate the numerical variance of the 8 neighboring grid points around the critical point, when the variance value is less than 0.01, it is determined to be a degenerate critical point, calculate the numerical gradient of the neighbors in the 4 diagonal directions around the point, select the direction with the largest absolute gradient value as the adjustment benchmark, add the gradient value to the critical point value multiplied by 0.5 times the grid spacing, forcibly change the value of the point so that the changed gradient direction is consistent with the surrounding trend, thereby eliminating unstructured small fluctuations and obtaining the meteorological monitoring field potential function; S12. Calculate the first-order difference of the meteorological monitoring field potential function in the meridional and zonal directions to construct the gradient vector field. Use the second-order difference method to calculate the second derivative to construct the Hessian matrix. Calculate the two eigenvalues ​​of the Hessian matrix. If both eigenvalues ​​are positive, it is determined to be a minimum point. If both eigenvalues ​​are negative, it is determined to be a maximum point. If one eigenvalue is positive and the other is negative, it is determined to be a saddle point. Using the saddle point coordinates as the starting position, move 0.1 grid spacing along the gradient descent direction and iterate until the extreme point is reached. Connect the coordinate points on the tracking path to form the meteorological feature skeleton. S13. Divide the interpolation calculation units by the meteorological feature skeleton, calculate the covariance function value of discrete observation data on the time axis, construct the time dimension variability function, calculate the semivariability function value of discrete observation data on the spatial plane, construct the spatial dimension variability function, use the weighted sum of the time dimension variability function and the spatial dimension variability function as the Kriging interpolation coefficient, perform weighted summation on the time series of discrete observation data, and generate the spatial attribute estimate of the meteorological grid entity. S14. The number of independent, unconnected regions in the meteorological feature skeleton is taken as the zero-dimensional Betti number, and the number of closed loop structures in the skeleton is taken as the one-dimensional Betti number. The Euler characteristic number is calculated by subtracting the total number of edges from the total number of vertices and adding the total number of top edges. The Betti number and the Euler characteristic number are used as topological invariants. The sum of squares of the difference between the observed value and the estimated value is calculated in each topological neighborhood divided by the meteorological feature skeleton. The variance is obtained by dividing the sum of squares of the difference by the sample size. A value of 1.96 times the standard deviation is selected as the error limit, and the confidence interval of the estimated spatial attribute of the meteorological grid entity is calculated. S15. Use the combination of latitude and longitude coordinates and elevation values ​​of meteorological grid entities as unique node identifiers, write the estimated values ​​of spatial attributes and the upper and lower limits of confidence intervals into the node attribute table, define the line segments connecting critical points in the meteorological feature skeleton as directed edges between two nodes, write the starting and ending coordinates of the edges into the edge relationship table, and use the resource description framework to serialize the node attribute table and edge relationship table to construct a meteorological service knowledge graph.

[0020] In this embodiment, S2 specifically includes: S21. Parse the text content in the user subscription request, traverse the preset meteorological terminology dictionary to match the strings in the text to extract keywords, read the latitude and longitude characters in the text and convert them into numerical values ​​as geographical location information, define a single meteorological element or combination of elements as an atomic proposition within the identification framework, construct an identification framework containing three types of propositions: high temperature, rainfall, and strong wind, count the number of high temperature keywords matched in the user subscription text as N1, the number of rainfall keywords matched as N2, and the number of strong wind keywords matched as N3, calculate the sum of the three values ​​N1, N2, and N3 as the denominator, divide N1, N2, and N3 as the numerators by the denominator, and obtain three values ​​that sum to 1 as the final values ​​of the demand probability allocation function; S22. Obtain the real-time GPS coordinate sequence uploaded by the user terminal, calculate the projected coordinates of each GPS coordinate point within the meteorological grid entity area, count the number of coordinate points within a specific meteorological grid entity area and divide by the total number of coordinate points to obtain the position proportion value of the grid entity, calculate the difference between 1 and the position proportion value as the position confidence value of non-grid entity, and combine the position proportion value with the position confidence value of non-grid entity to form the basic probability number of position evidence. S23. List all non-empty subsets of the three propositions of high temperature, rainfall, and strong wind in the identification framework to form a superpower set space. Read the probability values ​​corresponding to each subset from the demand probability allocation function as the first column vector. Read the probability values ​​corresponding to each subset from the location evidence as the second column vector. Traverse each pair of values ​​in the first and second column vectors and calculate the intersection of the two sets of propositions corresponding to the values. If the intersection is empty, add the product of the value pair to the conflict coefficient. If the intersection is not empty, add the product of the value pair to the composite probability value corresponding to the intersection. Calculate the difference between 1 and the conflict coefficient as the normalized denominator. Divide the composite probability value by the normalized denominator to obtain the corrected service confidence measure value. S24. Traverse all meteorological grid entities, compare the service confidence measure values ​​of each meteorological grid entity, select the meteorological grid entity with the largest service confidence measure value as the user's belonging unit, extract the meteorological element type label and attention threshold from the user's subscription request, write the meteorological element type label and attention threshold into the attribute field of the user's belonging unit, and complete the binding of user preference attributes. S25. Extract the latitude and longitude coordinates of the center point of the user's unit as a spatial location feature, extract the meteorological element type labels contained in the user's unit as a preference feature, combine the spatial location feature and preference feature into a two-dimensional data structure, construct a meteorological preference feature set, create user profile nodes using a graph database, and store the meteorological preference feature set as the attribute values ​​of the nodes.

[0021] In this embodiment, S3 specifically includes: S31. Parse the meteorological warning information text, extract the latitude and longitude coordinates contained therein as the center coordinates, extract the influence radius value as the buffer distance, extend the buffer distance in all directions with the center coordinates as the reference point to construct a circular buffer polygon, establish a planar grid covering the range of the buffer polygon, assign the grid cells covered by the buffer polygon to 1 and the grid cells not covered to 0, and generate the warning influence area composed of binary grid cells. S32. Traverse every grid cell within the area affected by the early warning, calculate the shortest distance value from the center point of the current grid cell to the boundary of the area affected by the early warning, assign the shortest distance value to the current grid cell as a distance scalar value, perform three-dimensional space fitting on the distance scalar values ​​of all grid cells to construct a distance scalar field, extract the grid cell with the largest distance scalar value in the distance scalar field as the local highest point, search for the ridge line formed by connecting grid cells with consecutive equal distance scalar values ​​in the distance scalar field, and connect the local highest point with the ridge line to form the early warning topology skeleton; S33. Construct a spatial adjacency graph within the area affected by the early warning. Take meteorological grid entities as graph nodes and the adjacency relationship between adjacent grid entities as edges. Use Dijkstra's algorithm to calculate the path cost from each meteorological grid entity node to the early warning topology skeleton node, and select the value with the minimum path cost as the shortest graph distance. S34. Compare the shortest map distance value with the nearest threshold value of 5 kilometers. If the shortest map distance value is less than 5 kilometers, mark the corresponding meteorological grid entity as the warning target grid, generate a string containing the unique identifier of the warning event, write the string into the attribute field of the warning target grid, establish the association between the warning target grid and the warning event, and generate service association edges.

[0022] In this embodiment, the disaster morphology fingerprint quantification extraction mechanism based on continuous homology groups specifically includes: Traverse all grid cells marked as early warning target grids, merge adjacent early warning target grids with common edges or common vertices into a single polygon, generate disaster area entities, construct the spatial adjacency matrix of disaster area entities, map disaster area entities to vertices, map spatial adjacency relationships to edges, set the scale parameter threshold to 10 kilometers, connect all vertex pairs with a spatial distance of less than 10 kilometers, and construct the Vietoris-Rips complex. Using the scale threshold as a filtering parameter, the filter parameter sequence is traversed in ascending order of value. Under each filtering parameter, the existing 0-dimensional simplex (vertices), 1-dimensional simplex (edges), and 2-dimensional simplex (triangles) in the Vietoris-Rips complex are identified. The simplexes are sorted according to their dimension size and lexicographical order, and column and row indices are assigned. The intersection relationship between two simplexes is calculated. If one simplex is a proper subset of another and their dimension difference is 1, the corresponding matrix element is assigned a value of 1; otherwise, it is assigned a value of 0, generating a boundary matrix composed of 0 and 1 elements. The boundary matrix is ​​then traversed from the leftmost column to the right to find the current column. If there are no non-zero elements in the current column or the row index of a non-zero element is less than the current column index, then the current column remains unchanged. If there are non-zero elements and their row index is greater than the current column index, then the column containing the non-zero element with the largest row index is XORed with the current column, and the row index of the non-zero element in the current column is updated to the largest row index of the XOR sum. This XORing and adding operation is repeated until the current column becomes a zero vector or only non-zero elements with row indices less than the current column index remain, thus completing the Gaussian elimination operation of the boundary matrix. The number of non-zero columns in the boundary matrix after elimination is counted as a topological invariant, generating a persistent homology group composed of homology group generators. Based on the persistent homology group, the scale parameter corresponding to the generation of each topological hole is recorded as the birth time, and the scale parameter corresponding to the disappearance of the topological hole is recorded as the death time. The difference between the birth time and the death time is calculated as the continuous lifespan. The Betti number sequence of topological holes in each dimension is extracted, and the continuous barcode of the topological hole is drawn according to the birth time and death time. The Betti number sequence and the continuous barcode are combined to form a disaster morphology fingerprint. The mean, variance, and maximum value of the lifetime in the disaster morphology fingerprint are calculated as statistical features. The standard disaster code and warning level corresponding to the entity in the disaster area are read as attribute information. The average rainfall and maximum wind speed in the entity in the disaster area are calculated as meteorological statistical features. The statistical features, attribute information and meteorological statistical features are concatenated into vectors to construct the warning push feature vector.

[0023] In this embodiment, S5 specifically includes: S51. Using the disaster-stricken area entity as the retrieval source point, traverse and retrieve the associated user profile nodes in the meteorological service knowledge graph to obtain the set of meteorological preference features stored in the user profile nodes. S52. Read the spatial boundary coordinate point set of entities in the disaster-stricken area, calculate the Euclidean distance between each coordinate point in the point set, and construct the topological feature complex of the disaster-stricken area; read the meteorological preference feature set stored in the user profile node, extract the multi-dimensional vector data in the feature set, calculate the cosine similarity between each pair of vectors in the multi-dimensional vector data, convert the cosine similarity into a cosine distance value as a metric, set a distance threshold of 0.1, determine the feature vectors with a cosine distance value less than 0.1 as neighboring points, connect the neighboring points to construct edges, identify the triangles formed by the closed edges as 2-dimensional simplexes, and combine all vertices, edges and triangles to generate the topological feature complex of user preferences; S53. Set a filtering step size of 500 meters for the topological feature complex of the disaster-stricken area and a filtering step size of 0.05 for the topological feature complex of the user preference. Traverse the corresponding scale parameter sequence in ascending order of filtering step size, identify the simplex currently existing in the complex, sort each simplex according to dimension size and lexicographical order and assign column index and row index, construct the boundary matrix, perform XOR elimination operation on the boundary matrix, count the number of non-zero columns in the boundary matrix after elimination, and extract the continuous barcode topological fingerprint. S54. Construct a set of matching pairs between the topological fingerprints of the disaster area topological feature complex and the user preference topological feature complex of the continuous barcode. Calculate the maximum value of the difference between the coordinates of the two barcode endpoints in the matching pair as the point cost. Calculate half the length of the unmatched barcode as the edge cost. Select the maximum value among the point cost of all matching pairs and the edge cost of the unmatched barcode as the bottleneck distance. Quantify the preference matching difference based on the bottleneck distance value. S55. Compare the bottleneck distance value with the preset matching threshold value of 0.5. If the bottleneck distance value is less than 0.5, it is determined that the corresponding user profile node meets the push conditions. The user profile nodes that meet the push conditions are aggregated to form the target push user set.

[0024] In this embodiment, S6 specifically includes: S61. Using the disaster type code as the index key, traverse and retrieve the associated meteorological science popularization guidance nodes in the meteorological service knowledge graph, and extract the disaster avoidance guide text, science popularization video links and defense diagrams stored in the meteorological science popularization guidance nodes as a set of game participants. S62. Construct a cooperative game model for push content features. Specifically, use the BERT model to extract semantic feature vectors from the risk avoidance guide text, use the ResNet-50 model to extract image feature vectors from the defense diagram, and use the 3D-CNN model to extract visual feature vectors from the popular science video. Define the calculation process of the profit function as follows: for any participant alliance subset, calculate the cosine similarity between all feature vectors in the alliance subset and the warning push feature vector. Select feature vectors with values ​​greater than 0.5 to construct a feature combination matrix. Calculate the covariance matrix of the feature combination matrix. Perform singular value decomposition on the covariance matrix to extract the maximum singular value. Multiply the maximum singular value by the feature fusion coefficient of 0.85 to obtain the profit function value of the current alliance subset. Set the calculation objective of the profit function to maximize the matching degree between the push content and the warning features. S63. Set the risk avoidance guide text, popular science video link and defense diagram as three participant nodes. List all possible permutations and combinations of participants, a total of 6. Under each permutation, calculate the increment of the benefit function value after the current participant node joins the existing participant alliance. Record this increment as the marginal contribution under the current order. Traverse all permutations, accumulate the marginal contribution values ​​of each participant node under different orders and divide by the total number of permutations, 6, to obtain the content contribution Shapley value of each node. The sum of the Shapley values ​​of the three nodes S64, the cumulative risk avoidance guide text, the science video link, and the defense diagram is obtained. The ratio of the Shapley value of each node to the sum is calculated, and the ratio result is used as the contribution weight. The sum of the values ​​of all contribution weights is 1. The resource identifiers are sorted in descending order according to the size of the contribution weight. The top 3 resources with the largest contribution weight values ​​are selected and fused with the early warning push feature vector to generate rich media weather push content.

[0025] In this embodiment, S7 specifically includes: S71. Parse the communication address information of the user profile node in the target push user set, extract the domain name suffix and protocol type in the communication address field to determine the user channel attribute. If the channel attribute is a mobile application, call the message push interface to distribute rich media weather push content. If the channel attribute is a web page, call the hypertext transfer protocol interface to render and display the content. S72. Collect reading status identifiers and click and favorite interaction data returned by user terminals. Construct a behavior sequence based on the timestamp order of the behavior occurrence. Set the sliding window length to 5 behavior steps. Within the sliding window, map a single user feedback behavior to a single vertex in a high-dimensional space. Calculate the Euclidean distance between the behavior feature vectors of any two vertices. If the Euclidean distance is less than the preset proximity threshold of 0.3, connect an edge between the two vertices. If there are edges between every pair of the three vertices, fill the triangle facets to construct the user feedback topological complex. Construct the boundary matrix of the user feedback topological complex. Use Gaussian elimination to perform upper triangularization on the boundary matrix. Specifically, start from the last column of the matrix and traverse backwards to the first column. If there are non-zero elements in the current column, eliminate the non-zero elements above the main diagonal in the column through row transformation to make the matrix present an upper triangular state. If the current column is all zero, retain it as a zero vector. Record the distribution changes of non-zero columns and the generation and disappearance times of topological features during the elimination process. Generate a feedback continuous homology group containing information on the changes in the Betti number of each dimension. S73. Based on the dimension value and the time point of disappearance of the topological holes statistically analyzed by the feedback continuous homology group, the time difference between the hole generation time point and the time point of disappearance is calculated as the duration. The dimension threshold is set to 1 and the duration threshold is set to 4 time units. Topological features with a dimension value greater than the dimension threshold and a duration greater than the duration threshold are identified as feedback abnormal holes. S74. Abstract the user feedback data corresponding to the feedback anomaly hole into feedback nodes, extract the user identifier, anomaly type and interaction timestamp attributes from the feedback nodes, abstract the association between the feedback nodes and the early warning target grid into service feedback association edges, and append the feedback nodes and service feedback association edges to the storage layer of the meteorological service knowledge graph to complete the dynamic update of the graph data.

[0026] Example 1: To verify the feasibility of this invention in the accurate delivery of meteorological early warning information and the evaluation of service effectiveness, the method of this invention was applied to the meteorological early warning system for sudden disasters of a provincial meteorological service center (hereinafter referred to as "Center M"). In traditional meteorological early warning service systems, template-based mass SMS messaging or single-image / text push methods are typically used. The content is monotonous, lacks in-depth content correlation for different disaster types, and cannot effectively perceive users' actual feedback and behavioral anomalies after receiving the warning, resulting in low readability of the warning information and difficulty in quantifying the effectiveness of emergency guidance. To solve the above problems, Center M decided to adopt the meteorological early warning content generation and feedback analysis method based on knowledge graphs and game theory proposed in this invention.

[0027] During implementation, Center M first constructed a multimodal meteorological service knowledge graph. Using disaster type codes as index keys, it traversed and retrieved associated meteorological science popularization guidance nodes, extracting evacuation guide texts, science popularization video links, and defense diagrams as a set of game participants. Center M's technical team used BERT, ResNet-50, and 3D-CNN models to extract semantic and visual feature vectors from text, images, and videos, respectively, and constructed a collaborative game model for push content features. For any subset of participant alliances, the system calculated the cosine similarity between the feature vector and the early warning push feature vector, selected features with values ​​greater than 0.5 to construct a combination matrix, and obtained the payoff function by multiplying the maximum singular value extracted through singular value decomposition with a coefficient of 0.85. By listing six permutation and combination orders, the Shapley value of each node was calculated and normalized to generate contribution weights. Based on this, the top three resources with the highest weights were selected for fusion, realizing intelligent generation and accurate matching of rich media meteorological push content.

[0028] In the content distribution and feedback analysis phase, center M analyzes channel attributes based on the communication address information of user profile nodes to accurately distribute rich media content to mobile or web terminals. Subsequently, the system collects reading status and interaction behavior data returned by user terminals, constructs behavior sequences based on timestamp order, and uses a sliding window and Euclidean distance threshold to construct a user feedback topological complex. The boundary matrix is ​​processed by upper triangularization using Gaussian elimination, the feedback continuous homology group is calculated, and the dimension and duration of topological holes are statistically analyzed. Feedback anomaly holes with a dimension greater than 1 and a duration greater than 4 time units are successfully identified and abstracted as feedback nodes to update the meteorological service knowledge graph, achieving topological awareness and dynamic closed-loop of service effectiveness.

[0029] During implementation, the technical team at Center M discovered that, compared to traditional single-channel push notifications and simple statistical analysis methods, the method of this invention significantly improves the relevance of weather warning content and the sensitivity of user feedback. Traditional methods generate warning content that lacks effective integration of multimodal resources, making it difficult to attract user attention and exhibiting a lag in identifying abnormal feedback behavior. In contrast, the method of this invention optimizes content composition through a Shapley value game model and accurately captures potential abnormal patterns in user behavior through continuous homology calculation.

[0030] To further verify the actual performance of the method of the present invention, Center M conducted a detailed comparative test between the method of the present invention and the traditional method. The specific performance data is shown in Table 1: Table 1 Performance Comparison of Central M Meteorological Early Warning Service System

[0031] As shown in Table 1, the performance of the meteorological early warning service system was comprehensively improved after applying the method of this invention. The matching accuracy of early warning content increased from 75.2% with traditional methods to 94.6%, and the average user reading time increased significantly from 12.5 seconds to 35.8 seconds, indicating that the generated rich media content better meets user needs and significantly enhances the dissemination effectiveness of early warning information. The accuracy of anomaly feedback identification increased from 68.5% to 96.2%, and the false alarm interference rate decreased from 15.0% to 2.2%, proving that the feedback mechanism based on topological data analysis can effectively eliminate noise and accurately locate service blind spots. The knowledge graph update delay was shortened from 120 minutes to 5 minutes, enabling rapid iterative optimization of the early warning service system. In addition, the operation and maintenance manpower cost decreased from 900,000 yuan / year to 350,000 yuan / year, and the service satisfaction rate increased from 79.0% to 96.5%, significantly improving social and economic benefits.

[0032] Through the method of this invention, Center M has successfully realized the intelligent generation of meteorological warning content and the deep perception of service feedback, effectively solving the problems of inaccurate content delivery and insufficient feedback analysis, greatly improving the level of precision of meteorological services and emergency response capabilities, and providing strong technical support for the construction of smart meteorological services.

[0033] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for intelligent push of meteorological service information based on knowledge graphs, characterized in that, Includes the following steps: S1. Construct the meteorological monitoring field potential function, use Morse theory to extract meteorological feature skeleton constraints for spatiotemporal interpolation, construct confidence intervals based on topological invariants, and generate a meteorological service knowledge graph and meteorological grid entities. S2. Analyze user subscription requests to construct a demand probability allocation function, synthesize location evidence in the superpower set space, calculate the service confidence measure of meteorological grid entities and bind user preferences to generate user profile nodes; S3. Analyze meteorological early warning information to construct the early warning impact area, use the Morse function to extract the early warning topology skeleton, calculate the shortest graph distance of meteorological grid entities to filter early warning target grids, and establish service association edges between early warning events and early warning target grids; S4. Aggregate the early warning target grid to generate entities of the disaster area, establish a disaster morphology fingerprint quantification extraction mechanism based on continuous homology groups, construct regional topological complex and track the topological evolution trajectory, extract disaster morphology fingerprints, and integrate meteorological statistical features to generate early warning push feature vectors. S5. Traverse user profile nodes with disaster-stricken area entities as the source point, construct topological feature complexes and extract topological fingerprints respectively, use bottleneck distance to measure preference matching differences, and filter target push user sets. S6. Retrieve associated meteorological science popularization guidance nodes by disaster type code, construct a cooperative game model of push content features, calculate the content contribution Shapley value, and generate rich media meteorological push content by fusing feature vectors based on contribution weight; S7. Distribute rich media weather push content based on the target user set, construct the user feedback topology complex and calculate the continuous homology group to identify feedback anomalies and holes, and update the weather service knowledge graph.

2. The method for intelligent push of meteorological service information based on knowledge graphs according to claim 1, characterized in that, S1 specifically includes: S11. Obtain discrete observation data from meteorological monitoring stations, construct an initial meteorological monitoring field using inverse distance weighted interpolation, and perform smoothing and noise reduction processing and Morse regularization processing on the initial meteorological monitoring field in sequence to obtain the meteorological monitoring field potential function. S12. Calculate the gradient vector field and Hessian matrix of the meteorological monitoring field potential function. Determine the critical point type based on the eigenvalue sign of the Hessian matrix. Construct an integral curve along the gradient vector field that starts at the saddle point and ends at the extreme point to form the meteorological feature skeleton. S13. Using meteorological feature skeleton as constraint boundary conditions, construct a spatiotemporal separation variability function model, perform interpolation calculations on the time series of discrete observation data, and generate spatial attribute estimates of meteorological grid entities. S14. Extract the Betti number and Euler characteristic number of the meteorological feature skeleton as topological invariants, construct a spatial distribution model of the estimation error based on the neighborhood constraint relationship of the meteorological feature skeleton, and calculate the meteorological data confidence interval of the spatial attribute estimation of the meteorological grid entity. S15. Construct a meteorological service knowledge graph by taking meteorological grid entities as node objects, spatial attribute estimates and meteorological data confidence intervals as node attributes, and meteorological feature skeletons as topological relationship edges between entities.

3. The method for intelligent push of meteorological service information based on knowledge graphs according to claim 1, characterized in that, S2 specifically includes: S21. Analyze the keywords and geographic location information in the user's subscription request, identify the types and spatial range of meteorological elements that the user is interested in, construct an identification framework and determine the set of propositions, and define the demand probability allocation function based on the set of propositions. S22. Obtain the real-time GPS coordinate sequence uploaded by the user terminal, calculate the projection distribution probability of the coordinate sequence on the meteorological grid entity, and map the projection distribution probability to the basic probability number of the location evidence. S23. In the superpower set space, the demand probability assignment function and the basic probability number are orthogonally summed according to the Dempster-Shafer composition rule to generate the service confidence measure of the meteorological grid entity. S24. Compare the service confidence measure values ​​of each meteorological grid entity, select the meteorological grid entity with the largest service confidence measure as the user's home unit, and bind the user's preference attribute to the user's home unit. S25. Extract the spatiotemporal features of the user's home unit and the key tags in the user's preference attributes, construct a set of meteorological preference features, and integrate the spatial location of the user's home unit to generate user profile nodes.

4. The intelligent push method for meteorological service information based on knowledge graphs according to claim 1, characterized in that, S3 specifically includes: S31. Analyze the latitude and longitude coordinates and influence radius parameters in the meteorological warning information, construct the buffer polygon, rasterize the buffer polygon, and generate the warning influence area. S32. Use the geographic elevation or warning level of the area affected by the warning as the input scalar value of the Morse function, calculate the contour topology of the scalar field, and construct the warning topology skeleton based on the contour topology. S33. Construct a spatial adjacency graph within the area affected by the early warning, and use Dijkstra's algorithm to calculate the shortest graph distance from meteorological grid entities to early warning topology skeleton nodes; S34. Mark meteorological grid entities whose shortest graph distance is less than the preset proximity threshold as early warning target grids, establish a relationship between the unique identifier of the early warning event and the early warning target grid, and generate service association edges.

5. The method for intelligent push of meteorological service information based on knowledge graphs according to claim 1, characterized in that, The disaster morphology fingerprint quantification extraction mechanism based on continuous homology groups specifically includes: Aggregate the early warning target grid to generate entities in the disaster area, calculate the spatial adjacency relationship of the entities in the disaster area, and construct the Vietoris-Rips complex based on the spatial adjacency relationship; Persistent homology calculations were performed on the Vietoris-Rips complex to generate persistent homology groups by tracking topological invariants under different scale parameters. Based on the statistical analysis of the generation and disappearance of topological holes in various dimensions using persistent homology groups, Betti number sequences and persistent barcodes are extracted as disaster morphological fingerprints. Statistical features of disaster morphology fingerprints are extracted, and attribute information corresponding to standard disaster codes and meteorological statistical features are integrated to construct early warning push feature vectors.

6. The intelligent push method for meteorological service information based on knowledge graphs according to claim 1, characterized in that, S5 specifically includes: S51. Using the disaster-stricken area entity as the retrieval source point, traverse and retrieve the associated user profile nodes in the meteorological service knowledge graph to obtain the set of meteorological preference features stored in the user profile nodes. S52. Using the spatial extent of the disaster-stricken area entity and the vector space of the meteorological preference feature set as objects respectively, construct the corresponding Vietoris-Rips complex to generate the disaster-stricken area topological feature complex and the user preference topological feature complex. S53. Perform continuous cohomology calculation on the topological feature complex of the disaster-stricken area and the topological feature complex of user preferences, and extract the continuous barcode topological fingerprint; S54. Calculate the bottleneck distance between the topological feature complex of the disaster-stricken area and the persistent barcode topological fingerprint of the topological feature complex of the user preference, and quantify the difference in preference matching based on the bottleneck distance value. S55. Filter user profile nodes whose preference matching difference is less than the preset matching threshold, and summarize the user profile nodes to form a target push user set.

7. The intelligent push method for meteorological service information based on knowledge graphs according to claim 1, characterized in that, S6 specifically includes: S61. Using the disaster type code as the index key, traverse the meteorological service knowledge graph to retrieve the associated meteorological science popularization guidance nodes, and extract the text and multimedia resources of the meteorological science popularization guidance nodes as a set of game participants. S62. Construct a cooperative game model for push content features, and define the feature fusion function of the early warning push feature vector and the set of game participants as the payoff function. S63. Calculate the marginal contribution of each node in the set of game participants in the feature fusion function, and calculate the content contribution Shapley value of each node based on the marginal contribution. S64. Normalize the Shapley value of the content contribution to generate contribution weights. Based on the contribution weights, perform a weighted summation of the feature vector of the early warning push and the feature vector of the game participant set to generate rich media weather push content.

8. The intelligent push method for meteorological service information based on knowledge graphs according to claim 1, characterized in that, Specifically, S7 includes: S71. Analyze the communication address information of the user profile nodes in the target push user set, determine the user channel attributes, and distribute the rich media weather push content to the corresponding user terminals according to the user channel attributes. S72. Collect reading status and interaction behavior data returned by user terminals, construct user feedback topological complex, perform continuous homology calculation on user feedback topological complex, and generate feedback continuous homology group. S73. Analyze the dimension and duration of topological holes based on feedback persistent homology groups, and identify feedback anomalous holes based on the dimension and duration. S74. Abstract the user feedback data corresponding to the feedback anomaly holes into feedback nodes, and abstract the relationship between feedback nodes and early warning target grids into service feedback association edges, and update the meteorological service knowledge graph.

Images