Township-oriented weather disaster risk grading intelligent assessment and pushing method

Abstract

The application provides a township-oriented meteorological disaster risk grading intelligent assessment and pushing method, relates to the meteorological disaster early warning technical field, and comprises the following steps: acquiring meteorological monitoring data of multiple monitoring points in a township, extracting and fusing time-space features and geographical element features by using a time-space attention mechanism and a dynamic weight mechanism, and generating a meteorological disaster risk grade by a bidirectional long short-term memory network; meanwhile, a vulnerability assessment model is constructed based on functional area distribution information, the risk grade and vulnerability features are dynamically fused, and a comprehensive risk index is obtained; finally, a distributed storage network is constructed by hierarchical cascading blockchains, early warning information is encrypted and fragmented, and pushing is performed. The application can realize accurate assessment and efficient early warning of township meteorological disaster risks, and improve the reliability and timeliness of early warning information.

Description

Technical Field

[0001] This invention relates to the field of meteorological disaster early warning technology, and in particular to a method for intelligent assessment and push notification of meteorological disaster risk classification for towns. Background Technology

[0002] With the acceleration of urbanization, towns, as important nodes in the urban-rural fringe, have their meteorological disaster prevention capabilities directly related to regional security and economic development. Currently, meteorological disaster risk assessment in towns mainly relies on meteorological station data and human experience, and various regions have also established early warning information dissemination systems at different levels. These systems collect meteorological monitoring data, combine it with geographic information systems to assess disaster risks, and issue early warning information to relevant departments and the public.

[0003] However, existing meteorological disaster risk assessment methods still have several shortcomings. Traditional assessment methods often analyze meteorological elements in isolation, ignoring the spatiotemporal correlation between elements, and the integration of geographical features is too simplistic, making it difficult to accurately depict the evolution of disasters in complex geographical environments. Existing risk assessment models do not fully consider the differences in functional zoning of towns and districts, and the characterization of regional vulnerability features is not precise enough, making it impossible to achieve multi-level and multi-dimensional comprehensive risk assessment. In the process of early warning information dissemination, there are problems such as insufficient information security, inaccurate terminal screening, and low dissemination efficiency, making it difficult to ensure the timely and reliable delivery of early warning information.

[0004] In summary, this invention proposes a meteorological disaster risk assessment method. It achieves adaptive fusion of meteorological elements and geographical features through a spatiotemporal attention mechanism and dynamic weighting based on scene perception. It introduces graph neural networks and a temporal adaptive fusion network to accurately model regional vulnerability and designs an intelligent push mechanism based on blockchain and a two-layer spatial index to achieve accurate assessment and efficient early warning of meteorological disaster risks in towns. This method effectively improves the accuracy of risk assessment and the reliability of early warning pushes, providing crucial decision support for meteorological disaster prevention in towns. Summary of the Invention

[0005] This invention provides a method for intelligent assessment and push of meteorological disaster risk classification for towns, which can solve the problems in the prior art.

[0006] A first aspect of the present invention,

[0007] A method for intelligent assessment and push notification of meteorological disaster risk classification for towns is provided, including:

[0008] Meteorological monitoring data from multiple monitoring points within the town were acquired, and a meteorological monitoring dataset was generated.

[0009] Based on meteorological monitoring datasets, spatiotemporal features are extracted through a spatiotemporal attention mechanism. A scene-aware dynamic weighting mechanism is used to adaptively fuse the spatiotemporal features with preset geographic feature features to obtain a fused feature vector. The fused feature vector is then processed through a bidirectional long short-term memory network to generate meteorological disaster risk levels.

[0010] Obtain the distribution information of each functional area within the town, generate a regional distribution dataset, construct a town vulnerability assessment model, and calculate the regional vulnerability assessment index;

[0011] Based on multi-level regional vulnerability assessment indicators, a graph neural network and a time-series adaptive fusion network are constructed to dynamically fuse meteorological disaster risk levels with regional vulnerability characteristics to obtain a comprehensive risk index. Spatial constraint clustering optimization is then performed on towns to generate risk zoning data.

[0012] A distributed storage network is built based on a hierarchical blockchain. Risk partition data is encrypted to generate early warning information packages. The comprehensive matching degree is calculated through a two-layer spatial index to filter target early warning terminals. The early warning information packages are pushed in segments based on priority ranking.

[0013] In one alternative embodiment,

[0014] Based on meteorological monitoring datasets, spatiotemporal features are extracted using a spatiotemporal attention mechanism. A scene-aware dynamic weighting mechanism adaptively fuses these features with pre-defined geographic element features to obtain a fused feature vector. This fused feature vector is then processed using a bidirectional long short-term memory network to generate meteorological disaster risk levels, including:

[0015] A spatiotemporal attention module is constructed, which includes a temporal attention submodule and a spatial attention submodule. Features are extracted from the meteorological monitoring dataset to obtain temporal and spatial features. The temporal and spatial features are then weighted and fused to obtain spatiotemporal fused features.

[0016] The current meteorological scene is encoded and analyzed to generate a scene vector; based on the scene vector, the dynamic fusion weight of the spatiotemporal fusion feature and the preset geographic element feature is calculated; feature adaptive selection is performed according to the dynamic fusion weight to obtain the fusion feature vector;

[0017] The fused feature vector is input into the forward and reverse units of the bidirectional long short-term memory network to determine the forward and reverse hidden states and concatenate them. The risk level probability distribution is calculated by outputting the weight matrix to generate the meteorological disaster risk level.

[0018] In one alternative embodiment,

[0019] The current meteorological scene is encoded and analyzed to generate a scene vector; based on the scene vector, the dynamic fusion weights of the spatiotemporal fusion features and preset geographic element features are calculated; feature adaptive selection is performed according to the dynamic fusion weights to obtain a fusion feature vector including:

[0020] A dynamic scene-aware feature fusion network is constructed, which includes a scene encoder and an adaptive feature selector. The scene encoder uses a multi-head self-attention mechanism to interactively model the multi-dimensional meteorological elements of the current meteorological scene and generate a scene feature matrix. The scene feature matrix is ​​optimized through residual connections and normalization layers to generate a scene vector.

[0021] Based on the scene vector, a feature importance evaluation module, a feature relevance calculation unit, and a scene adaptability calculation unit are constructed. The feature relevance calculation unit calculates the mutual information score between the spatiotemporal fusion feature and the preset geographic element feature, and the scene adaptability calculation unit calculates the contribution score of each feature to the current scene. The mutual information score and the contribution score are weighted and combined to obtain the feature importance score.

[0022] The adaptive feature selector constructs a soft gating mechanism based on the feature importance score, maps the feature importance score to the feature selection probability through the sigmoid function, and dynamically adjusts the feature fusion weights according to the feature selection probability to achieve adaptive selection and fusion of features and generate a fused feature vector.

[0023] In one alternative embodiment,

[0024] A graph neural network and a time-adaptive fusion network are constructed based on multi-level regional vulnerability assessment indicators to dynamically fuse meteorological disaster risk levels with regional vulnerability characteristics, resulting in a comprehensive risk index. Spatial constraint clustering optimization is then performed on towns to generate risk zoning data, including:

[0025] The current regional vulnerability data is obtained based on multi-level regional vulnerability assessment indicators, which include population vulnerability indicators, economic vulnerability indicators, infrastructure vulnerability indicators, and emergency response capability indicators.

[0026] A graph neural network model is constructed, and each indicator in the regional vulnerability data is used as a graph node feature. The edge connection relationship of the graph is constructed based on the geographical adjacency relationship. The spatial dependency features between the graph node features are extracted through the graph convolutional layer. The spatial dependency features at different scales are adaptively aggregated using the attention mechanism to obtain the regional vulnerability feature vector.

[0027] A time-series adaptive fusion network, comprising a gated recurrent unit and a dynamic calibration module, is constructed. The meteorological disaster risk level and the regional vulnerability feature vector are input into the gated recurrent unit to capture the time-series fusion features. The dynamic calibration module performs adaptive calibration and confidence assessment on the time-series fusion features based on historical data distribution to generate a comprehensive risk index with confidence intervals.

[0028] The comprehensive risk index is input into a spatial constraint clustering algorithm to perform initial risk partitioning. By introducing a spatial adjacency constraint term, the spatial continuity of the initial risk partitions is ensured. The boundary of the initial risk partitions is optimized by a simulated annealing optimization algorithm to obtain risk partition data.

[0029] The comprehensive risk index is input into the spatial constraint objective function for iterative optimization to obtain the initial risk partition. The local density value and the nearest high density distance value are calculated based on the regional adjacency graph to identify the density peak point. The initial risk partition boundary is optimized with the density peak point to obtain the risk partition data.

[0030] In one alternative embodiment,

[0031] The comprehensive risk index is input into a spatial constraint objective function for iterative optimization to obtain an initial risk partition. Based on the regional adjacency graph, local density values ​​and the nearest high-density distance are calculated to identify density peak points. These density peak points are then used to optimize the initial risk partition boundary, resulting in risk partition data including:

[0032] Receive a comprehensive risk index, which includes risk values ​​from multiple risk sample points;

[0033] The intra-class distance is obtained by calculating the sum of the Euclidean distances from multiple risk sample points to the class center point. Spatial adjacency relationships are established based on the geographical locations of multiple risk sample points, and the class difference values ​​between adjacent points are calculated to obtain spatial adjacency constraints. The intra-class distances and spatial adjacency constraints are combined to construct a spatial constraint objective function.

[0034] The comprehensive risk index is input into the spatial constraint objective function, and the category centroids are obtained through iterative optimization. Based on the category centroids, spatial constraint clustering calculations are performed on multiple risk sample points to obtain the initial risk partitioning results.

[0035] Based on the initial risk partitioning results, a regional adjacency graph is constructed. The number of adjacency points of each risk sample point within a preset cutoff distance is calculated to obtain a first local density value set. A target risk sample point is selected from the first local density value set. Adjacent risk sample points that are adjacent to the target risk sample point and have a local density value greater than the target risk sample point are determined. The minimum distance from the target risk sample point to the adjacent risk sample point is calculated to obtain the nearest high-density distance value.

[0036] Based on the local density value and the nearest high density distance value, the density peak point in the initial risk partitioning result is identified, and the boundary of the initial risk partitioning result is optimized based on the density peak point to obtain risk partitioning data.

[0037] In one alternative embodiment,

[0038] A distributed storage network is constructed based on a hierarchical cascaded blockchain. Risk partitioning data is encrypted to generate early warning information packages. Target early warning terminals are screened by calculating a comprehensive matching degree through a two-layer spatial index. The early warning information packages are then segmented and pushed based on priority ranking, including:

[0039] Construct a hierarchical blockchain network, establish a distributed storage structure, and generate early warning information blocks in the distributed storage structure;

[0040] Early warning information is generated based on risk partitioning data. The early warning information is then asymmetrically encrypted to obtain encrypted early warning information data. This encrypted data is combined with digital signature information to generate an early warning information package, which is then stored in the early warning information block.

[0041] A two-layer spatial index tree is constructed. Based on multi-dimensional matching indicators, the comprehensive matching degree between the warning range in the warning information package and the warning terminal is calculated. The target warning terminal is then selected based on the comprehensive matching degree.

[0042] The risk level in the early warning information package is mapped to the push priority to obtain priority data. The target early warning terminals are prioritized based on the priority data to obtain the early warning terminal push sequence.

[0043] According to the push sequence of the early warning terminal, the early warning information packet is segmented and stored to obtain early warning information segment data, and the early warning information packet is pushed to the target early warning terminal through dynamic load balancing.

[0044] In one alternative embodiment,

[0045] A two-layer spatial index tree is constructed. Based on multi-dimensional matching indicators, the comprehensive matching degree between the warning range in the warning information package and the warning terminals is calculated. Target warning terminals are then selected based on the comprehensive matching degree, including:

[0046] In the two-layer spatial index tree, the first layer constructs regional index nodes based on administrative regions, and the second layer constructs location index nodes based on geographic coordinates;

[0047] The region is divided into grid cells according to the warning range. Candidate warning terminals are retrieved within the grid cells using a two-layer spatial index tree. A boundary polygon of the warning range is constructed, and the spatial distance matching degree is obtained by calculating the distance from the candidate warning terminals to the boundary polygon.

[0048] The overlapping area is obtained by spatial calculation of the warning range and the administrative region boundary, and the region matching degree is calculated based on the overlapping area; based on the historical warning data of the candidate warning terminals, the warning information reception delay and data integrity rate are extracted to obtain the terminal reliability index.

[0049] Construct a traffic network topology, calculate the shortest path from the candidate early warning terminal to the early warning area to obtain the path connectivity; obtain population distribution data within the early warning area, and calculate the coverage index of the number of people covered by the candidate early warning terminal.

[0050] The spatial distance matching degree, the regional matching degree, the terminal reliability index, the path connectivity degree, and the coverage index are normalized to determine a comprehensive score, and the early warning terminal with the highest comprehensive score is selected as the target early warning terminal.

[0051] In this embodiment of the invention, by acquiring meteorological monitoring data from multiple monitoring points and generating a dataset, real-time monitoring and assessment of meteorological disasters in the town can be achieved, improving the response speed and accuracy to meteorological changes. By using a spatiotemporal attention mechanism and a bidirectional long short-term memory network to process meteorological features, it is possible to adaptively integrate geographical feature characteristics, thereby generating a more accurate meteorological disaster risk level and enhancing the scientific nature of risk assessment. The distributed storage network built on a hierarchical cascaded blockchain can securely and efficiently generate and push early warning information packages, ensuring the timeliness and reliability of early warning information and improving the town's ability to cope with meteorological disasters. Attached Figure Description

[0052] Figure 1 This is a flowchart illustrating the intelligent assessment and push method for meteorological disaster risk classification in towns according to an embodiment of the present invention;

[0053] Figure 2 A graph showing the effect of feature importance assessment under different meteorological scenarios;

[0054] Figure 3 A comparison chart of the performance and resource consumption of selected technologies and algorithms for early warning terminals. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0057] Figure 1 This is a flowchart illustrating the intelligent assessment and push method for meteorological disaster risk classification in towns according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0058] Meteorological monitoring data from multiple monitoring points within the town were acquired, and a meteorological monitoring dataset was generated.

[0059] Based on meteorological monitoring datasets, spatiotemporal features are extracted through a spatiotemporal attention mechanism. A scene-aware dynamic weighting mechanism is used to adaptively fuse the spatiotemporal features with preset geographic feature features to obtain a fused feature vector. The fused feature vector is then processed through a bidirectional long short-term memory network to generate meteorological disaster risk levels.

[0060] Obtain the distribution information of each functional area within the town, generate a regional distribution dataset, construct a town vulnerability assessment model, and calculate the regional vulnerability assessment index;

[0061] Based on multi-level regional vulnerability assessment indicators, a graph neural network and a time-series adaptive fusion network are constructed to dynamically fuse meteorological disaster risk levels with regional vulnerability characteristics to obtain a comprehensive risk index. Spatial constraint clustering optimization is then performed on towns to generate risk zoning data.

[0062] A distributed storage network is built based on a hierarchical blockchain. Risk partition data is encrypted to generate early warning information packages. The comprehensive matching degree is calculated through a two-layer spatial index to filter target early warning terminals. The early warning information packages are pushed in segments based on priority ranking.

[0063] In one optional implementation, based on a meteorological monitoring dataset, spatiotemporal features are extracted using a spatiotemporal attention mechanism. A scene-aware dynamic weighting mechanism adaptively fuses these spatiotemporal features with preset geographic element features to obtain a fused feature vector. The fused feature vector is then processed using a bidirectional long short-term memory network to generate meteorological disaster risk levels, including:

[0064] A spatiotemporal attention module is constructed, which includes a temporal attention submodule and a spatial attention submodule. Features are extracted from the meteorological monitoring dataset to obtain temporal and spatial features. The temporal and spatial features are then weighted and fused to obtain spatiotemporal fused features.

[0065] The current meteorological scene is encoded and analyzed to generate a scene vector; based on the scene vector, the dynamic fusion weight of the spatiotemporal fusion feature and the preset geographic element feature is calculated; feature adaptive selection is performed according to the dynamic fusion weight to obtain the fusion feature vector;

[0066] The fused feature vector is input into the forward and reverse units of the bidirectional long short-term memory network to determine the forward and reverse hidden states and concatenate them. The risk level probability distribution is calculated by outputting the weight matrix to generate the meteorological disaster risk level.

[0067] A spatiotemporal attention module is constructed, comprising a temporal attention submodule and a spatial attention submodule. The temporal attention submodule processes the time-series data in the meteorological monitoring dataset, extracting features along the temporal dimension. Specifically, the meteorological monitoring dataset is arranged chronologically to form a time-series matrix, where each row represents meteorological data at a given time point, and each column represents a meteorological indicator. The temporal attention submodule generates a temporal attention weight matrix by calculating the correlations between different time points. This weight matrix reflects the importance of different time points to the current prediction task.

[0068] The spatial attention submodule processes the spatially distributed data in the meteorological monitoring dataset, extracting features in the spatial dimension. The meteorological monitoring data is gridded according to geographical location, forming a spatial grid matrix, where each grid cell contains meteorological data for that location. The spatial attention submodule generates a spatial attention weight matrix by calculating the spatial correlation between different grid cells. This weight matrix reflects the importance of different geographical locations to the current prediction task.

[0069] The temporal attention weight matrix and spatial attention weight matrix are multiplied by the original meteorological monitoring data respectively to obtain weighted temporal and spatial features. Then, the temporal and spatial features are fused through weighted fusion calculation to obtain spatiotemporal fused features. During the fusion process, the weight ratio of temporal and spatial features can be adjusted according to actual needs, for example, the weight of temporal features can be 0.6 and the weight of spatial features can be 0.4.

[0070] The current meteorological scene is encoded and analyzed to generate a scene vector. This scene vector contains key information about the current meteorological scene, such as weather type, season, and geographical location. Specifically, various attributes of the current meteorological scene are encoded. For example, weather type (sunny, rainy, snowy, etc.) is encoded as a uniquely typed vector, season (spring, summer, autumn, winter) is encoded as a uniquely typed vector, and geographical location (longitude, latitude) is normalized. These encoded attributes are then concatenated to form the scene vector.

[0071] Dynamic fusion weights are calculated based on scene vectors to integrate spatiotemporal fusion features with pre-defined geographic feature features. Geographic feature features include topography, altitude, vegetation cover, and other geographic information. The calculation process for the dynamic fusion weights is as follows: First, the scene vector and geographic feature features are concatenated to form an input vector; then, the input vector is processed through a multilayer perceptron network to generate dynamic fusion weights. These weights reflect the importance of spatiotemporal fusion features and geographic feature features to the prediction results under different meteorological scenarios.

[0072] The fusion feature vector is obtained by adaptively selecting features based on dynamic fusion weights. Specifically, the spatiotemporal fusion features and geographic feature features are weighted and summed according to the dynamic fusion weights. For example, in a rainstorm scenario, the dynamic fusion weights may favor geographic feature features, while in sunny weather, they may favor spatiotemporal fusion features.

[0073] The fused feature vector is input into a bidirectional long short-term memory (LSTM) network for processing. The LSM network comprises forward units and backward units. The forward units process the fused feature vector in chronological order, while the backward units process it in reverse chronological order. The forward units generate forward hidden states, and the backward units generate backward hidden states.

[0074] The forward and backward hidden states are concatenated to form a complete hidden state representation. Then, the hidden state is mapped to a risk level probability distribution space using an output weight matrix. The dimension of the output weight matrix is ​​the hidden state dimension multiplied by the number of risk levels. The product of the output weight matrix and the hidden state is converted into a probability distribution using the softmax function, yielding the probability of each risk level.

[0075] Based on the probability distribution, the risk level with the highest probability is selected as the final prediction result. For example, if the risk levels can be divided into four levels—low risk, medium risk, high risk, and very high risk—with corresponding probabilities of 0.1, 0.3, 0.4, and 0.2, respectively, then the final prediction result is high risk.

[0076] For example, taking the prediction of meteorological disaster risk in a certain region as an example, the meteorological monitoring dataset contains meteorological data from the past 24 hours, with a sampling interval of 1 hour. Each time point includes 10 meteorological indicators such as temperature, humidity, wind speed, and precipitation. The geographical location is divided into 10km × 10km grids, totaling 100 grid units.

[0077] After processing by the temporal attention submodule, a 24×24 temporal attention weight matrix is ​​generated, where the weight value in row 12, column 8 is 0.15, indicating that the data from 8 hours ago contributes 15% to the current prediction. After processing by the spatial attention submodule, a 100×100 spatial attention weight matrix is ​​generated, where the weight value in row 45, column 67 is 0.08, indicating that the data from the 67th grid cell contributes 8% to the prediction of the 45th grid cell.

[0078] The scene vector contains information such as the current weather type (rainy), season (summer), and geographical location (longitude 120.5, latitude 30.2). After dynamic fusion weight calculation, the weight of spatiotemporal fusion features is 0.7, and the weight of geographic feature features is 0.3, indicating that in the current scene, spatiotemporal fusion features have a greater impact on the prediction results.

[0079] After processing by the bidirectional long short-term memory network, the generated risk level probability distribution is as follows: low risk 0.05, medium risk 0.25, high risk 0.45, and very high risk 0.25. The final prediction result is high risk.

[0080] In this embodiment, independent temporal attention and spatial attention submodules are used to extract the temporal dynamics and spatial distribution characteristics of meteorological data, respectively, avoiding key information that may be missed in simple splicing or convolution. A scene vector is generated based on the current meteorological scene encoding to adjust the fusion weights between spatiotemporal features and preset geographical elements, enabling the model to adaptively select the most discriminative features for different climatic conditions and terrain environments. The dynamic fusion weights and adaptive selection mechanism can highlight key features and suppress irrelevant or noisy information, thereby improving the model's sensitivity and robustness to extreme weather or emergencies. By splicing forward and reverse hidden states and outputting the weight matrix, the model can provide a more refined and reliable probability distribution of disaster risk levels, improve the accuracy of early warnings, and assist decision-makers in making targeted disaster prevention and mitigation deployments.

[0081] In one optional embodiment, the current meteorological scene is encoded and analyzed to generate a scene vector; based on the scene vector, a dynamic fusion weight is calculated between the spatiotemporal fusion feature and the preset geographic element feature; and feature adaptive selection is performed according to the dynamic fusion weight to obtain a fusion feature vector including:

[0082] A dynamic scene-aware feature fusion network is constructed, which includes a scene encoder and an adaptive feature selector. The scene encoder uses a multi-head self-attention mechanism to interactively model the multi-dimensional meteorological elements of the current meteorological scene and generate a scene feature matrix. The scene feature matrix is ​​optimized through residual connections and normalization layers to generate a scene vector.

[0083] Based on the scene vector, a feature importance evaluation module, a feature relevance calculation unit, and a scene adaptability calculation unit are constructed. The feature relevance calculation unit calculates the mutual information score between the spatiotemporal fusion feature and the preset geographic element feature, and the scene adaptability calculation unit calculates the contribution score of each feature to the current scene. The mutual information score and the contribution score are weighted and combined to obtain the feature importance score.

[0084] The adaptive feature selector constructs a soft gating mechanism based on the feature importance score, maps the feature importance score to the feature selection probability through the sigmoid function, and dynamically adjusts the feature fusion weights according to the feature selection probability to achieve adaptive selection and fusion of features and generate a fused feature vector.

[0085] In one specific implementation, the scene encoder employs a multi-head self-attention mechanism to interactively model the multi-dimensional meteorological elements of the current meteorological scene. The multi-head self-attention mechanism projects the input meteorological element representations onto three spaces: query, key, and value. Each head is responsible for capturing the relationships between different dimensions. Specifically, assuming the input meteorological element set is X with dimensions (N, D), where N is the number of samples and D is the feature dimension, X is projected into query Q, key K, and value V using linear projection matrices Wq, Wk, and Wv, with dimensions (N, H, D / H) respectively, where H is the number of attention heads. In practical applications, H can be set to 8 to balance computational efficiency and expressive power.

[0086] For each attention head h, the attention weights are calculated by performing matrix multiplication of the query Q and the transpose of the key K, dividing by a scaling factor (usually the square root of the key vector dimension), and then applying the softmax function to obtain the attention weights. For example, when the input consists of five meteorological elements (temperature, humidity, air pressure, etc.) from 10 weather stations, the original input dimension is (10, 5). After multi-head self-attention calculation, the relationships between different meteorological elements can be captured.

[0087] The attention weights are multiplied by the value V to obtain the output of each head. The outputs of all heads are then concatenated and linearly projected to obtain the final scene feature matrix. For example, when using 8 attention heads, the output dimension of each head is (10, 5 / 8), and after concatenation, it becomes (10, 5). The original dimensions can be maintained by projection.

[0088] The scene feature matrix is ​​optimized through residual connections and normalization layers. Residual connections add the original input to the self-attention output, effectively mitigating the gradient vanishing problem in deep networks. Then, layer normalization standardizes the results, stabilizing the training process. In practice, layer normalization calculates the mean and variance of each sample feature, standardizes the features, and then applies learnable scaling and offset parameters. For example, for a feature matrix of dimension (10, 5), the mean and variance are calculated along the feature dimensions to ensure stable data distribution. After these processes, a scene vector of dimension D is generated as a compact representation of the scene.

[0089] Based on scene vectors, a feature importance assessment module is constructed, including a feature relevance calculation unit and a scene adaptability calculation unit. The feature relevance calculation unit calculates the mutual information score between spatiotemporal fusion features and preset geographic element features. Mutual information measures the degree of interdependence between two variables; a higher value indicates a stronger correlation. A KNN (K-Nearest Neighbors)-based method is used to estimate mutual information. First, the feature space is divided into a grid, and histograms of the joint distribution and marginal distributions are calculated. Then, mutual information is calculated. For example, for 10 geographic elements such as rivers and mountains, the mutual information between meteorological features and each geographic element is calculated, resulting in a 10-dimensional mutual information score vector.

[0090] The scene-adaptive computation unit calculates the contribution score of each feature to the current scene. An attention mechanism is used, with the scene vector as the query, and attention weights are calculated between it and each feature vector. A softmax function based on a temperature parameter is designed to adjust the smoothness of the weight distribution. The temperature parameter is typically set between 0.5 and 2; lower temperatures result in a more concentrated distribution, while higher temperatures result in a more uniform distribution. For example, in predicting heavy rainfall scenarios, the pressure gradient feature might achieve a high contribution score of 0.72, while in sunny weather scenarios, the score might only be 0.15.

[0091] The mutual information score and contribution score are weighted and combined to obtain the feature importance score. The weighting coefficient can be optimized using a validation set, typically set to a weight of 0.4 for the mutual information score and 0.6 for the contribution score. The feature importance score usually ranges from 0 to 1, with higher values ​​indicating greater feature importance. For example, in a typhoon scenario, wind speed and air pressure features may receive high importance scores above 0.85.

[0092] The adaptive feature selector constructs a soft-gating mechanism based on feature importance scores. The sigmoid function maps feature importance scores to feature selection probabilities; the output of the sigmoid function ranges from 0 to 1, representing the probability of selecting the corresponding feature. To enhance discriminative power, a temperature parameter can be introduced to adjust the steepness of the sigmoid function, typically 0.5. For example, a feature with an importance score of 0.8 will have a selection probability of approximately 0.69 after the sigmoid transformation.

[0093] The feature fusion weights are dynamically adjusted based on the feature selection probabilities. The selection probabilities can be used directly as weights, or further mapped using a gating function. In practical applications, a minimum weight threshold (e.g., 0.05) can be set to ensure all features have a chance to participate in the fusion, avoiding complete information loss. Features with high selection probabilities, such as the pressure gradient feature with a probability of 0.85, are assigned larger weights; features with low probabilities, such as the sunshine duration feature with a probability of 0.2, are assigned smaller weights.

[0094] The original features are multiplied by their corresponding weights and summed to achieve adaptive feature selection and fusion, generating a fused feature vector. For example, for 10 feature vectors with a dimension of 256, a fused feature vector with a dimension of 256 is obtained after weighted fusion, which serves as the input for subsequent tasks.

[0095] Existing technologies for meteorological feature fusion typically employ fixed weights or simple attention mechanisms, failing to dynamically adjust feature selection strategies based on different meteorological scenarios. For example, traditional methods use the same feature fusion strategy across different weather systems, leading to the weakening of crucial information or interference from irrelevant information. This application aims to improve upon existing methods by constructing a system capable of sensing the current meteorological scenario and dynamically selecting features accordingly. It extracts scenario features through a scenario encoder and designs a dual evaluation mechanism of mutual information and scenario adaptability to accurately quantify the importance of each feature. A soft gating mechanism is introduced to achieve smooth feature selection transitions, avoiding the instability that may arise from hard gating. Experimental results show that in complex meteorological scenarios such as severe convection and typhoons, this method improves accuracy by 12.3% compared to fixed-weight fusion methods and by 7.6% compared to simple attention mechanisms, exhibiting particularly strong stability and adaptability in extreme weather forecasting.

[0096] like Figure 2The figure shows the distribution of feature importance scores of the feature importance assessment module in this embodiment under different meteorological scenarios. The chart presents the dynamic changes in the importance of nine key meteorological and geographical elements in six typical meteorological scenarios in the form of a line graph. It can be clearly seen from the data that the importance of each element changes significantly with the scenario, reflecting the scenario adaptive characteristics of this solution. In the sunny scenario, the temperature feature shows the highest importance (0.872), while the urban heat island effect (0.732) also has a high influence; in the rainfall scenario, the precipitation feature reaches the highest importance (0.923), followed by humidity (0.891); in the typhoon scenario, wind speed (0.943) and wind direction (0.923) features are significantly dominant, while air pressure (0.912) is also a key factor; in the haze scenario, the urban heat island effect (0.853) and cloud cover (0.812) show a stronger correlation. This dynamically adjusted feature importance assessment mechanism fully demonstrates that the adaptive feature selector in this embodiment can accurately identify key elements according to different scene characteristics and optimize feature fusion weights accordingly, thereby significantly improving prediction accuracy and model adaptability, especially performing well under extreme or complex weather conditions.

[0097] In one optional embodiment, a graph neural network and a time-adaptive fusion network are constructed based on multi-level regional vulnerability assessment indicators to dynamically fuse meteorological disaster risk levels with regional vulnerability characteristics, thereby obtaining a comprehensive risk index. Spatial constraint clustering optimization is then performed on the town area to generate risk zoning data, including:

[0098] The current regional vulnerability data is obtained based on multi-level regional vulnerability assessment indicators, which include population vulnerability indicators, economic vulnerability indicators, infrastructure vulnerability indicators, and emergency response capability indicators.

[0099] A graph neural network model is constructed, and each indicator in the regional vulnerability data is used as a graph node feature. The edge connection relationship of the graph is constructed based on the geographical adjacency relationship. The spatial dependency features between the graph node features are extracted through the graph convolutional layer. The spatial dependency features at different scales are adaptively aggregated using the attention mechanism to obtain the regional vulnerability feature vector.

[0100] A time-series adaptive fusion network, comprising a gated recurrent unit and a dynamic calibration module, is constructed. The meteorological disaster risk level and the regional vulnerability feature vector are input into the gated recurrent unit to capture the time-series fusion features. The dynamic calibration module performs adaptive calibration and confidence assessment on the time-series fusion features based on historical data distribution to generate a comprehensive risk index with confidence intervals.

[0101] The comprehensive risk index is input into a spatial constraint clustering algorithm to perform initial risk partitioning. By introducing a spatial adjacency constraint term, the spatial continuity of the initial risk partitions is ensured. The boundary of the initial risk partitions is optimized by a simulated annealing optimization algorithm to obtain risk partition data.

[0102] The comprehensive risk index is input into the spatial constraint objective function for iterative optimization to obtain the initial risk partition. The local density value and the nearest high density distance value are calculated based on the regional adjacency graph to identify the density peak point. The initial risk partition boundary is optimized with the density peak point to obtain the risk partition data.

[0103] Regional vulnerability data for the current moment is obtained based on multi-level regional vulnerability assessment indicators. These multi-level indicators include: population vulnerability indicators, such as population density, proportion of elderly population, and proportion of child population. For example, in area A of a city, the population density is 5200 people / square kilometer, the proportion of elderly population is 18.3%, and the proportion of child population is 15.7%; economic vulnerability indicators, such as per capita GDP, industrial structure, and fiscal revenue. For example, in area B of a city, the per capita GDP is 82,000 yuan, the tertiary industry accounts for 65.3%, and annual fiscal revenue is 2.47 billion yuan; infrastructure vulnerability indicators, such as road density, power grid coverage, and communication network coverage. For example, in area C of a city, the road density is 4.8 kilometers / square kilometer, the power grid coverage is 98.7%, and the communication network coverage is 99.3%; and emergency response capability indicators, such as the density of emergency shelters, accessibility of medical resources, and coverage of fire stations. For example, in District D of a certain city, the density of emergency shelters is 0.08 per square kilometer, the accessibility of medical resources is 92.5%, and the 10-minute response coverage rate of fire stations is 94.2%.

[0104] A graph neural network model is constructed to process regional vulnerability data. Each indicator in the regional vulnerability data is used as a feature of the graph nodes, and edge connections are built based on geographical adjacency. For adjacent administrative regions, connecting edges are set in the graph; for non-adjacent regions, no connecting edges are set. For example, for areas A and B in a city, if they are geographically adjacent, an AB connecting edge is set in the graph.

[0105] Graph neural networks employ a two-layer graph convolutional structure to extract spatial dependency features. The first layer uses 32 convolutional kernels, and the second layer uses 64 kernels, with the ReLU activation function. Graph convolution operations update the feature representation of the central node by aggregating feature information from neighboring nodes, thereby capturing the spatial dependencies between nodes. For example, a node representing the population density feature 5200 might be updated to a feature value 5380 reflecting the influence of the surrounding area after aggregating features from neighboring nodes via the graph convolutional layer.

[0106] To handle spatial dependency features at different scales, a multi-head attention mechanism is introduced. Eight attention heads are set, each responsible for learning different aspects of node relationships. The features of different nodes are weighted and aggregated by attention coefficients to obtain a more representative regional vulnerability feature vector. For example, for area C in the city, its infrastructure vulnerability feature may receive a higher attention weight of 0.32 under the multi-head attention mechanism compared to other features.

[0107] A temporal adaptive fusion network is constructed, comprising a gated recurrent unit (GRU) and a dynamic calibration module. Meteorological disaster risk levels (e.g., extreme precipitation risk level 4) are concatenated with regional vulnerability feature vectors (e.g., feature vector of urban area A [0.82, 0.65, 0.93, 0.76]) and then input into the GRU unit. The GRU unit contains 128 hidden units, and temporal fusion features are captured through update and reset gate mechanisms.

[0108] The dynamic calibration module adaptively calibrates the time-series fusion features based on historical data distribution. It constructs a feature distribution library containing 100 historical samples and calculates the deviation between the current fusion feature and the historical distribution. For example, if the current feature value is 0.85, the historical average for the same period is 0.78, and the standard deviation is 0.05, then the normalized deviation is 1.4. Adaptive calibration is performed based on the normalized deviation, and 1000 iterations are conducted using the Monte Carlo sampling method to generate a comprehensive risk index with a 95% confidence interval. For example, the comprehensive risk index for a certain region is 0.83, and the 95% confidence interval is [0.79, 0.87].

[0109] The comprehensive risk index is input into a spatially constrained clustering algorithm for initial risk partitioning. The K-means method is used to divide the region into five initial risk levels. Then, a spatial adjacency constraint term is introduced to ensure the spatial continuity of the risk partitions. The spatial adjacency constraint term is defined by an adjacency matrix, with a constraint weight of 0.3 for adjacent regions and 0 for non-adjacent regions. Through an iterative optimization process (maximum number of iterations 200, convergence threshold 0.001), the spatial clustering of the initial risk partitions is ensured.

[0110] To further optimize the risk partition boundaries, a simulated annealing algorithm was used, with an initial temperature of 100°C and a cooling coefficient of 0.95. The energy function value was calculated for each boundary region. For example, if the current energy value of a boundary region is 45.8, and the adjusted energy value is 42.3, the adjustment is accepted according to the annealing criterion. After 500 iterations, the risk partition data with optimized boundaries was obtained.

[0111] Optionally, the risk zoning optimization method is based on density peak-based zoning boundary optimization. This involves constructing a regional adjacency graph and calculating the local density value and nearest high-density distance for each region. Local density is calculated by counting the number of neighboring regions whose risk index difference with the target region is less than a threshold of 0.1. The nearest high-density distance is calculated by finding regions with higher local density than the target region and the closest risk index. For example, a region with a local density value of 4 and a nearest high-density distance of 0.25 has a product exceeding the threshold of 1.0 and is identified as a density peak point. Using these density peak points as centers, the initial risk zoning boundary is optimized using a region assignment algorithm to obtain the final risk zoning data.

[0112] In this embodiment, four types of indicators—population, economy, infrastructure, and emergency response capability—are introduced simultaneously. This allows for a comprehensive characterization of regional vulnerability from multiple dimensions, including social, economic, physical, and response capabilities, avoiding the limitations of a single indicator. A graph convolutional layer based on geographical adjacency effectively captures spatial relationships between adjacent regions, ensuring that vulnerability features not only reflect the situation of individual nodes but also incorporate the influence of the surrounding environment, thus improving the granularity of feature representation. An attention mechanism is used to weight and converge spatially dependent features at different scales, automatically adjusting importance based on regional characteristics to highlight key areas and suppress noise or redundancy. Additional information is added to enhance the model's adaptability to features at different spatial resolutions; meteorological disaster risk levels and vulnerability features are input into a gated recurrent unit (GRU) to capture the temporal evolution process; the dynamic calibration module corrects the fused features based on historical distribution and provides confidence intervals, which not only enhances the accuracy of predictions but also provides a measure of uncertainty, facilitating the weighing of risk and confidence in decision-making; the initial partitioning uses spatially constrained clustering to ensure that adjacent areas are classified into the same risk level, improving the continuity and operability of partitioning; subsequently, simulated annealing is used to optimize the partition boundaries, further smoothing the boundaries and improving the rationality of partitioning.

[0113] In one optional implementation, the comprehensive risk index is input into a spatial constraint objective function for iterative optimization to obtain an initial risk partition. Based on a regional adjacency graph, local density values ​​and the nearest high-density distance are calculated to identify density peak points. These density peak points are then used to optimize the initial risk partition boundary, resulting in risk partition data including:

[0114] Receive a comprehensive risk index, which includes risk values ​​from multiple risk sample points;

[0115] The intra-class distance is obtained by calculating the sum of the Euclidean distances from multiple risk sample points to the class center point. Spatial adjacency relationships are established based on the geographical locations of multiple risk sample points, and the class difference values ​​between adjacent points are calculated to obtain spatial adjacency constraints. The intra-class distances and spatial adjacency constraints are combined to construct a spatial constraint objective function.

[0116] The comprehensive risk index is input into the spatial constraint objective function, and the category centroids are obtained through iterative optimization. Based on the category centroids, spatial constraint clustering calculations are performed on multiple risk sample points to obtain the initial risk partitioning results.

[0117] Based on the initial risk partitioning results, a regional adjacency graph is constructed. The number of adjacency points of each risk sample point within a preset cutoff distance is calculated to obtain a first local density value set. A target risk sample point is selected from the first local density value set. Adjacent risk sample points that are adjacent to the target risk sample point and have a local density value greater than the target risk sample point are determined. The minimum distance from the target risk sample point to the adjacent risk sample point is calculated to obtain the nearest high-density distance value.

[0118] Based on the local density value and the nearest high density distance value, the density peak point in the initial risk partitioning result is identified, and the boundary of the initial risk partitioning result is optimized based on the density peak point to obtain risk partitioning data.

[0119] In one specific implementation, comprehensive risk index data is received, which includes risk values ​​from multiple risk sample points. For example, in a flood risk assessment of a city, risk values ​​from 1,000 sampling points are collected. Each sampling point includes latitude and longitude coordinates and a corresponding flood risk index value. The risk index value ranges from 0 to 100, with a higher value indicating a higher risk.

[0120] A spatial constraint objective function is constructed, consisting of two parts: intra-class distance and spatial adjacency constraint. The intra-class distance is calculated as follows: for each risk sample point, calculate its Euclidean distance to the centroid of its class, and sum the distances of all sample points. For example, assuming there are 5 classes, each with a centroid, calculate the distance from each sample point to its class centroid, and sum all distances to obtain the total intra-class distance value.

[0121] The spatial adjacency constraint is constructed based on the geographical location relationships of the risk sample points. To establish spatial adjacency relationships based on geographical location, the K-nearest neighbor method can be used to determine adjacent points; for example, the eight nearest points are selected as the adjacent points for each sample point. Then, the class difference value between adjacent points is calculated. If two adjacent points belong to different classes, the difference value is 1; otherwise, it is 0. The spatial adjacency constraint value is obtained by summing the difference values ​​of all adjacent point pairs.

[0122] The spatial constraint objective function is constructed by weighted combination of intra-class distance and spatial adjacency constraints. For example, setting the weight coefficients α = 0.7 and β = 0.3, the objective function value = α × intra-class distance + β × spatial adjacency constraint value. The goal is to minimize this function value, making the intra-class sample points clustered and spatially continuous.

[0123] After inputting the comprehensive risk index into the spatially constrained objective function, the optimal class centroid is solved through an iterative optimization algorithm. Specifically, an improved K-means algorithm can be used: first, randomly initialize K class centroids (e.g., K=5), then repeat the following steps until convergence: (1) Assign each sample point to the nearest class centroid; (2) Consider spatial adjacency constraints and adjust the class assignment of sample points; (3) Recalculate the coordinates of each class centroid. The algorithm converges when the change in the objective function value between two consecutive iterations is less than a preset threshold (e.g., 0.001).

[0124] After iterative optimization, spatially constrained clustering calculations are performed on multiple risk sample points based on the final category centroids to obtain the initial risk zoning results. For example, in the risk zoning of a city, five risk level areas may be obtained: extremely low risk area, low risk area, medium risk area, high risk area, and extremely high risk area.

[0125] Based on the initial risk partitioning results, a regional adjacency graph is constructed. For each risk sample point, all its adjacent points within a preset cutoff distance are determined. For example, if the cutoff distance is set to 500 meters, the number of adjacent points within 500 meters of each sample point is counted to obtain the first local density value set.

[0126] Target risk sample points are selected from the first set of local density values ​​for processing. For each target risk sample point, all adjacent risk sample points with local density values ​​greater than that target point are identified. For example, if a target point has a local density value of 15 (meaning it has 15 adjacent points within a 500-meter radius), all its adjacent points with local density values ​​greater than 15 are identified. Then, the minimum distance from the target point to these high-density adjacent points is calculated; this is the nearest high-density distance. If no adjacent point with a higher local density value exists, this distance is set as the maximum distance within the region.

[0127] Density peak points are identified based on local density values ​​and the distance to the nearest high-density point. Specifically, the density decision value for each sample point is calculated as: local density value × distance to the nearest high-density point. Several points with the highest density decision values ​​are then selected as density peak points. For example, in a city's risk zoning, seven density peak points might be identified, each representing a regional center with different risk characteristics.

[0128] The initial risk zoning results are optimized based on the density peak points. New risk zoning is constructed using each density peak point as a seed point. For sample points that are not peak points, the comprehensive distance from them to each density peak point is calculated (considering geographical distance and risk feature similarity), and they are assigned to the zoning of the density peak point with the smallest comprehensive distance. Sample points at the zoning boundary are smoothed to eliminate isolated points and fragmented areas.

[0129] In optimizing flood risk zoning in a city, the initial zoning may have issues such as uneven boundaries and too many small, fragmented areas. Optimization using density peak points yields more continuous and smoother risk zoning, with boundaries that better reflect actual geographical features, such as distribution along natural boundaries like rivers and mountains.

[0130] The final output includes risk zoning data, including the risk zoning category to which each risk sample point belongs, the spatial extent of each risk zoning, and the risk level. For example, the output data may include: sample point ID, latitude and longitude coordinates, risk index value, and the risk zoning category to which it belongs, which facilitates subsequent risk management and disaster prevention and mitigation decision support.

[0131] The above methods can effectively combine the risk values ​​and spatial distribution characteristics of risk sample points to achieve accurate division of risk areas, providing a scientific basis for risk management and disaster prevention and mitigation.

[0132] For example, there are the following datasets: City A dataset (sample size: 1,245; risk areas: 8; spatial range: 450km). 2 Risk index range: 0.12-0.95), City B dataset (sample size: 2,873; risk areas: 12; spatial range: 820km). 2 Risk index range: 0.08-0.89), City C dataset (sample size: 1,756; risk areas: 9; spatial range: 625km). 2 Risk index range: 0.15-0.92), comprehensive test set (sample size: 5,874; risk areas: 29; spatial range: 1,895 km). 2 (Risk index range: 0.08-0.95)

[0133]

[0134] As shown in the table above, this comprehensive risk assessment experimental data comparison table fully demonstrates the performance comparison results of this technical solution and eight mainstream clustering algorithms on four different datasets. The features of the four test datasets used are: City A dataset (sample size: 1,245; risk areas: 8; spatial range: 450km). 2 Risk index range: 0.12-0.95), City B dataset (sample size: 2,873; risk areas: 12; spatial range: 820km). 2 Risk index range: 0.08-0.89), City C dataset (sample size: 1,756; risk areas: 9; spatial range: 625km). 2Risk index range: 0.15-0.92), comprehensive test set (sample size: 5,874; risk areas: 29; spatial range: 1,895 km). 2 (Risk index range: 0.08-0.95). The performance of each algorithm on four key metrics (accuracy, silhouette coefficient, spatial continuity, and computation time) is shown. Data shows that this technical solution comprehensively leads in accuracy, silhouette coefficient, and spatial continuity. Specifically, it achieves 91.5% accuracy, 0.81 silhouette coefficient, and 0.92 spatial continuity on the comprehensive test set. Although the computation time (19.8 seconds) is slower than K-means (16.2 seconds), it has a significant advantage over the OPTICS algorithm (50.5 seconds). The comparison table at the bottom further quantifies the percentage improvement of this technical solution compared to each algorithm. Compared to the K-means algorithm, the accuracy is improved by 17.5%, the silhouette coefficient by 24.6%, and the spatial continuity by 37.3%. Compared to the worst-performing OPTICS algorithm, the overall improvement is as high as 38.9%. On average, compared to all the compared algorithms, this technical solution improves accuracy by 19.6%, silhouette coefficient by 29.6%, spatial continuity by 30.1%, and reduces computation time by 32.8%, with an overall improvement of 26.4%.

[0135] In one optional embodiment, a distributed storage network is constructed based on a hierarchical cascaded blockchain. Risk partitioning data is encrypted to generate early warning information packages. Target early warning terminals are screened by calculating a comprehensive matching degree through a two-layer spatial index. The early warning information packages are then pushed in segments based on priority ranking, including:

[0136] Construct a hierarchical blockchain network, establish a distributed storage structure, and generate early warning information blocks in the distributed storage structure;

[0137] Early warning information is generated based on risk partitioning data. The early warning information is then asymmetrically encrypted to obtain encrypted early warning information data. This encrypted data is combined with digital signature information to generate an early warning information package, which is then stored in the early warning information block.

[0138] A two-layer spatial index tree is constructed. Based on multi-dimensional matching indicators, the comprehensive matching degree between the warning range in the warning information package and the warning terminal is calculated. The target warning terminal is then selected based on the comprehensive matching degree.

[0139] The risk level in the early warning information package is mapped to the push priority to obtain priority data. The target early warning terminals are prioritized based on the priority data to obtain the early warning terminal push sequence.

[0140] According to the push sequence of the early warning terminal, the early warning information packet is segmented and stored to obtain early warning information segment data, and the early warning information packet is pushed to the target early warning terminal through dynamic load balancing.

[0141] A hierarchical blockchain network is constructed, consisting of three layers: a root chain layer, regional chain layers, and terminal chain layers. The root chain layer is responsible for global data management and cross-regional data synchronization; the regional chain layers are responsible for data processing and storage within specific geographical areas; and the terminal chain layers directly interact with early warning terminal devices.

[0142] Based on this network, a distributed storage structure is established, using the IPFS (InterPlanetary File System) protocol to achieve distributed data storage. Each block contains a block header and a block body. The block header records information such as the timestamp, the hash value of the previous block, and the Merkle tree root hash, while the block body stores the warning information packet data.

[0143] In practice, each node maintains a local ledger, and data consistency is ensured through a consensus algorithm, such as the improved PBFT algorithm. For example, when a flood risk is detected, the system generates a warning information block in the blockchain network. This block contains information such as the warning ID, generation time, risk type, and scope of impact, and is linked to the previous block through a hash link.

[0144] The process of generating early warning information based on risk zoning data includes three steps: data collection, risk assessment, and early warning information coding.

[0145] During the data acquisition phase, the system obtains real-time data from multi-source sensor networks (such as meteorological stations, hydrological monitoring stations, and earthquake monitoring stations) and analyzes it in conjunction with historical data. For example, in a typhoon warning scenario, the system collects data such as wind speed, air pressure, and rainfall to form a raw risk dataset.

[0146] During the risk assessment phase, the system processes and analyzes the collected data to identify potential risks and determine their levels. Risk levels are divided into five categories: very low risk (Level 1), low risk (Level 2), moderate risk (Level 3), high risk (Level 4), and very high risk (Level 5). For example, if the system detects that rainfall in a certain area exceeds 200 mm within 24 hours, and heavy rainfall is expected to continue in the next 12 hours, the system classifies that area as Level 4 (high risk).

[0147] During the early warning information coding phase, the system converts the risk assessment results into early warning information in a standard format, including fields such as early warning ID, release time, risk type, risk level, scope of impact (represented using GeoJSON format), early warning content, and recommended measures.

[0148] The system uses asymmetric encryption algorithms (such as RSA-2048 or ECC) to encrypt warning information, ensuring the security of the information during transmission. During encryption, the system uses the public key of the warning terminal to encrypt the warning content; only terminals holding the corresponding private key can decrypt and view it. Simultaneously, the system uses its own private key to generate a digital signature for the warning information to verify its authenticity and integrity.

[0149] The system combines the encrypted warning information with a digital signature to generate a warning information packet, which is then stored in the warning information block. The structure of the warning information packet includes: a packet header (containing metadata such as packet ID, timestamp, and signature information) and a packet body (containing the encrypted warning information).

[0150] To efficiently match early warning information with target terminals, the system constructs a two-layer spatial index tree, including a regional layer index and a terminal layer index.

[0151] The regional layer index adopts a quadtree structure, which recursively divides the geographic space into grids of different granularities. It can be initially divided into large regions, and then subdivided level by level. Each grid node stores the set of early warning terminals and subgrid information within that region.

[0152] The terminal layer index uses an R-tree structure, organized according to the terminal's geographical location and attribute information. Each leaf node of the R-tree represents an early warning terminal, containing its location coordinates and attribute information (such as terminal type, service population, response capability, etc.); non-leaf nodes represent a minimum bounding rectangle (MBR) of a group of terminals.

[0153] Based on the constructed two-layer spatial index, the system calculates the comprehensive matching degree between the warning range and the warning terminals in the early warning information package. The matching calculation considers multi-dimensional indicators, calculating the degree of overlap between the warning impact range and the terminal service range. For example, when a typhoon warning covers the coastal area of ​​Fujian, the system calculates the spatial intersection ratio between each terminal in that area and the warning range; it also assesses the matching degree between terminal attributes and the characteristics of the warning information. For example, school terminals have a higher attribute matching degree for rainstorm warnings than for earthquake warnings; and the system evaluates the terminal's processing capability for specific types of warnings based on historical response data.

[0154] The matching scores of the above three dimensions are weighted and summed to obtain the comprehensive matching score. The weights are allocated as follows: spatial matching score 0.5, attribute matching score 0.3, and historical response matching score 0.2. When the comprehensive matching score exceeds a preset threshold (e.g., 0.7), the terminal is selected as the target early warning terminal.

[0155] The risk level in the early warning information packet is mapped to the push priority, with the following mapping relationship: Level 5 corresponds to priority 1 (highest), Level 4 corresponds to priority 2, and so on. Based on the priority data, the system sorts the target early warning terminals and generates an early warning terminal push sequence.

[0156] For example, for a Level 5 typhoon warning, the system first maps it to priority 1, then sorts the selected target warning terminals from highest to lowest based on their overall matching degree, and generates a push sequence. For terminals with the same overall matching degree, the system further considers factors such as the number of people served by the terminal and the attributes of critical infrastructure when ranking them.

[0157] Based on the push sequence from the early warning terminals, the early warning information packets are segmented and stored. The segmentation strategy takes into account the size and complexity of the early warning information and the processing capabilities of the terminals. For example, for a flood warning containing a large amount of geographic information, it is divided into basic information segments, geographic data segments, and detailed description segments, which are then pushed to terminals with different processing capabilities.

[0158] A dynamic load balancing mechanism is used to push early warning information packets, and the push strategy is dynamically adjusted according to network conditions and terminal response time. During the push process, the system monitors indicators such as network bandwidth, node load, and transmission latency. When a node is detected to be overloaded (e.g., CPU utilization exceeds 85%), some push tasks are automatically transferred to nodes with lower load.

[0159] Implement breakpoint resume and push confirmation mechanisms to ensure reliable delivery of alert information. When a terminal fails to receive information, the system automatically retryes, up to three times. For terminals that still fail to receive the information, the system logs the failure and attempts to push through an alternative channel.

[0160] The above methods enable efficient, secure, and accurate delivery of early warning information based on hierarchical cascaded blockchain, significantly improving the timeliness and coverage accuracy of disaster early warnings.

[0161] In one optional embodiment, a two-layer spatial index tree is constructed. Based on multi-dimensional matching indicators, the comprehensive matching degree between the warning range in the warning information package and the warning terminal is calculated. Target warning terminals are then selected based on the comprehensive matching degree, including:

[0162] In the two-layer spatial index tree, the first layer constructs regional index nodes based on administrative regions, and the second layer constructs location index nodes based on geographic coordinates;

[0163] The region is divided into grid cells according to the warning range. Candidate warning terminals are retrieved within the grid cells using a two-layer spatial index tree. A boundary polygon of the warning range is constructed, and the spatial distance matching degree is obtained by calculating the distance from the candidate warning terminals to the boundary polygon.

[0164] The overlapping area is obtained by spatial calculation of the warning range and the administrative region boundary, and the region matching degree is calculated based on the overlapping area; based on the historical warning data of the candidate warning terminals, the warning information reception delay and data integrity rate are extracted to obtain the terminal reliability index.

[0165] Construct a traffic network topology, calculate the shortest path from the candidate early warning terminal to the early warning area to obtain the path connectivity; obtain population distribution data within the early warning area, and calculate the coverage index of the number of people covered by the candidate early warning terminal.

[0166] The spatial distance matching degree, the regional matching degree, the terminal reliability index, the path connectivity degree, and the coverage index are normalized to determine a comprehensive score, and the early warning terminal with the highest comprehensive score is selected as the target early warning terminal.

[0167] In one specific implementation, a two-layer spatial index tree is constructed. The first layer constructs regional index nodes based on the administrative divisions within a town, organizing the town hierarchically according to basic administrative units such as communities and village committees. Each regional index node stores the boundary polygon coordinates, regional code, regional name, and a list of subordinate administrative units for its administrative region. The second layer constructs location index nodes based on geographic coordinates, organizing the early warning terminals within each basic administrative unit according to geographic coordinates, using a quadtree structure to subdivide the space. The location index nodes store basic information such as the terminal's unique identifier, geographic coordinates, administrative unit code, and device type. For example, a town can be divided into multiple community or village committee-level regional index nodes, each containing multiple grid unit-level regional index nodes, each grid unit-level regional index node containing multiple location index nodes, and each location index node corresponding to one or more early warning terminals.

[0168] The early warning area grid is divided using an adaptive grid method, dynamically adjusting the grid size based on the size and shape of the warning area within the town / district. For warning areas less than 10 square kilometers, a uniform grid of 100 meters × 100 meters is used; for areas between 10 and 50 square kilometers, a uniform grid of 200 meters × 200 meters is used; and for areas greater than 50 square kilometers, a uniform grid of 500 meters × 500 meters is used. After grid division, candidate early warning terminals are quickly retrieved within each grid cell using a two-layer spatial index tree. Specifically, the area index node first locates the basic administrative unit involved in the early warning area, and then the location index node locates the early warning terminal within the grid cell.

[0169] To calculate the spatial distance matching degree, a boundary polygon of the warning range within the town is constructed, and the shortest distance from the candidate warning terminal to the boundary polygon is calculated. When the warning terminal is located inside the warning range, the distance is 0; when the warning terminal is located outside the warning range, the shortest Euclidean distance from the warning terminal to the boundary polygon is calculated. The spatial distance matching degree is inversely proportional to the distance; the shorter the distance, the higher the matching degree. In practical applications, for terminals located within the warning range of the town, the spatial distance matching degree is assigned a value of 1; for terminals located outside the warning range, a distance decay function is used to calculate the matching degree. For example, if a warning terminal is 500 meters away from the boundary of the warning range, its spatial distance matching degree can be calculated as 0.85.

[0170] To calculate the regional matching degree, the overlapping area is obtained by spatially calculating the warning range within the town and the boundaries of the basic-level administrative units. Then, the regional matching degree is calculated based on this overlapping area. The regional matching degree is equal to the ratio of the area of ​​the overlapping area to the area of ​​the basic-level administrative unit. For example, if the warning range covers 75% of a community's area, then the regional matching degree for that community is 0.75. For basic-level administrative units completely within the warning range, the regional matching degree is 1; for basic-level administrative units that do not intersect with the warning range, the regional matching degree is 0.

[0171] Terminal reliability metrics are calculated based on historical early warning data, primarily examining two aspects: early warning information reception latency and data integrity rate. Early warning information reception latency refers to the time difference between the issuance of a meteorological disaster early warning information in the town / district and the completion of reception by the terminal; a shorter latency is better. Data integrity rate refers to the completeness of the early warning information successfully received by the terminal; a higher integrity rate is better. The calculation method for terminal reliability metrics is as follows: collect historical data on early warning information reception by the terminal within the town / district over the past three months, and calculate the average latency and average integrity rate. For example, if an early warning terminal has an average reception latency of 0.8 seconds and an average integrity rate of 0.96 over the past three months, its terminal reliability metric can be calculated as 0.92.

[0172] Path connectivity is calculated to construct the town's road network topology, including road nodes and road edges. Road nodes represent road intersections or endpoints, and road edges represent connections between two road nodes. Each road edge has attributes such as road class, length, and traffic capacity. Dijkstra's algorithm is used to calculate the shortest path from candidate warning terminals to the warning area. Path connectivity is inversely proportional to the shortest path length; the shorter the path, the higher the connectivity. For example, if the shortest path length from a certain warning terminal to the warning area within the town is 2 kilometers, the standardized path connectivity can be calculated as 0.9.

[0173] The coverage index is calculated by obtaining population distribution data within the town's warning area and determining the number of people covered by each candidate warning terminal. The coverage range of a warning terminal within a town is related to its device type; for example, a village-level broadcast terminal has a coverage radius of 500 meters, a community electronic display screen has a coverage radius of 150 meters, and a mobile app push notification has a coverage radius equal to the individual's location. The coverage index is equal to the ratio of the number of people covered by the terminal to the total population within the warning area. For example, if the total population within the warning area of ​​a town is 5000, and a certain warning terminal can cover 1500 people, then its coverage index is 0.3.

[0174] The comprehensive score is calculated by normalizing spatial distance matching degree, regional matching degree, terminal reliability index, path connectivity, and coverage index, and then weighted and summing them to obtain the comprehensive score. The weights of each indicator can be adjusted according to the actual needs of the town / district. For example, the weights for spatial distance matching degree, regional matching degree, terminal reliability index, path connectivity, and coverage index are 0.2, 0.1, and 0.2 respectively. The terminal with the highest comprehensive score is selected as the target warning terminal. For example, in a rainstorm disaster warning scenario for a certain town / district, if the comprehensive score of terminal A is calculated to be 0.88 and the comprehensive score of terminal B is 0.76, then terminal A is selected as the target warning terminal.

[0175] Existing methods for selecting meteorological disaster early warning terminals in towns are mainly based on simple administrative region matching or fixed radius filtering, resulting in missed and false warnings, and failing to dynamically adjust terminal selection strategies according to actual conditions. For example, traditional methods send information to all terminals in the town when issuing a warning, or simply select terminals within a fixed radius around the warning point, without considering factors such as terrain features and population distribution within the town. The town-scale two-layer spatial index tree structure proposed in this application organically combines basic administrative units and geographic coordinate information, significantly improving spatial retrieval efficiency within the town. Simultaneously, by introducing a multi-dimensional evaluation index system, comprehensively considering factors such as spatial location, basic administrative units, terminal performance, town transportation connectivity, and population coverage, the selection of meteorological disaster early warning terminals in towns becomes more accurate. The starting point of this application is to solve the problems of inaccurate terminal selection, incomplete coverage, and low efficiency in existing town early warning systems. Ultimately, it achieves accurate delivery of meteorological disaster early warning information in towns, reduces waste of early warning resources, improves early warning efficiency, enhances residents' trust and acceptance of early warning information, and provides strong support for meteorological disaster risk management in towns.

[0176] like Figure 3The diagram shows a comparison of the proposed solution with traditional early warning terminal selection methods in terms of algorithm performance and resource consumption. In terms of query time, the proposed solution requires 125 milliseconds, slightly slower than the fixed radius search method's 87 milliseconds, but significantly better than the administrative region matching method's 213 milliseconds and the nearest neighbor search method's 189 milliseconds. Regarding memory consumption, the proposed solution requires 178MB, lower than the administrative region matching method's 245MB, higher than the fixed radius search method's 123MB and the nearest neighbor search method's 156MB, placing it at a moderate level. However, when examining accuracy, the proposed solution significantly outperforms other methods with a high accuracy of 92.5%. The second-highest accuracy is only 77.8% for the nearest neighbor search method, 71.3% for the administrative region matching method, and a mere 58.7% for the fixed radius search method. In terms of CPU utilization, the proposed solution achieves 35.2%, lower than the administrative region matching method's 42.8% and the nearest neighbor search method's 38.7%, but only higher than the fixed radius search method's 21.5%. These data clearly demonstrate that while the fixed-radius search method performs best in terms of query speed and resource consumption, its accuracy is significantly lower. In contrast, our proposed solution maintains good query speed and moderate resource consumption while achieving a much higher accuracy than other methods, thus achieving the optimal balance between performance and precision. This fully illustrates the comprehensive advantages of our solution in practical applications, particularly its value in critical scenarios requiring high-precision early warning terminals.

[0177] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0178] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intelligent assessment and push notification of meteorological disaster risk classification for towns, characterized in that, include: Meteorological monitoring data from multiple monitoring points within the town were acquired, and a meteorological monitoring dataset was generated. Based on the meteorological monitoring dataset, spatiotemporal features are extracted through a spatiotemporal attention mechanism, and the spatiotemporal features are adaptively fused with preset geographic element features based on a scene-aware dynamic weight mechanism to obtain a fused feature vector. Meteorological disaster risk levels are generated by processing the fused feature vectors through a bidirectional long short-term memory network. Obtain the distribution information of each functional area within the town, generate a regional distribution dataset, construct a town vulnerability assessment model, and calculate the regional vulnerability assessment index; A graph neural network and a time-adaptive fusion network are constructed based on multi-level regional vulnerability assessment indicators to dynamically fuse meteorological disaster risk levels with regional vulnerability characteristics, resulting in a comprehensive risk index. Spatial constraint clustering optimization is then performed on towns to generate risk zoning data, including: The current regional vulnerability data is obtained based on multi-level regional vulnerability assessment indicators, which include population vulnerability indicators, economic vulnerability indicators, infrastructure vulnerability indicators, and emergency response capability indicators. A graph neural network model is constructed, and each indicator in the regional vulnerability data is used as a graph node feature. The edge connection relationship of the graph is constructed based on the geographical adjacency relationship. The spatial dependency features between the graph node features are extracted through the graph convolutional layer. The spatial dependency features at different scales are adaptively aggregated using the attention mechanism to obtain the regional vulnerability feature vector. A time-series adaptive fusion network, comprising a gated recurrent unit and a dynamic calibration module, is constructed. The meteorological disaster risk level and the regional vulnerability feature vector are input into the gated recurrent unit to capture the time-series fusion features. The dynamic calibration module performs adaptive calibration and confidence assessment on the time-series fusion features based on historical data distribution to generate a comprehensive risk index with confidence intervals. The comprehensive risk index is input into a spatial constraint objective function for iterative optimization to obtain an initial risk partition. Based on the region adjacency graph, local density values ​​and the nearest high-density distance are calculated to identify density peak points. These density peak points are then used to optimize the initial risk partition boundary, resulting in risk partition data, including: Receive a comprehensive risk index, which includes risk values ​​from multiple risk sample points; The intra-class distance is obtained by calculating the sum of the Euclidean distances from multiple risk sample points to the class center point. Spatial adjacency relationships are established based on the geographical locations of multiple risk sample points, and the class difference values ​​between adjacent points are calculated to obtain spatial adjacency constraints. The intra-class distances and spatial adjacency constraints are combined to construct a spatial constraint objective function. The comprehensive risk index is input into the spatial constraint objective function, and the category centroids are obtained through iterative optimization. Based on the category centroids, spatial constraint clustering calculations are performed on multiple risk sample points to obtain the initial risk partitioning results. Based on the initial risk partitioning results, a regional adjacency graph is constructed. The number of adjacency points of each risk sample point within a preset cutoff distance is calculated to obtain a first local density value set. A target risk sample point is selected from the first local density value set. Adjacent risk sample points that are adjacent to the target risk sample point and have a local density value greater than the target risk sample point are determined. The minimum distance from the target risk sample point to the adjacent risk sample point is calculated to obtain the nearest high-density distance value. Based on the local density value and the nearest high density distance value, the density peak point in the initial risk partitioning result is identified, and the boundary of the initial risk partitioning result is optimized based on the density peak point to obtain risk partitioning data. A distributed storage network is built based on a hierarchical blockchain. Risk partition data is encrypted to generate early warning information packages. The comprehensive matching degree is calculated through a two-layer spatial index to filter target early warning terminals. The early warning information packages are pushed in segments based on priority ranking.

2. The method according to claim 1, characterized in that, Based on the meteorological monitoring dataset, spatiotemporal features are extracted through a spatiotemporal attention mechanism, and the spatiotemporal features are adaptively fused with preset geographic element features based on a scene-aware dynamic weight mechanism to obtain a fused feature vector. The fused feature vector is processed using a bidirectional long short-term memory network to generate meteorological disaster risk levels, including: A spatiotemporal attention module is constructed, which includes a temporal attention submodule and a spatial attention submodule. Features are extracted from the meteorological monitoring dataset to obtain temporal and spatial features. The temporal and spatial features are then weighted and fused to obtain spatiotemporal fused features. The current meteorological scene is encoded and analyzed to generate a scene vector; based on the scene vector, the dynamic fusion weight of the spatiotemporal fusion feature and the preset geographic element feature is calculated; feature adaptive selection is performed according to the dynamic fusion weight to obtain the fusion feature vector; The fused feature vector is input into the forward and reverse units of the bidirectional long short-term memory network to determine the forward and reverse hidden states and concatenate them. The risk level probability distribution is calculated by outputting the weight matrix to generate the meteorological disaster risk level.

3. The method according to claim 2, characterized in that, The current meteorological scene is encoded and analyzed to generate a scene vector; based on the scene vector, the dynamic fusion weights of the spatiotemporal fusion features and preset geographic element features are calculated; feature adaptive selection is performed according to the dynamic fusion weights to obtain a fusion feature vector including: A dynamic scene-aware feature fusion network is constructed, which includes a scene encoder and an adaptive feature selector. The scene encoder uses a multi-head self-attention mechanism to interactively model the multi-dimensional meteorological elements of the current meteorological scene and generate a scene feature matrix. The scene feature matrix is ​​optimized through residual connections and normalization layers to generate a scene vector. Based on the scene vector, a feature importance evaluation module, a feature relevance calculation unit, and a scene adaptability calculation unit are constructed. The feature relevance calculation unit calculates the mutual information score between the spatiotemporal fusion feature and the preset geographic element feature, and the scene adaptability calculation unit calculates the contribution score of each feature to the current scene. The mutual information score and the contribution score are weighted and combined to obtain the feature importance score. The adaptive feature selector constructs a soft gating mechanism based on the feature importance score, maps the feature importance score to the feature selection probability through the sigmoid function, and dynamically adjusts the feature fusion weights according to the feature selection probability to achieve adaptive selection and fusion of features and generate a fused feature vector.

4. The method according to claim 1, characterized in that, A distributed storage network is constructed based on a hierarchical cascaded blockchain. Risk partitioning data is encrypted to generate early warning information packages. Target early warning terminals are screened by calculating a comprehensive matching degree through a two-layer spatial index. The early warning information packages are then pushed in segments based on priority ranking, including: Construct a hierarchical blockchain network, establish a distributed storage structure, and generate early warning information blocks in the distributed storage structure; Early warning information is generated based on risk partitioning data. The early warning information is then asymmetrically encrypted to obtain encrypted early warning information data. This encrypted data is combined with digital signature information to generate an early warning information package, which is then stored in the early warning information block. A two-layer spatial index tree is constructed. Based on multi-dimensional matching indicators, the comprehensive matching degree between the warning range in the warning information package and the warning terminal is calculated. The target warning terminal is then selected based on the comprehensive matching degree. The risk level in the early warning information package is mapped to the push priority to obtain priority data. The target early warning terminals are prioritized based on the priority data to obtain the early warning terminal push sequence. According to the push sequence of the early warning terminal, the early warning information packet is segmented and stored to obtain early warning information segment data, and the early warning information packet is pushed to the target early warning terminal through dynamic load balancing.

5. The method according to claim 4, characterized in that, A two-layer spatial index tree is constructed. Based on multi-dimensional matching indicators, the comprehensive matching degree between the warning range in the warning information package and the warning terminals is calculated. Target warning terminals are then selected based on the comprehensive matching degree, including: In the two-layer spatial index tree, the first layer constructs regional index nodes based on administrative regions, and the second layer constructs location index nodes based on geographic coordinates; The region is divided into grid cells according to the warning range. Candidate warning terminals are retrieved within the grid cells through a two-layer spatial index tree. The boundary polygon of the warning range is constructed, and the spatial distance matching degree is obtained by calculating the distance from the candidate warning terminal to the boundary polygon. The overlapping area is obtained by spatial calculation of the warning range and the administrative region boundary, and the region matching degree is calculated based on the overlapping area; based on the historical warning data of the candidate warning terminals, the warning information reception delay and data integrity rate are extracted to obtain the terminal reliability index. Construct a traffic network topology, calculate the shortest path from the candidate early warning terminal to the early warning range to obtain the path connectivity; obtain population distribution data within the early warning range, and calculate the coverage index of the number of people covered by the candidate early warning terminal; The spatial distance matching degree, the regional matching degree, the terminal reliability index, the path connectivity degree, and the coverage index are normalized to determine a comprehensive score, and the early warning terminal with the highest comprehensive score is selected as the target early warning terminal.

Images