Method for predicting spatial distribution of storm intensity and demarcating falling area based on graph neural network

Abstract

The application belongs to the technical field of weather forecast, and discloses a storm intensity spatial distribution prediction and falling area demarcation method based on a graph neural network, which comprises the following steps: dividing a target area into multiple spatial units, taking the geometric center of each spatial unit as a node, collecting multi-source meteorological data and mapping the multi-source meteorological data to each node to construct a spatial node set; calculating the edge weight value between any two nodes to construct an adaptive adjacency matrix; generating a node feature vector of each node to form a node feature matrix; inputting the adaptive adjacency matrix and the node feature matrix into a pre-constructed graph convolutional neural network model to output a storm intensity prediction matrix and identify storm nodes; performing connected domain analysis on the storm nodes to extract the peripheral boundary contour of each connected domain and generate a storm falling area set; and the application can realize intelligent processing of the whole process from storm prediction to emergency response, improve the accuracy of storm early warning and the fine degree of falling area demarcation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of weather prediction, more particularly, the present application relates to a rainstorm intensity spatial distribution prediction and falling area demarcation method based on a graph neural network. BACKGROUND

[0002] As an extreme weather phenomenon, rainstorm has the characteristics of strong burst, wide influence range and great destructive power, often causing secondary disasters such as urban waterlogging and mountain flood mudslides, and causing serious threats to people's life and property safety and social and economic development. Therefore, accurately predicting the spatial distribution of rainstorm intensity and accurately demarcating its influence area are of great significance for meteorological disaster prevention and mitigation decision-making, emergency resource allocation and public warning. However, traditional rainstorm prediction methods mainly rely on numerical weather prediction models or statistical regression analysis, which have obvious shortcomings in dealing with complex nonlinear atmospheric systems: numerical weather prediction models are sensitive to initial conditions and consume huge computing resources, making it difficult to capture the fine features of local heavy rain; traditional statistical regression methods are difficult to effectively model the complex spatio-temporal correlation between meteorological elements, resulting in limited prediction accuracy, especially in the demarcation of rainstorm area boundaries and the description of intensity distribution.

[0003] In recent years, deep learning technology has been widely applied in the field of meteorological prediction, and methods such as convolutional neural networks and recurrent neural networks have made some progress in precipitation prediction. However, the above methods still have key problems: on the one hand, most methods treat meteorological data as regular grid images, ignoring the non-Euclidean spatial relationship between different geographical locations and the meteorological transmission characteristics, and cannot accurately depict the spatial propagation law of rainstorm systems; on the other hand, existing methods lack effective integration of topographic factors, wind direction transmission and other physical mechanisms, making it difficult to reflect the real atmospheric dynamic process; in addition, in the identification of rainstorm area, simple threshold judgment is often used, lacking fine analysis of the shape characteristics of the falling area and quantitative evaluation of the risk level, and unable to provide effective support for graded disaster prevention response. Therefore, there is an urgent need for a comprehensive method that combines multi-source meteorological data, graph neural network prediction and falling area demarcation to improve the prediction accuracy of rainstorm intensity spatial distribution and the reliability of falling area demarcation.

[0004] In view of the above problems, the present application proposes a rainstorm intensity spatial distribution prediction and falling area demarcation method based on a graph neural network to solve the above problems. SUMMARY

[0005] In order to overcome the above-mentioned defects of the prior art and achieve the above-mentioned purposes, the present application provides the following technical scheme: a rainstorm intensity spatial distribution prediction and falling area demarcation method based on a graph neural network, comprising:

[0006] Step S1: Divide the target area into multiple spatial units evenly, and use the geometric center of each spatial unit as a node. Collect multi-source meteorological data of the target area and map the multi-source meteorological data to each node to construct a set of spatial nodes.

[0007] Step S2: Based on the spatial node set, calculate the geographical distance, terrain similarity and prevailing wind direction transmission probability between any two nodes, and obtain the edge weights through fusion analysis. Construct an adaptive adjacency matrix based on all edge weights.

[0008] Step S3: For each node in the spatial node set, generate a node feature vector based on the multi-source meteorological data mapped to the corresponding node, and summarize them to form a node feature matrix;

[0009] Step S4: Input the adaptive adjacency matrix and node feature matrix into the pre-constructed graph convolutional neural network model. The graph convolutional neural network model performs weighted propagation on the node feature matrix based on the adaptive adjacency matrix, extracts the temporal evolution pattern in the node feature matrix, and outputs the future... The predicted rainfall intensity values ​​for each node at each time point are used to form a rainfall intensity prediction matrix.

[0010] Step S5: Perform spatial interpolation and threshold segmentation on the rainstorm intensity prediction matrix, identify rainstorm nodes based on the threshold segmentation results, perform connected component analysis on all rainstorm nodes, extract the outer boundary contours of each connected component based on the spatial interpolation results, and generate a rainstorm landing area set.

[0011] Step S6: Statistically analyze the location attribute indicators of each rainstorm location in the rainstorm location cluster, calculate the comprehensive risk level of each rainstorm location based on the location attribute indicators, and generate differentiated early warning response plans for rainstorm locations with different comprehensive risk levels.

[0012] Furthermore, multi-source meteorological data includes dynamic observation data and static observation data; dynamic observation data includes... Rainfall, radar echo intensity, and cloud water content at consecutive moments; static observation data includes topographic elevation data.

[0013] Methods for constructing a spatial node set include:

[0014] Raster registration was performed on radar echo intensity and cloud water content in static and dynamic observation data to obtain the terrain elevation data corresponding to each node. The radar echo intensity and cloud water content at consecutive time points; spatial interpolation of rainfall in dynamic observation data to obtain the corresponding data for each node. Rainfall data at consecutive time points; topographic elevation data corresponding to each node and... The rainfall amount, the radar echo intensity and the cloud water content of each continuous time are integrated to obtain node meteorological data of each node.

[0015] Further, the method for calculating the geographical distance, the terrain similarity and the dominant wind direction transmission probability between any two nodes comprises:

[0016] The latitude and longitude coordinates of each node are obtained, and the spherical distance between any two nodes is calculated based on the latitude and longitude coordinates using the formula for calculating the spherical distance, and is taken as the geographical distance.

[0017] According to the terrain elevation data of each node, the cosine similarity between any two nodes is calculated in sequence, and is taken as the terrain similarity.

[0018] The historical wind field data corresponding to each node are obtained, and the historical wind field data includes the wind direction angle and the wind speed at multiple historical times; the value range of the wind direction angle is uniformly divided according to a preset interval width to obtain multiple wind direction intervals; the dominant interval of each node is determined in sequence from the multiple wind direction intervals, and the mean value of the wind direction angle and the wind speed in the dominant interval corresponding to each node is calculated to obtain the dominant wind direction and the average wind speed of each node; the direction angle of the connecting line between any two nodes is calculated according to the latitude and longitude coordinates of each node; and the dominant wind direction transmission probability between any two nodes is calculated according to the geographical distance and the direction angle of the connecting line between any two nodes, and the dominant wind direction and the average wind speed of each node.

[0019] Further, the method for constructing the adaptive adjacency matrix comprises:

[0020] The geographical distance between any two nodes is sequentially standardized to obtain the standard distance between any two nodes; different correlation weights are set for the standard distance, the terrain similarity and the dominant wind direction transmission probability, and the reciprocal of the standard distance, the terrain similarity and the dominant wind direction transmission probability between any two nodes are sequentially weighted and summed based on the correlation weights to obtain the edge weight value between any two nodes.

[0021] An initial adjacency matrix is constructed, and all edge weight values are sequentially added to the initial adjacency matrix; a screening threshold is calculated according to the edge weight values in the initial adjacency matrix; and the edge weight values in the initial adjacency matrix that are less than the screening threshold are all adjusted to 0.

[0022] The initial adjacency matrix after the edge weight value adjustment is taken as the adaptive adjacency matrix.

[0023] Further, the method for summarizing the node feature matrix comprises:

[0024] According to the node meteorological data of each node, the rainfall, radar echo intensity and cloud water content corresponding to the same time in the same node are integrated to obtain the time meteorological data of each node, and the terrain elevation data of each node is sequentially added to the corresponding time meteorological data; each set of time meteorological data is input into the trained feature extraction model to predict the time feature data corresponding to each set of time meteorological data, and the feature extraction model is a deep neural network model;

[0025] The time feature data corresponding to each node is integrated to obtain the node feature vector of each node; the node feature vectors of each node are summarized to form the node feature matrix of the target area.

[0026] Further, the method for forming the rainstorm intensity prediction matrix comprises:

[0027] The adaptive adjacency matrix and the node feature matrix are input into the pre-constructed graph convolutional neural network model, the graph convolutional neural network model performs weighted propagation on the node feature matrix by using the adaptive adjacency matrix through a graph convolutional layer to obtain a weighted feature matrix; wherein the weighted feature matrix includes a weighted feature vector corresponding to each node; the graph convolutional neural network model processes the change of the weighted feature matrix at consecutive time points through a time convolutional layer to extract the time evolution law in the weighted feature matrix;

[0028] The graph convolutional neural network model combines the weighted feature matrix and the time evolution law to generate the rainstorm intensity prediction value of each node at future time points; the rainstorm intensity prediction values of each node at future time points are integrated to form the rainstorm intensity prediction matrix.

[0029] Further, the method for identifying the rainstorm node comprises:

[0030] Each rainstorm intensity prediction value corresponding to each node is compared with a preset rainstorm threshold value, and the future time point at which the rainstorm intensity prediction value is greater than the rainstorm threshold value is marked as a rainstorm time point; the maximum rainstorm duration corresponding to each node is counted according to the rainstorm time point corresponding to each node; the maximum rainstorm duration corresponding to each node is compared with a preset duration threshold value, and the node with a maximum rainstorm duration greater than the duration threshold value is marked as a rainstorm node.

[0031] Further, the method for generating the rainstorm falling area set comprises:

[0032] The rainstorm nodes corresponding to each future time point in future time points are integrated to obtain A set of rainstorm nodes is identified; connected component analysis is performed on each set of rainstorm nodes to obtain the connected component at each future time; the rainstorm intensity prediction values ​​with the same future time in the rainstorm intensity prediction matrix are integrated to obtain... A set of rainfall intensity ensembles is generated; spatial interpolation is then performed on the predicted rainfall intensity values ​​within each set of rainfall intensity ensembles to obtain the predicted rainfall intensity for each spatial location within the target area in the future. The predicted intensity of the rainstorm at a given moment;

[0033] Each predicted rainfall intensity value corresponding to each spatial location is compared with a rainfall threshold. Spatial locations where the predicted rainfall intensity value is greater than the rainfall threshold are taken as the rainfall locations within the corresponding connected components. Based on the rainfall locations within each connected component, the boundary contours of each connected component are extracted to obtain the outer boundary contours of each connected component, which are then used as the time-domain areas. The time-domain areas of each future time are merged to obtain the rainfall-domain areas. All rainfall-domain areas are then integrated to generate a rainfall-domain area set.

[0034] Furthermore, the methods for fusing the time regions of each future moment include:

[0035] For the future Each future moment in the time frame is assigned a corresponding time weight, and the predicted rainfall intensity values ​​for different future moments at the same spatial location are weighted and fused based on the time weight to obtain the comprehensive rainfall intensity for each spatial location.

[0036] Each spatial location within the rainfall area at each time point is marked as a candidate location. Based on the predicted rainfall intensity for each candidate location at different future times, the duration of the rainfall at each candidate location is calculated. For each candidate location, the corresponding rainfall duration is compared with a duration threshold, and the corresponding comprehensive rainfall intensity is compared with a rainfall threshold. Candidate locations with a rainfall duration greater than the duration threshold and a comprehensive rainfall intensity greater than the rainfall threshold are marked as selected locations. Connectivity analysis is performed on all selected locations to obtain multiple candidate rainfall areas.

[0037] The number of selected locations within each candidate landing area is counted and marked as the landing area size. The landing area size of each candidate landing area is compared with a preset size threshold, and the candidate landing areas with landing area sizes greater than the size threshold are designated as rainstorm landing areas.

[0038] Furthermore, the method for calculating the comprehensive risk level of each rainstorm area includes:

[0039] Each indicator in the area attribute index corresponding to each rainstorm area is standardized to obtain the area standard index corresponding to each rainstorm area; the entropy weight method is used to calculate the index weight corresponding to each standard index in the area standard index corresponding to each rainstorm area.

[0040] Setting positive ideal solution and negative ideal solution; calculating the positive score between each rainstorm falling area corresponding falling area standard index and the positive ideal solution and the negative score between each rainstorm falling area corresponding falling area standard index and the negative ideal solution in turn, and calculating the comprehensive score of each rainstorm falling area according to the positive score and the negative score in turn; calculating the ratio between the negative score and the comprehensive score corresponding to each rainstorm falling area, and obtaining the risk score of each rainstorm falling area;

[0041] Setting different risk intervals for different comprehensive risk levels, and comparing the risk score of each rainstorm falling area with each risk interval in turn; taking the comprehensive risk level corresponding to the risk interval where the risk score falls as the comprehensive risk level of the corresponding rainstorm falling area;

[0042] Among them, the falling area attribute index includes falling area size, intensity gradient entropy and fluctuation intensity index.

[0043] The technical effects and advantages of the rainstorm intensity spatial distribution prediction and falling area delimitation method based on the graph neural network of the present application are as follows:

[0044] By introducing the graph neural network model, combining multi-source meteorological data and wind field characteristics, high-precision prediction of the spatial distribution of rainstorm intensity and automatic delimitation of rainstorm falling area are realized; by constructing an adaptive adjacency matrix that integrates geographical distance, terrain similarity and dominant wind direction transmission probability, the limitations of traditional methods that regard meteorological data as regular grids are broken through, the non-Euclidean spatial relationship of the rainstorm system and the meteorological transmission physical mechanism are effectively described, and the expression ability of the spatial propagation characteristics of the rainstorm system is improved; the graph convolutional neural network model is used to simultaneously capture the spatio-temporal dependence relationship of the rainstorm evolution, significantly improving the prediction accuracy of the spatial distribution of rainstorm intensity and the capture ability of local heavy precipitation characteristics; a multiple screening mechanism (including intensity threshold, double determination of duration, time weighted fusion and size constraint) is used to realize accurate delimitation of rainstorm falling area, effectively avoiding misjudgment and omission caused by simple threshold determination; based on multi-dimensional attribute indexes such as falling area size, intensity gradient entropy and fluctuation intensity index, the comprehensive risk level is quantified scientifically, and a differentiated early warning response scheme is generated; the whole process of intelligent processing from rainstorm prediction to risk assessment and emergency response is realized, effectively overcoming the shortcomings of existing technologies in complex nonlinear atmospheric system modeling, spatio-temporal correlation expression, falling area fine description and risk quantitative assessment, thereby improving the accuracy of rainstorm warning, the fine degree of falling area delimitation and the scientificity of disaster prevention and mitigation decision-making. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 The flow chart of the rainstorm intensity spatial distribution prediction and falling area delimitation method based on the graph neural network of the present application is shown in Figure 1. DETAILED DESCRIPTION

[0046] With reference to the accompanying drawings, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all the other embodiments obtained by those ordinarily skilled in the art without creative effort belong to the scope of the present application.

[0047] Embodiment 1

[0048] Please refer to Figure 1 As shown in the figure, the method for predicting the spatial distribution of storm intensity and delimiting the storm area based on the graph neural network comprises the following steps.

[0049] Step S1, uniformly divide the target area into a plurality of spatial units, and take the geometric center of each spatial unit as a node, collect multi-source meteorological data of the target area, and map the multi-source meteorological data to each node to construct a spatial node set.

[0050] In this embodiment, the target area is uniformly divided based on the unit area to obtain a plurality of spatial units with the same area by a preset unit area. Each spatial unit is a rectangle, and the unit area is set in advance by those skilled in the art according to the scale of the target area, the prediction accuracy requirement and other factors. The target area refers to the geographical area range for storm prediction. The geometric center refers to the centroid inside the spatial unit.

[0051] In this embodiment, the multi-source meteorological data includes dynamic observation data and static observation data.

[0052] Specifically, the dynamic observation data includes rainfall, radar echo intensity and cloud water content at N continuous time points, N is an integer greater than 1; N continuous time points with the current time as the end point, that is, The last time point of N continuous time points is the current time, and the current time is the specific time for predicting the spatial distribution of storm intensity and delimiting the storm area. The static observation data is terrain elevation data.

[0053] Rainfall refers to the actual precipitation measured per unit time by rain gauges at various meteorological stations within the target area, reflecting the actual intensity of precipitation on the ground. Radar echo intensity refers to the signal strength value returned by the electromagnetic waves emitted by weather radar after encountering precipitation particles (raindrops) within the target area, characterizing the concentration and size of airborne precipitation particles. Cloud water content refers to the mass of liquid water contained in a unit volume of air within the target area, calculated by inversion from cloud echo signals detected by cloud radar, characterizing the concentration of liquid water droplets in clouds and potential precipitation capacity. Topographic elevation data refers to the height values ​​of various locations on the land surface within the target area relative to a reference surface (such as sea level), obtained through remote sensing or ground mapping, characterizing topographic relief features and terrain distribution.

[0054] In this embodiment, the method for constructing a spatial node set includes:

[0055] Raster registration was performed on radar echo intensity and cloud water content in static and dynamic observation data to obtain the terrain elevation data corresponding to each node. The radar echo intensity and cloud water content at consecutive time points; spatial interpolation of rainfall in dynamic observation data to obtain the corresponding data for each node. Rainfall at consecutive time points; among which, grid registration methods include affine transformation, bilinear interpolation resampling, etc., and spatial interpolation methods include inverse distance weighting, Kriging interpolation, etc.

[0056] The terrain elevation data corresponding to each node and The rainfall, radar echo intensity, and cloud water content at consecutive moments are integrated to obtain the node meteorological data for each node; based on all nodes and their corresponding node meteorological data, a spatial node set is constructed.

[0057] It should be noted that the terrain elevation data for each node is derived from the local area where the corresponding node is located. It consists of several height values, used to characterize the terrain feature distribution of the local area where the corresponding node is located; where the local area corresponding to the node is centered on the node. It is an integer greater than 1.

[0058] Step S2: Based on the set of spatial nodes, calculate the geographical distance, terrain similarity and prevailing wind direction transmission probability between any two nodes, and obtain the edge weights through fusion analysis. Construct an adaptive adjacency matrix based on all edge weights.

[0059] In this embodiment, the method for calculating the geographical distance between any two nodes includes:

[0060] The latitude and longitude coordinates of each node are obtained, and the spherical distance between any two nodes is calculated based on the latitude and longitude coordinates using the Haversine formula, and is taken as the geographic distance.

[0061] The latitude and longitude coordinates of each node are calculated by a person skilled in the art according to the latitude and longitude range of the target area and the unit area; the Haversine formula is prior art, and the specific calculation process is not described in detail here.

[0062] In this embodiment, the method for calculating the terrain similarity between any two nodes includes:

[0063] According to the terrain elevation data of each node, the cosine similarity between any two nodes is calculated in turn, and is taken as the terrain similarity.

[0064] In this embodiment, the method for calculating the dominant wind direction transmission probability between any two nodes includes:

[0065] The historical wind field data corresponding to each node is obtained, and the historical wind field data includes the wind direction angle and the wind speed at multiple historical time points. The wind direction angle is obtained by a wind direction sensor arranged in each space unit, and the wind speed is obtained by a wind speed sensor arranged in each space unit.

[0066] The value range of the wind direction angle is uniformly divided according to a preset interval width, to obtain multiple wind direction intervals. The interval width is preset by a person skilled in the art according to the actual situation, and the value range of the wind direction angle is The number of wind directions corresponding to each node in each wind direction interval is counted according to the wind direction angle corresponding to each node, and the number of wind directions is the number of wind direction angles; the number of wind directions of the same node is compared, and the wind direction interval with the largest number of wind directions is taken as the dominant interval of the corresponding node; the mean value of the wind direction angle and the wind speed in the dominant interval corresponding to each node is calculated, to obtain the dominant wind direction and the average wind speed corresponding to each node.

[0067] The direction angle of the connecting line between any two nodes is calculated according to the latitude and longitude coordinates of each node. The expression of the direction angle of the connecting line is:

[0068] ;

[0069] In the formula, is the latitude coordinate of node , is the longitude coordinate of node , is the direction angle of the connecting line pointing from node to node , is the latitude coordinate of node , is the longitude coordinate of node , is the latitude coordinate of node , is the latitude coordinate of node , For nodes latitude coordinates It is a bivariate arctangent function.

[0070] Based on the geographical distance and the directional angle of the line connecting any two nodes, as well as the prevailing wind direction and average wind speed at each node, calculate the probability of prevailing wind direction transmission between any two nodes. The expression for the probability of prevailing wind direction transmission is:

[0071] ;

[0072] In the formula, For nodes For nodes The probability of prevailing wind direction transmission. These are normalization coefficients used to ensure that the probability of prevailing wind transmission is within a certain range. between, It is an exponential function. For nodes The prevailing wind direction, This is a wind direction dispersion parameter used to reflect wind direction stability. For nodes With nodes Geographical distance between them For nodes average wind speed, For transmission time scale, The wind speed is the highest value among the historical wind field data corresponding to all nodes. The normalization coefficient, wind direction dispersion parameter, and transmission time scale are all set by those skilled in the art based on the historical wind field data and practical experience for each node. In this embodiment, the preferred value range for the wind direction dispersion parameter is [value range missing]. .

[0073] It should be noted that when , indicating nodes Located at node The direction of the headwind, therefore the node For nodes The prevailing wind direction transmission is almost non-existent, and the corresponding prevailing wind direction transmission probability is set to 0; at the same time, due to the existence of angle cyclicity issues, therefore when It also needs to be calculated and will and Compare; Updated to and The minimum value in, and with By comparing the results, it is possible to accurately determine whether the probability of the prevailing wind direction should be set to 0.

[0074] In the embodiment, the method for constructing the adaptive adjacency matrix comprises:

[0075] The geographical distance between any two nodes is sequentially standardized to obtain a standard distance between any two nodes; different correlation weights are set for the standard distance, the terrain similarity and the dominant wind direction transmission probability, and the inverse of the standard distance, the terrain similarity and the dominant wind direction transmission probability between any two nodes are sequentially weighted and summed based on the correlation weights to obtain an edge weight value between any two nodes. The correlation weights are set in advance by the person skilled in the art according to the importance of the standard distance, the terrain similarity and the dominant wind direction transmission probability in the correlation between nodes in the rainstorm scenario;

[0076] An initial adjacency matrix is constructed, and all edge weight values are sequentially added to the initial adjacency matrix; wherein the size of the initial adjacency matrix is , , where n is the number of nodes; the mean and the standard deviation of all edge weight values are calculated to obtain an edge mean and an edge standard deviation; the edge standard deviation is multiplied by a preset adjustment coefficient, and then the edge mean is added to obtain a screening threshold; the adjustment coefficient is set in advance by the person skilled in the art according to the actual situation; the edge weight values in the initial adjacency matrix are compared with the screening threshold, and the edge weight values less than the screening threshold are adjusted to 0, and the edge weight values greater than or equal to the screening threshold are not adjusted;

[0077] The initial adjacency matrix with the adjusted edge weight values is used as the adaptive adjacency matrix.

[0078] It should be understood that the geographical distance, the terrain similarity and the dominant wind direction transmission probability between two nodes are calculated to depict the correlation between different nodes; since the occurrence and evolution of rainstorm has significant spatial correlation and propagation, the meteorological data of a single node cannot reflect the overall dynamics of the meteorological system; therefore, by comprehensively considering the spatial proximity, the terrain feature consistency and the wind direction transmission characteristics, an edge weight value more consistent with the meteorological physical law can be constructed to form the adaptive adjacency matrix, thereby providing a reliable spatial relationship constraint for the subsequent graph convolutional neural network model for predicting the rainstorm intensity.

[0079] Step S3, for each node in the spatial node set, a node feature vector is generated according to the multi-source meteorological data mapped to the corresponding node, and a node feature matrix is formed by summarizing.

[0080] In the embodiment, the method for forming the node feature matrix comprises:

[0081] According to the node meteorological data of each node, the rainfall, radar echo intensity and cloud water content corresponding to the same time in the same node are integrated to obtain the time meteorological data of each node, and the terrain elevation data of each node is sequentially added to the corresponding time meteorological data; each group of time meteorological data is input into the trained feature extraction model to predict the time feature data corresponding to each group of time meteorological data. The time feature data includes but is not limited to cloud-rain conversion efficiency index, terrain forced convergence index, vertical development comprehensive index, and water vapor convergence intensity.

[0082] Specifically, the cloud-rain conversion efficiency index refers to the efficiency of the conversion of cloud water content in the atmosphere into surface precipitation, which is obtained by fusing rainfall, radar echo intensity and cloud water content, and is used to reflect whether the precipitation potential of the rainstorm system is enhanced or attenuated; the terrain forced convergence index refers to the quantitative value of the blocking, lifting and convergence of airflow by terrain, which is obtained by fusing rainfall, radar echo intensity and terrain elevation data, and is used to reflect the enhancement effect of terrain on rainfall, so as to determine whether the rainstorm is likely to be rapidly enhanced due to the effect of terrain; the vertical development comprehensive index refers to the development potential of the convective system in the vertical direction, which is obtained by fusing rainfall, radar echo intensity, cloud water content and terrain elevation data, and is used to reflect the vertical development of the convective cloud system, so as to determine whether the rainstorm system is in the development period, mature period or decay period; the water vapor convergence intensity refers to the degree of local water vapor aggregation under the action of terrain and meteorological system, which is obtained by fusing rainfall, radar echo intensity, cloud water content and terrain elevation data, and is used to reflect the water vapor supply condition of the rainstorm system, so as to determine whether the rainfall system has sufficient water vapor to maintain or enhance the precipitation intensity.

[0083] It should be noted that the feature extraction model is a deep neural network model, which includes an input layer, a hidden layer and an output layer; each hidden layer includes a plurality of neurons, and each neuron is connected to the next layer of neurons, and the connection includes a weight that determines the importance and influence of data transmission in the neural network; an activation function is applied to each neuron between the hidden layer and the output layer, which introduces nonlinearity and allows the network to learn more complex patterns and features; the training process of the feature extraction model includes:

[0084] Pre-collecting Group different time meteorological data, and sequentially set corresponding time feature data for each group of time meteorological data, is an integer greater than 1; each group of time meteorological data and the corresponding time feature data are respectively converted into a group of feature vectors; the time feature data corresponding to the time meteorological data is collected by those skilled in the art in the process of historically predicting the spatial distribution of rainstorm intensity. By analyzing different sets of meteorological data at different times, and based on practical experience, each set of meteorological data is analyzed sequentially to obtain the corresponding time-specific characteristics of each set of meteorological data, thereby completing the process. Setting the time-specific characteristic data corresponding to the group of time-based meteorological data;

[0085] Each set of feature vectors is used as input to the feature extraction model, which outputs a set of predicted time-time feature data corresponding to each set of meteorological data at each time moment. The actual time-time feature data corresponding to each set of meteorological data at each time moment is used as the prediction target. The actual time-time feature data is the pre-set time-time feature data corresponding to the meteorological data at each time moment. The training objective is to minimize the sum of prediction errors for all meteorological data at all times. The formula for calculating the prediction error is as follows: ,in For prediction error, This represents the group number of the feature vector corresponding to the meteorological data at that time. For the first The characteristic data of the predicted time corresponding to the meteorological data at each time point. For the first The actual time-specific feature data corresponding to the meteorological data at each time point are used to train the feature extraction model until the sum of the prediction errors converges.

[0086] The time-specific feature data corresponding to each node are integrated to obtain the node feature vector of each node; the node feature vectors of each node are summarized to form the node feature matrix of the target region.

[0087] Step S4: Input the adaptive adjacency matrix and node feature matrix into the pre-constructed graph convolutional neural network model. The graph convolutional neural network model performs weighted propagation on the node feature matrix based on the adaptive adjacency matrix, extracts the temporal evolution pattern in the node feature matrix, and outputs the future... The predicted rainfall intensity values ​​for each node at each time point are used to form a rainfall intensity prediction matrix.

[0088] In this embodiment, the method for forming a rainstorm intensity prediction matrix includes:

[0089] The adaptive adjacency matrix and the node feature matrix are input into a pre-constructed graph convolutional neural network model. The graph convolutional neural network model uses the graph convolutional layer to perform weighted propagation on the node feature matrix using the adaptive adjacency matrix to obtain a weighted feature matrix, thereby realizing the spatial information fusion of node feature vectors between different nodes. The weighted feature matrix includes a weighted feature vector corresponding to each node. Each weighted feature vector is the result of weighted fusion of the node feature vector of the corresponding node and the node feature vectors of all its corresponding neighboring nodes (i.e., nodes with non-zero edge weights) according to the corresponding edge weights in the adaptive adjacency matrix.

[0090] The expression for the weighted characteristic matrix is: In the formula, This represents the weighted characteristic matrix, i.e., This represents the activation function. Represents an adaptive adjacency matrix. Indicates the first The node feature matrix of the layer Indicates the first The weight matrix of the layer, , This represents the number of graph convolutional layers; the activation function and the number of graph convolutional layers are both preset by those skilled in the art based on the task characteristics and practical experience of the graph convolutional neural network model; the weight matrix is ​​initially set by those skilled in the art according to the actual situation, and is continuously optimized through the backpropagation algorithm during the training process of the graph convolutional neural network model;

[0091] The graph convolutional neural network model applies weighted feature matrices through temporal convolutional layers. The changes at consecutive moments are processed to extract the temporal evolution pattern in the weighted feature matrix. The temporal evolution pattern refers to the dynamic trend and mutual coupling relationship of each data in the moment feature data with time, which is used to reflect the development and evolution process of the rainstorm system in the time dimension.

[0092] Graph convolutional neural network models combine weighted feature matrices and temporal evolution patterns to generate future... The predicted rainfall intensity for each node at each time point. It is an integer greater than 1; it will be the future. The predicted rainfall intensity values ​​for each node at each time point are integrated to form a rainfall intensity prediction matrix.

[0093] It should be noted that the graph convolutional neural network model includes an input layer, a graph convolutional layer, a temporal convolutional layer, activation functions, pooling / normalization layers, and an output layer. The input layer receives the node feature matrix and the adaptive adjacency matrix. The graph convolutional layer performs weighted propagation and spatial fusion of the node feature vectors based on the adaptive adjacency matrix to capture the spatial correlation characteristics between nodes. The temporal convolutional layer processes the changes in the node feature vectors over consecutive time steps to extract the temporal evolution pattern of the rainstorm system. Activation functions are typically applied between each convolutional layer to introduce nonlinearity, and normalization or pooling layers can be used to stabilize training and enhance feature representation capabilities. The output layer generates the predicted rainstorm intensity value for each node at multiple future times based on the spatial propagation and temporal evolution results, forming the final rainstorm intensity prediction matrix. The graph convolutional neural network model is an existing technology, and its specific construction process will not be elaborated upon here.

[0094] Step S5, spatial interpolation and threshold segmentation processing are performed on the rainstorm intensity prediction matrix, rainstorm nodes are identified based on the threshold segmentation processing result, connected component analysis is performed on all rainstorm nodes, the peripheral boundary contour of each connected component is extracted based on the spatial interpolation result, and a rainstorm falling area set is generated.

[0095] In this embodiment, the method for identifying rainstorm nodes includes:

[0096] Each rainstorm intensity prediction value corresponding to each node is compared with a preset rainstorm threshold respectively, a future time at which the rainstorm intensity prediction value is greater than the rainstorm threshold is marked as a rainstorm time, and a future time at which the rainstorm intensity prediction value is less than or equal to the rainstorm threshold is not marked; the maximum rainstorm duration corresponding to each node is counted according to the rainstorm time corresponding to each node; the maximum rainstorm duration corresponding to each node is compared with a preset duration threshold respectively; a node at which the maximum rainstorm duration is greater than the duration threshold is marked as a rainstorm node, and a node at which the maximum rainstorm duration is less than or equal to the duration threshold is not marked.

[0097] The rainstorm threshold and the duration threshold are both preset by a person skilled in the art according to actual conditions; the maximum rainstorm duration is the maximum continuous number of rainstorm times; for example, a node corresponds to eight times, among which time 1, time 2, time 4, time 5 and time 6 are rainstorm times, the continuous number of time 1 and time 2 is 2, and the continuous number of time 4, time 5 and time 6 is 3, so the maximum continuous number of rainstorm times is 3, that is, the maximum rainstorm duration is 3.

[0098] In this embodiment, the method for generating the rainstorm falling area set includes:

[0099] Each rainstorm node corresponding to each future time in the future times is integrated to obtain a set of rainstorm node sets; each set of rainstorm node sets is sequentially subjected to connected component analysis to obtain a connected component of each future time; rainstorm intensity prediction values corresponding to the same future time in the rainstorm intensity prediction matrix are integrated to obtain a set of rainstorm intensity sets; each rainstorm intensity prediction value in each set of rainstorm intensity sets is sequentially subjected to spatial interpolation to obtain a rainstorm intensity prediction value of each spatial position in the target area at the future times; it should be noted that the connected component analysis is a prior art, and the specific process will not be described in detail here;

[0100] Each rainstorm intensity prediction value corresponding to each spatial position is compared with the rainstorm threshold respectively; a spatial position at which the rainstorm intensity prediction value is greater than the rainstorm threshold is taken as a rainstorm position in the corresponding connected component; and a spatial position at which the rainstorm intensity prediction value is less than or equal to the rainstorm threshold is not operated;

[0101] Based on the location of the rainstorm within each connected component, the boundary contours of each connected component are extracted using algorithms such as Moore neighborhood tracing and Suzuki-Abe, and the outer boundary contours of each connected component are obtained and used as the time-based landing area. The time-based landing areas of each future time are merged to obtain the rainstorm landing area. All rainstorm landing areas are integrated to generate a rainstorm landing area set.

[0102] In this embodiment, the method for fusing the time regions of each future time step includes:

[0103] For the future Each future moment in the time frame is assigned a corresponding time weight, and the predicted rainfall intensity values ​​for different future moments at the same spatial location are weighted and fused based on the time weights to obtain the comprehensive rainfall intensity for each spatial location. The time weights are preset by those skilled in the art according to the actual situation. The closer the future moment is to the current moment, the greater the time weight, and the farther the future moment is from the current moment, the smaller the time weight, and the sum of all time weights is 1.

[0104] Each spatial location within the rainfall area at each time point is marked as a candidate location. Based on the predicted rainfall intensity for each candidate location at different future times, the duration of rainfall at each candidate location is calculated. For each candidate location, the corresponding rainfall duration is compared with a duration threshold, and the corresponding comprehensive rainfall intensity is compared with a rainfall threshold. Candidate locations with a rainfall duration greater than the duration threshold and a comprehensive rainfall intensity greater than the rainfall threshold are marked as selected locations. Candidate locations with a rainfall duration less than or equal to the duration threshold, or a comprehensive rainfall intensity less than or equal to the rainfall threshold, are not marked. Connectivity analysis is performed on all selected locations to obtain multiple candidate rainfall areas. The expression for calculating the rainfall duration at each candidate location is:

[0105] ;

[0106] In the formula, For the duration of the rainstorm, For the first A future moment, , For the first The predicted intensity of the rainstorm at a future time. Let this be an indicator function for the occurrence of heavy rain. The threshold for heavy rainfall; where, when hour, ,when hour, ;

[0107] Count the number of selected positions in each candidate drop zone and mark it as the drop zone size; compare the drop zone size of each candidate drop zone with the preset size threshold respectively, and the size threshold is preset by the person skilled in the art according to the actual situation; the candidate drop zone with the drop zone size greater than the size threshold is taken as the heavy rain drop zone, and the candidate drop zone with the drop zone size less than or equal to the size threshold is not taken as the heavy rain drop zone.

[0108] Step S6, count the drop zone attribute index of each heavy rain drop zone in the heavy rain drop zone cluster, and calculate the comprehensive risk level of each heavy rain drop zone based on the drop zone attribute index, and generate a differentiated early warning response scheme for heavy rain drop zones with different comprehensive risk levels.

[0109] In the embodiment, the method for calculating the comprehensive risk level of each heavy rain drop zone comprises:

[0110] Each index in the drop zone attribute index corresponding to each heavy rain drop zone is respectively standardized to obtain the drop zone standard index corresponding to each heavy rain drop zone; the index weight corresponding to each standard index in the drop zone standard index corresponding to each heavy rain drop zone is calculated in turn by using the entropy weight method; wherein the entropy weight method is a prior art, and the specific process will not be described in detail here.

[0111] At this time, the positive ideal solution and the negative ideal solution are set, the positive ideal solution is , and the negative ideal solution is ; the weighted Euclidean distance between the drop zone standard index corresponding to each heavy rain drop zone and the positive ideal solution is calculated in turn to obtain the positive score; the weighted Euclidean distance between the drop zone standard index corresponding to each heavy rain drop zone and the negative ideal solution is calculated in turn to obtain the negative score; the sum of the positive score and the negative score corresponding to the same heavy rain drop zone is calculated to obtain the comprehensive score of each heavy rain drop zone; the ratio between the negative score and the comprehensive score corresponding to each heavy rain drop zone is calculated to obtain the risk score of each heavy rain drop zone;

[0112] Wherein, the calculation method of the weighted Euclidean distance is: for each standard index in the drop zone standard index, first calculate the index difference between the corresponding value in the positive ideal solution or the negative ideal solution, then add each index difference square multiplied by the corresponding index weight in turn, and finally take the square root to obtain the weighted Euclidean distance;

[0113] Different risk intervals are set for different comprehensive risk levels, and the risk score of each heavy rain drop zone is compared with each risk interval in turn; the comprehensive risk level corresponding to the risk interval into which the risk score falls is taken as the comprehensive risk level of the corresponding heavy rain drop zone; the risk interval of different comprehensive risk levels is set by the person skilled in the art according to the risk score calculated in the historical drop zone delimitation process.

[0114] Among them, the landing area attribute indicators include landing area size, intensity gradient entropy, and pulsation intensity index; landing area size is used to describe the spatial size of the rainstorm landing area; intensity gradient entropy is used to measure the unevenness or complexity of the distribution of rainstorm intensity prediction values ​​within the rainstorm landing area; pulsation intensity index is used to measure the degree of fluctuation of rainstorm intensity prediction values ​​within the rainstorm landing area over time, reflecting the instability of the rainstorm.

[0115] Among them, the intensity gradient entropy is specifically the information entropy. The calculation method for information entropy is existing technology, and the specific calculation process will not be elaborated here. The expression for the pulsation intensity index is:

[0116] ;

[0117] In the formula, The pulsation intensity index, This represents the standard deviation of the predicted rainfall intensity for all selected locations within the rainfall area. This represents the average of the predicted rainfall intensity values ​​for all selected locations within the rainfall area. This represents the average duration of the heavy rainfall at all selected locations within the heavy rainfall area. This is the average of the peak counts for all selected locations within the rainstorm area. The peak count refers to the number of times a local maximum value is reached in the rainstorm intensity time series corresponding to the selected location (i.e., the rainstorm intensity prediction value at a certain future time is higher than the rainstorm intensity prediction values ​​at the previous and next times). It is used to reflect the fluctuation frequency of the rainstorm intensity prediction value over time.

[0118] In this embodiment, the method for generating differentiated early warning response schemes for rainstorm areas with different comprehensive risk levels includes:

[0119] A pre-set early warning response set includes early warning response plans corresponding to different comprehensive risk levels. Based on the comprehensive risk level corresponding to different rainstorm areas, the corresponding early warning response plan is obtained from the early warning response set. The comprehensive risk level includes, but is not limited to, low-risk, medium-risk, and high-risk levels. When the comprehensive risk level is low, the corresponding early warning response plan includes, for example, strengthening meteorological monitoring and sending reminder text messages to the public. When the comprehensive risk level is medium, the corresponding early warning response plan includes, for example, activating the flood control duty system and dredging the drainage system in advance. When the comprehensive risk level is high, the corresponding early warning response plan includes, for example, activating the emergency response mechanism and organizing the evacuation of people from dangerous areas.

[0120] The early warning response set is pre-set by a person skilled in the art according to national and local emergency management norms and preplan systems, historical storm disaster data statistical analysis, target region hazard-bearing body vulnerability assessment and other multidimensional factors, so as to ensure that the early warning response scheme is scientific, targeted and executable.

[0121] In this embodiment, by introducing a graph neural network model, combining multi-source meteorological data and wind field characteristics, high-precision prediction of the spatial distribution of storm intensity and automatic delineation of the storm area are realized; by constructing an adaptive adjacency matrix that fuses geographical distance, terrain similarity and dominant wind direction transmission probability, the limitation of traditional methods that regard meteorological data as regular grids is broken through, the non-Euclidean space relationship of the storm system and the meteorological transmission physical mechanism are effectively described, and the expression ability of the spatial propagation characteristics of the storm system is improved; the graph convolutional neural network model is used to simultaneously capture the spatiotemporal dependence relationship of the storm evolution, and the prediction accuracy of the spatial distribution of storm intensity and the capture ability of local heavy precipitation characteristics are significantly improved; a multiple screening mechanism (including intensity threshold, double judgment of duration, time weighted fusion and scale constraint) is used to realize accurate delineation of the storm area, effectively avoiding misjudgment and omission caused by simple threshold judgment; based on multi-dimensional attribute indexes such as area size, intensity gradient entropy and pulsation intensity index, the comprehensive risk level is scientifically quantified, and a differentiated early warning response scheme is generated; the whole-process intelligent processing from storm prediction to risk assessment and emergency response is realized, effectively overcoming the shortcomings of existing technologies in complex nonlinear atmospheric system modeling, spatiotemporal correlation expression, area fine description and risk quantification assessment, thereby improving the accuracy of storm warning, the fine degree of area delineation and the scientificity of disaster prevention and mitigation decision-making.

[0122] Embodiment 2

[0123] The present application also provides an electronic device. The electronic device can include one or more processors and one or more memories. Wherein the memory has computer readable code stored therein, the computer readable code, when executed by the one or more processors, can perform the graph neural network-based storm intensity spatial distribution prediction and area delineation method as described above.

[0124] The method or system according to the embodiments of the present application can also be implemented by means of the architecture of the electronic device shown in the present application. The electronic device can include a bus, one or more CPUs, ROM, RAM, a communication port connected to a network, input / output, a hard disk, etc. The storage device in the electronic device, such as ROM or hard disk, can store the graph neural network-based storm intensity spatial distribution prediction and area delineation method provided by the present application. Further, the electronic device can also include a user interface. Of course, the architecture shown in the present application is only exemplary, and when implementing different devices, one or more components of the electronic device shown in the present application can be omitted according to actual needs.

[0125] Embodiment 3

[0126] Referring to the drawings, one embodiment of the present application discloses a computer readable storage medium. The computer readable storage medium stores computer readable instructions. When the computer readable instructions are run by a processor, the method for storm intensity spatial distribution prediction and rain area demarcation based on a graph neural network according to the embodiments of the present application described above with reference to the drawings can be executed. The storage medium includes, but is not limited to, for example, volatile memory and / or non-volatile memory. The volatile memory may, for example, include random access memory (RAM) and cache memory, etc. The non-volatile memory may, for example, include read-only memory (ROM), hard disk, flash memory, etc.

[0127] In addition, according to the embodiments of the present application, the processes described above with reference to the flowcharts can be implemented as a computer software program. For example, the present application provides a non-transitory machine readable storage medium storing machine readable instructions capable of being run by a processor to execute instructions corresponding to the method steps provided by the present application, for example: the method for storm intensity spatial distribution prediction and rain area demarcation based on a graph neural network. When the computer program is executed by a central processing unit (CPU), the above-mentioned functions defined in the method of the present application are executed.

[0128] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art can modify the technical solutions described in the foregoing embodiments or make equivalent replacements to some technical features. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

[0129] The formulas in the present specification are dimensionless numerical calculations, the formulas are obtained by software simulation of a large amount of data to obtain a formula of the most recent real situation, and the preset parameters and threshold values in the formula are set by a person skilled in the art according to the actual situation.

[0130] Although the embodiments of the present application have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, replacements and variations can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the claims and their equivalents.

Claims

1. A method for storm intensity spatial distribution prediction and rain area demarcation based on a graph neural network, characterized in that, The method comprises the following steps: Step S1: Divide the target area evenly into multiple spatial units, and use the geometric center of each spatial unit as a node. Collect multi-source meteorological data of the target area, and map the multi-source meteorological data to each node to construct a spatial node set. The multi-source meteorological data includes dynamic observation data and static observation data. The dynamic observation data includes... Rainfall, radar echo intensity, and cloud density at consecutive moments; static observation data includes topographic data. Step S2: Based on the set of spatial nodes, calculate the geographical distance, terrain similarity and dominant wind direction transmission probability between any two nodes, and obtain the edge weight value through fusion analysis, and construct an adaptive adjacency matrix according to all edge weight values; The method for calculating the geographical distance, terrain similarity and dominant wind direction transmission probability between any two nodes comprises: Obtain the latitude and longitude coordinates of each node, and calculate the spherical distance between any two nodes based on the latitude and longitude coordinates by using the formula of the secant of the half chord, and take the spherical distance as the geographical distance; According to the terrain and landform data of each node, the cosine similarity between any two nodes is calculated in turn, and taken as the terrain similarity; Obtain the historical wind field data corresponding to each node, the historical wind field data including wind direction angle and wind speed at multiple historical time points; according to a preset interval width, uniformly divide the value interval of the wind direction angle to obtain multiple wind direction intervals; from the multiple wind direction intervals, determine the dominant interval of each node in turn, and perform mean value calculation on the wind direction angle and wind speed in the dominant interval corresponding to each node respectively to obtain the dominant wind direction and average wind speed of each node; according to the latitude and longitude coordinates of each node, the direction angle of the connecting line between any two nodes is calculated; according to the geographical distance and the direction angle of the connecting line between any two nodes, and the dominant wind direction and average wind speed of each node, the dominant wind direction transmission probability between any two nodes is calculated; Step S3: For each node in the set of spatial nodes, generate a node feature vector according to the multi-source meteorological data mapped to the corresponding node, and aggregate to form a node feature matrix; Step S4: input the adaptive adjacency matrix and the node feature matrix into the pre-constructed graph convolutional neural network model, the graph convolutional neural network model performs weighted propagation on the node feature matrix based on the adaptive adjacency matrix, extracts the time evolution law in the node feature matrix, and outputs the storm intensity prediction value of each node at each future moment to form a storm intensity prediction matrix. Step S4: input the adaptive adjacency matrix and the node feature matrix into the pre-constructed graph convolutional neural network model, the graph convolutional neural network model performs weighted propagation on the node feature matrix based on the adaptive adjacency matrix, extracts the time evolution law in the node feature matrix, and outputs the storm intensity prediction value of each node at each future moment to form a storm intensity prediction matrix. Step S5: Perform spatial interpolation and threshold segmentation processing on the rainstorm intensity prediction matrix, identify rainstorm nodes based on the threshold segmentation processing result, and perform connected component analysis on all rainstorm nodes, extract the peripheral boundary contour of each connected component based on the spatial interpolation result, and generate a rainstorm area set; The method for generating the rainstorm area set comprises: Future By integrating the rainstorm nodes corresponding to each future moment in the given time period, we obtain... A set of rainstorm nodes is identified; connected component analysis is performed on each set of rainstorm nodes to obtain the connected component at each future time; the rainstorm intensity prediction values ​​with the same future time in the rainstorm intensity prediction matrix are integrated to obtain... A set of rainfall intensity ensembles is generated; spatial interpolation is then performed on the predicted rainfall intensity values ​​within each set of rainfall intensity ensembles to obtain the predicted rainfall intensity for each spatial location within the target area in the future. The predicted intensity of the rainstorm at a given moment; Compare each rainstorm intensity prediction value corresponding to each spatial position with the rainstorm threshold value respectively; the spatial position with a rainstorm intensity prediction value greater than the rainstorm threshold value is taken as a rainstorm position in the corresponding connected component; the boundary contour of each connected component is extracted according to the rainstorm position in each connected component, the peripheral boundary contour of each connected component is obtained, and taken as the time area; the time areas of each future time are fused to obtain a rainstorm area; all rainstorm areas are integrated to generate a rainstorm area set; Step S6: Statistically analyze the area attribute indexes of each rainstorm area in the rainstorm area set, calculate the comprehensive risk level of each rainstorm area based on the area attribute indexes, and generate a differentiated early warning response scheme for rainstorm areas with different comprehensive risk levels.

2. The storm intensity spatial distribution prediction and falling area delimitation method based on a graph neural network according to claim 1, characterized in that, The method for constructing the set of spatial nodes comprises: The radar echo intensity and the cloud layer density in the static observation data and the dynamic observation data are raster registered to obtain the topography and geomorphology data corresponding to each node and the radar echo intensity and the cloud layer density of each node at a continuous time The rainfall at a continuous time of each node is obtained by spatial interpolation on the rainfall in the dynamic observation data. The topography and geomorphology data corresponding to each node and the rainfall, the radar echo intensity and the cloud layer density of each node at a continuous time are integrated to obtain the node meteorological data of each node. The spatial node set is constructed according to all the nodes and the corresponding node meteorological data.

3. The storm intensity spatial distribution prediction and falling area delimitation method based on a graph neural network according to claim 2, characterized in that, The method for constructing the adaptive adjacency matrix comprises: Standardize the geographical distance between any two nodes in turn to obtain the standard distance between any two nodes; different correlation weights are set for the standard distance, terrain similarity and dominant wind direction transmission probability, and the inverse of the standard distance, terrain similarity and dominant wind direction transmission probability between any two nodes are weighted and summed in turn based on the correlation weights to obtain the edge weight value between any two nodes; The initial adjacency matrix is constructed, and all edge weights are sequentially added to the initial adjacency matrix; a screening threshold is calculated according to the edge weights in the initial adjacency matrix; the edge weights with values less than the screening threshold in the initial adjacency matrix are all adjusted to 0; The initial adjacency matrix with the adjusted edge weights is taken as the adaptive adjacency matrix.

4. The storm intensity spatial distribution prediction and falling area delimitation method based on a graph neural network according to claim 3, characterized in that, The method for aggregating the node feature matrix comprises: According to the node meteorological data of each node, the rainfall, radar echo intensity and cloud density at the same time in the same node are integrated to obtain the time point meteorological data of each node, and the topographic and geomorphic data of each node are sequentially added to the corresponding time point meteorological data; each group of time point meteorological data is input into the trained feature extraction model to predict the time point feature data corresponding to each group of time point meteorological data, and the feature extraction model is a deep neural network model; The time point feature data corresponding to each node is integrated to obtain the node feature vector of each node; the node feature vectors of each node are aggregated to form the node feature matrix of the target area.

5. The storm intensity spatial distribution prediction and falling area delimitation method based on a graph neural network according to claim 4, characterized in that, The method for forming the rainstorm intensity prediction matrix comprises: The adaptive adjacency matrix and the node feature matrix are input into a pre-constructed graph convolutional neural network model, the graph convolutional neural network model performs weighted propagation on the node feature matrix by a graph convolutional layer, and a weighted feature matrix is obtained; wherein the weighted feature matrix includes a weighted feature vector corresponding to each node; the graph convolutional neural network model processes the change of the weighted feature matrix at consecutive time points by a time convolutional layer, and extracts the time evolution law in the weighted feature matrix; The graph convolutional neural network model combines a weighted feature matrix and a time evolution law to generate a storm intensity prediction value of each node at a future time The storm intensity prediction value of each node at a future time is integrated to form a storm intensity prediction matrix. The storm intensity prediction value of each node at a future time is integrated to form a storm intensity prediction matrix.

6. The storm intensity spatial distribution prediction and falling area delimitation method based on a graph neural network according to claim 5, characterized in that, The method for identifying the rainstorm node comprises: Each rainstorm intensity prediction value corresponding to each node is compared with a preset rainstorm threshold, and the future time point with the rainstorm intensity prediction value greater than the rainstorm threshold is marked as a rainstorm time point; the maximum rainstorm duration corresponding to each node is counted according to the rainstorm time point corresponding to each node; the maximum rainstorm duration corresponding to each node is compared with a preset duration threshold, and the node with the maximum rainstorm duration greater than the duration threshold is marked as a rainstorm node.

7. The storm intensity spatial distribution prediction and falling area delimitation method based on a graph neural network according to claim 6, characterized in that, The method for fusing the time point landing area of each future time point comprises: For the future Each future time is respectively set with a corresponding time weight, and the rainstorm intensity prediction values corresponding to different future times at the same spatial position are weighted and fused based on the time weight to obtain a comprehensive rainstorm intensity corresponding to each spatial position. Each spatial position in each time point landing area is marked as a candidate position, and the rainstorm duration of each candidate position is counted according to the rainstorm intensity prediction value of each candidate position corresponding to different future time points; for each candidate position, the corresponding rainstorm duration is compared with the duration threshold, and the corresponding comprehensive rainstorm intensity is compared with the rainstorm threshold; the candidate position with the rainstorm duration greater than the duration threshold and the comprehensive rainstorm intensity greater than the rainstorm threshold is marked as a selected position; all selected positions are subjected to connected component analysis to obtain a plurality of candidate landing areas; The number of selected positions in each candidate landing area is counted and marked as a landing area size; the landing area size of each candidate landing area is compared with a preset size threshold, and the candidate landing area with the landing area size greater than the size threshold is taken as a rainstorm landing area.

8. The storm intensity spatial distribution prediction and falling area delimitation method based on a graph neural network according to claim 7, characterized in that, The method for calculating the comprehensive risk level of each rainstorm landing area comprises: Each index in the landing area attribute index corresponding to each rainstorm landing area is subjected to standardization processing to obtain the landing area standard index corresponding to each rainstorm landing area; the index weight corresponding to each standard index in the landing area standard index corresponding to each rainstorm landing area is sequentially calculated by using an entropy weight method. Setting positive ideal solution and negative ideal solution; calculating positive score between each rainstorm falling area corresponding to falling area standard index and positive ideal solution and negative score between each rainstorm falling area corresponding to falling area standard index and negative ideal solution, and calculating comprehensive score of each rainstorm falling area according to positive score and negative score; calculating ratio between negative score and comprehensive score of each rainstorm falling area, and obtaining risk score of each rainstorm falling area; Setting different risk intervals for different comprehensive risk grades, and comparing risk score of each rainstorm falling area with each risk interval in sequence; taking comprehensive risk grade corresponding to risk interval where risk score falls as comprehensive risk grade of corresponding rainstorm falling area; Wherein, falling area attribute indexes include falling area size, intensity gradient entropy and fluctuation intensity index.

Images