Multi-mode collaborative local intelligent thunder grading early warning method and system and storage medium

Abstract

The invention discloses a multi-mode cooperative local intelligent thunder and lightning grading early warning method and system and a storage medium, and relates to the technical field of meteorological disaster early warning, the multi-mode cooperative local intelligent thunder and lightning grading early warning method comprises the following steps: preprocessing three types of heterogeneous data sources of lightning positioning data, radar cloud picture data and atmospheric electric field data; a unified input reference of space-time alignment is constructed, and then the lightning movement state, the thundercloud movement direction and the lightning occurrence probability are obtained through a lightning movement track prediction module, a thundercloud movement direction obtaining module and a lightning occurrence probability calculation module; fusing the lightning moving state, the thundercloud moving direction and the lightning occurrence probability through a multi-mode dynamic fusion module to obtain a lightning comprehensive risk probability and lightning predicted arrival time, and analyzing the lightning comprehensive risk probability Prisk and the lightning predicted arrival time Tarrival through a grading early warning module to obtain a lightning grading early warning result. The system overcomes the limitation of a single data source, is high in adaptability, and achieves the precise protection and early warning of local thunder and lightning.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of meteorological disaster early warning, and in particular to a multi-modal cooperative local intelligent lightning hierarchical early warning method and system and a storage medium. BACKGROUND

[0002] Lightning disasters pose a great threat to lightning-sensitive areas such as cultural relics buildings, scenic spots, petrochemical plants, and wind farms. Lightning-induced major safety accidents occur frequently, so it is very important to provide early warning of lightning.

[0003] The prior art collects lightning data for analysis and early warning, but the protection capability is insufficient, mainly in the following three aspects: 1. Multi-source data cooperation aggravates local risks, the existing technology due to data island and fusion mechanism defects, leading to key scene protection failure; mainly manifested as radar, lightning positioning, atmospheric electric field data are usually scattered in different systems, the storage format is heterogeneous, and there is a lack of unified space-time reference, and most systems only realize data parallel connection (such as "radar + electric field" simple superposition), without building deep feature interaction model, single dependence is easy to cause high false alarm rate; in addition, in the early stage of thunderstorm, radar cloud tracking should be given priority to, and near warning needs to strengthen the electric field mutation signal, but the existing technology uses static threshold, which cannot adapt to the scene evolution, and lacks multi-modal weight self-adaptive mechanism. 2. The general model is seriously out of touch with the local scene, and the regional warning model of the meteorological department is usually divided into a grid greater than 5 kilometers. This rough warning mode cannot meet the fine demand of lightning-sensitive areas; for example, different terrains in scenic spots, complex facility layout of petrochemical plants, and high tower distribution of wind farms all require more accurate local warning; however, the general model cannot consider these local particularities, leading to protection failure, which makes the local area passive in lightning protection and difficult to effectively respond to sudden lightning disasters, increasing the risk of accidents. 3. The hierarchical strategy lacks scene adaptability: the existing lightning warning adopts a static hierarchical mechanism and cannot respond to local dynamic variables, increasing the risk of accidents; for example, when a thunderstorm strikes during a peak flow in a scenic spot, the warning needs to be dynamically upgraded, but the traditional system only relies on fixed thresholds. In addition, the flammable and explosive nature of petrochemical plants requires a more stringent response mechanism, but the general strategy cannot meet the hierarchical emergency needs of flammable and explosive places. SUMMARY

[0004] In order to overcome the defects of the prior art that lightning data is fragmented, has poor adaptability, and lacks scene customization, which makes it difficult to meet the demand for accurate local lightning protection, the present application provides a multi-modal cooperative local intelligent lightning hierarchical early warning method.

[0005] To achieve the above purpose, the present application adopts the following technical scheme, comprising:

[0006] A multi-modal cooperative local intelligent lightning hierarchical early warning method, comprising:

[0007] S1: collecting multi-modal lightning data of a local lightning geographical area, the multi-modal lightning data including lightning positioning data, radar cloud image data and atmospheric electric field data;

[0008] S2: preprocessing the multi-modal lightning data; the preprocessing including data cleaning, time and space reference unification, normalization processing and quality enhancement verification on the multi-modal lightning data;

[0009] S3: inputting the preprocessed multi-modal lightning data into a lightning early warning model to obtain a lightning grading early warning result;

[0010] The lightning early warning model includes a lightning moving track prediction module, a thundercloud moving direction obtaining module, a lightning occurrence probability calculation module, a multi-modal dynamic fusion processing module and a grading early warning module.

[0011] The lightning track prediction module is used for lightning track prediction on the lightning positioning data to obtain a lightning moving state T; the thundercloud moving direction obtaining module is used for processing the radar cloud image data to obtain a thundercloud moving direction D; the lightning occurrence probability calculation module is used for processing and analyzing the atmospheric electric field data to obtain a lightning occurrence probability p; the multi-modal dynamic fusion module is used for fusion processing of the lightning moving state T, the thundercloud moving direction D and the lightning occurrence probability p to obtain a lightning comprehensive risk probability and a lightning predicted arrival time; and the grading early warning module is used for analyzing the lightning comprehensive risk probability P risk and the lightning predicted arrival time T arrival to obtain a lightning grading early warning result.

[0012] Preferably, the processing process of the lightning track prediction module includes:

[0013] S311: performing grid division on the local lightning geographical area, regarding each grid as a node and regarding the relationship between adjacent grids as an edge to construct a lightning map structure, and each node feature including lightning frequency and lightning current intensity;

[0014] S312: obtaining a spatial embedding vector s v of each node based on the lightning map structure, including:

[0015] S3121: projecting each node v and its adjacent node u through a learnable weight matrix to obtain an original attention coefficient The calculation formula is:

[0016]

[0017] wherein, W k is a learnable weight matrix of the kth layer; a kis the learnable parameter vector of the attention mechanism of the kth layer; || represents concatenation; LeakyReLU is the leaky ReLU activation function; the superscript T is the transpose of the matrix; and are the k-1th layer features of node v and neighboring node u respectively;

[0018] S3122: Raw attention coefficients for all neighboring nodes of each node v Perform Softmax normalization to obtain normalized attention weights

[0019] S3123: Using normalized attention weights Projected features of neighboring nodes u Perform weighted summation and then use the activation function to get the k-th spatial embedding of node v

[0020] S3124: Get the spatial embedding vector s based on the spatial embedding of all layers of the node v ;

[0021] S313: Obtain the predicted node states of all nodes in the future based on the spatial embedding vector of each node, including:

[0022] S3131: For each time step t, the spatial embedding vectors of all nodes are stacked to form the spatial feature matrix of the current time step Where V is the total number of nodes;

[0023] S3132: The spatial feature matrix S of each time step t t Projection is query Q t , key K t Sum V t , and calculate the spatial attention weights between nodes by scaling the dot product attention, and weighted aggregation value V t , get the spatial attention A at the t-th time step t , the calculation formula is:

[0024] Q t =S t W q

[0025] K t =S t W k

[0026] V t =S t W v

[0027]

[0028] where Softmax() is row normalization; d v is the spatial embedding dimension; W q , W k and W v are learnable weight matrices.

[0029] S3133: input the spatial attention of all time steps into the feedforward network after stacking along the time dimension and perform time step segmentation to obtain the time attention T t .

[0030] S3134: fuse the spatial attention A t and the time attention T t through cross attention to obtain the spatio-temporal joint feature F t , including:

[0031] Calculate the time-guided spatial fusion matrix C t based on the spatial attention A t and the time attention T t :

[0032] C t = CrossAttention(Q = T t , K = A t , V = A t )

[0033] where Q comes from the time dimension, and K and V come from the spatial dimension.

[0034] Calculate the space-guided time fusion matrix D t based on the spatial attention A t and the time attention T t :

[0035] D t = CrossAttention(Q = A t , K = T t , V = T t )

[0036] where Q comes from the spatial dimension, and K / V come from the time dimension.

[0037] Concatenate the time-guided spatial fusion matrix and the space-guided time fusion matrix, and project it to the target dimension through a fully connected layer to obtain the spatio-temporal joint feature F t :

[0038] F t = [C t || D t ] · W f

[0039] wherein W f is a fusion weight matrix;

[0040] S3135: input the spatio-temporal joint feature matrix F t into the Transformer decoder to obtain all node states at the future K time steps:

[0041]

[0042] wherein, is the predicted state of all nodes at the future k-th time step;

[0043] S314: based on the predicted node state obtain the lightning movement track Track and the velocity vector w, including:

[0044] S3141: input the predicted node state into the fully connected layer to generate the lightning occurrence probability p of each grid at the future time step; t+k

[0045] S3142: extract the probability area exceeding the set probability threshold on the spatial grid to form a lightning activity thermal map;

[0046] S3143: based on the Viterbi algorithm, find the optimal path on the thermal map at consecutive time steps to obtain the lightning movement state T; the lightning movement state T includes the lightning movement track Track=(Tr T+1 ,Tr T+2 ,...,Tr T+K ) and the velocity vector w=(w T+1 ,w T+2 ,...,w T+K ).

[0047] Preferably, the processing procedure of the thundercloud movement direction acquisition module includes:

[0048] S321: based on the radar cloud image data, perform spatial feature extraction to obtain the thundercloud spatial feature f g ;

[0049] S322: based on the radar cloud image data, perform time series evolution prediction to obtain the time series evolution feature m T ;

[0050] S323: based on the thundercloud spatial feature f g and the time series evolution feature m T , perform feature fusion to obtain the thundercloud movement direction D.

[0051] Preferably, in step S323, based on the thundercloud spatial feature f g ​and time sequence evolution feature m T The feature fusion is performed to obtain a thundercloud moving direction D, including:

[0052] S3231: Based on the thundercloud spatial feature f g and time sequence evolution feature m T The fusion feature R is obtained fusion , and the calculation formula is:

[0053] R fusion =σ(W f ·[f g ;m T ])⊙f g +(1-σ(W f ·[f g ;m T ]))⊙m T

[0054] Wherein, sigma is a sigmoid function; W f is a learning weight;

[0055] S3232: The thundercloud fusion feature R fusion is input into a decoder, and k frames of future radar cloud images are output;

[0056] S3233: Each frame of radar cloud image is binarized, and the center point coordinates of the connected region are taken as the thundercloud center point coordinates;

[0057] S3234: The center point coordinates of the thundercloud in all frames of radar cloud images are used to form a centroid sequence, and linear regression is performed to obtain a thundercloud moving direction D, the thundercloud moving direction D including a direction angle theta and a speed v radar .

[0058] Preferably, the processing process of the lightning occurrence probability p calculation module includes:

[0059] S331: Different time scale features are extracted from the atmospheric electric field data E using a multi-scale convolution kernel, and then feature fusion is performed to obtain an atmospheric electric field feature matrix H;

[0060] S332: The lightning occurrence probability p is obtained based on the atmospheric electric field feature matrix H using a quantum support vector machine QSVM; including:

[0061] S3321: The atmospheric electric field feature matrix H is mapped to a quantum Hilbert space, and a kernel matrix K ij is defined:

[0062]

[0063] Wherein, k is the size of the convolution kernel; K ijrepresents the similarity between the i-th row and the j-th row; i and j represent the i-th row and the j-th row of the matrix, respectively; and represents the k-th element of the i-th row and the j-th row of the atmospheric electric field feature matrix H; θ k is an angle parameter that can be trained in the quantum kernel function; Π performs a multiplicative superposition on the extracted features of all convolution kernels; |·| 2 is the modulus square;

[0064] S3322: Obtain the optimal quantum support vector machine (QSVM) classification model and the optimal weight coefficient by optimizing the objective function, and the objective function is:

[0065]

[0066] wherein, α i , α j represent the weight coefficients of the i-th and j-th atmospheric electric field data samples, respectively, β i , β j represent the class labels of the i-th and j-th atmospheric electric field data samples, respectively;

[0067] S3323: Obtain the decision function f(x) based on the optimal weight coefficient, and the expression of the decision function is:

[0068]

[0069] wherein, b is a bias term;

[0070] S3324: Calculate the distance d of the atmospheric electric field data sample to the hyperplane based on the decision function, and the calculation formula is:

[0071]

[0072] wherein, ||ω|| is the norm of the weight vector ω;

[0073] S3325: Obtain the lightning occurrence probability p based on the sample distance d to the hyperplane, and the calculation formula is:

[0074]

[0075] wherein, r is a scaling parameter.

[0076] Preferably, the processing process of the multi-modal dynamic fusion module comprises:

[0077] S341: Introduce a dynamic weight distribution mechanism based on the lightning movement state T, the thundercloud movement direction D, and the lightning occurrence probability p to obtain the multi-modal fusion feature F fusion , and the calculation formula is:

[0078] F fusion= a T • T + a D • D + a p • p

[0079]

[0080] wherein q is a target modal identifier, taking values of T, D or p, h is a denominator summation index, traversing T, D and p; F fusion is a multi-modal fusion feature; a q is an attention mechanism weight; W q and W h are weight matrices; F q is a qthmodal feature, including lightning movement state T, thunder cloud movement direction D or lightning occurrence probability p; g is a gating function; sigma is a Sigmoid activation function; || represents splicing;

[0081] S342: obtaining lightning comprehensive risk probability P fusion and lightning estimated arrival time T risk based on multi-modal fusion feature F arrival ; including:

[0082] S3421: obtaining lightning comprehensive risk probability P fusion based on multi-modal fusion feature F risk , the calculation formula is:

[0083] P risk = sigma (W risk · F fusion + b risk )

[0084] wherein W risk is a risk weight vector; sigma is a sigmoid function; b risk is a risk probability bias term;

[0085] S3422: inputting multi-modal fusion feature F fusion into a GRU network to obtain lightning estimated arrival time T arrival .

[0086] Preferably, the processing process of the hierarchical warning module includes:

[0087] based on lightning comprehensive risk probability P risk and lightning estimated arrival time T arrival , the lightning warning result is obtained by dynamically adjusting the adaptive threshold; including:

[0088] S3511: obtaining three-level adaptive threshold P risk based on lightning comprehensive risk probability P adjusted , the calculation formula is:

[0089] P adjusted =P base ×[1+β×(P risk -1)]

[0090] Wherein, P base is the basic probability threshold; divided into first level P base1 , second level P base2 and third level P base3 , P base1 <P base2 <P base1 ; beta is the scene sensitivity coefficient; P risk is the scene risk index;

[0091] S3512: judging the comprehensive risk probability, when the comprehensive risk probability P risk <the first adaptive threshold, no warning;

[0092] When the first adaptive threshold is less than or equal to the comprehensive risk probability P risk <the second adaptive threshold, it is a lightning first level warning;

[0093] When the second adaptive threshold is less than or equal to the comprehensive risk probability P risk <the third adaptive threshold, it is a lightning second level warning;

[0094] When the third adaptive threshold is less than or equal to the comprehensive risk probability P risk , it is a lightning third level warning.

[0095] Preferably, it also includes:

[0096] S4: continuously monitoring the lightning activity, radar cloud data and atmospheric electric field data of the local lightning geographical area, setting the multi-modal lightning data threshold, and when the multi-modal lightning data of the local lightning geographical area is lower than the multi-modal lightning data threshold, the warning response is removed;

[0097] S5: updating the attention mechanism weight of the multi-modal dynamic fusion module based on the backtracking data, realizing the iterative optimization of the multi-modal collaborative local lightning grading warning.

[0098] A multi-modal collaborative local lightning grading warning system, comprising a data acquisition module, a data preprocessing module and a lightning warning model; the lightning warning model comprises a lightning moving track prediction module, a thundercloud moving direction acquisition module, a lightning occurrence probability calculation module, a cross-module dynamic fusion processing module and a grading warning module, to realize a multi-modal collaborative local intelligent lightning grading warning method.

[0099] A storage medium has a computer program stored thereon, the computer program is implemented when being executed to realize a multi-modal coordinated local intelligent lightning grading early warning method.

[0100] The advantages of the present application are:

[0101] (1) The present application preprocesses three types of heterogeneous data sources of lightning positioning data, radar cloud data and atmospheric electric field data, constructs a unified input reference of space-time alignment, then obtains lightning movement state, thunder cloud movement direction and lightning occurrence probability through lightning movement trajectory prediction module, thunder cloud movement direction acquisition module and lightning occurrence probability calculation module, then fuses lightning movement state, thunder cloud movement direction and lightning occurrence probability through multi-modal dynamic fusion module to obtain lightning comprehensive risk probability and lightning predicted arrival time, and the grading early warning module analyzes lightning comprehensive risk probability P risk and lightning predicted arrival time T arrival , and obtains lightning grading early warning results, overcoming the limitations of single data source, high adaptability, and realizing accurate protection and early warning of local lightning.

[0102] (2) The lightning movement trajectory prediction module of the present application constructs a lightning map structure, combines graph attention mechanism (GAT) and Transformer decoder, realizes dynamic prediction of lightning movement trajectory and speed, and solves the deficiency of traditional methods that can only predict static position.

[0103] (3) The thunder cloud movement direction acquisition module of the present application accurately captures the movement trend of thunder cloud through spatiotemporal feature fusion and centroid sequence regression of radar cloud.

[0104] (4) The present application introduces quantum support vector machine (QSVM) into atmospheric electric field data analysis, utilizes the parallel computing capability of quantum state, and still maintains high-precision lightning probability prediction under complex electric field environment.

[0105] (5) The present application introduces gating function and attention mechanism, realizes inter-modal nonlinear interaction gating attention network, dynamically fuses multi-modal features, inputs double-task layer, outputs lightning comprehensive risk probability and lightning predicted arrival time, predicts lightning arrival time through GRU network, designs dynamic weight distribution mechanism, automatically adjusts the contribution degree of each mode according to real-time data quality (such as reducing the weight of radar cloud when it is disturbed by precipitation), and ensures the stability of the fusion result.

[0106] (6) The present application obtains a three-level adaptive threshold P risk based on lightning comprehensive risk probability P adjusted , and obtains lightning comprehensive risk probability P riskDynamically adjust the adaptive threshold to obtain lightning warning results, which solves the problem of static threshold in traditional warnings and improves the practicality of warnings through dynamic fusion and scene adaptation. BRIEF DESCRIPTION OF THE DRAWINGS

[0107] Figure 1 This is a diagram of the steps for local lightning graded warning;

[0108] Figure 2 This is the architecture diagram of the lightning warning model;

[0109] Figure 3 This is a schematic diagram of a lightning locator;

[0110] Figure 4 This is a schematic diagram of weather radar;

[0111] Figure 5 Schematic diagram of atmospheric electric field data acquisition equipment. DETAILED DESCRIPTION

[0112] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0113] like Figures 1-5 As shown, the present invention proposes a multi-modal collaborative local intelligent lightning graded warning method, including:

[0114] S1: Collect multimodal lightning data, which includes lightning location data, radar cloud image data, and atmospheric electric field data.

[0115] Lightning location data is collected using high-precision lightning locators, such as Figure 2 As shown in the figure, each lightning locator at each monitoring site is equipped with a fast antenna and data acquisition system. The fast antenna has a reception time constant of approximately 200 μs, and the received electric field waveform has a frequency of 500 Hz to 500 kHz. To achieve high timing accuracy, a GPS clock is used to synchronize signals arriving at different sites. The data acquisition system uses a high-sampling analog-to-digital converter (ADC) with a sampling rate of 10 MS / s, achieving a timing accuracy of 0.1 μs.

[0116] When the monitoring site receives lightning data, it is returned to the central server through a 4G / 5G signal, and then the lightning data collected by multiple monitoring sites is used to accurately obtain the time and location (latitude, longitude, and height) information of lightning occurrence using the time of arrival (TOA) method.

[0117] Radar cloud data is mainly collected by S-band and C-band weather radars, such as Figure 3 As shown, after signal processing and algorithm inversion, the composite reflectivity (CR) is obtained, which is mainly used to identify severe convective weather (such as hail, heavy rain, etc.).

[0118] The CR product is a multi-layer reflectivity data obtained based on the radar volume scan mode. All elevation angle reflectivity data for the same geographic location are compared, and the maximum value is selected as the CR value (composite reflectivity) for that geographic location. Then, the polar coordinates are converted to Cartesian coordinates, and radar cloud data with latitude and longitude grid can be obtained. The radar cloud data includes several frames of radar cloud images.

[0119] Atmospheric electric field data is collected by a low false alarm thunderstorm warning instrument, such as Figure 4 The low false alarm thunderstorm warning instrument uses the principle of electrostatic shielding, and its core structure consists of a rotating metal rotor and a fixed stator. The rotor periodically cuts the external electric field, forming alternating shielding and exposure states, so that the stator induces an alternating charge signal proportional to the atmospheric electric field strength to measure the atmospheric electric field.

[0120] S2: Preprocessing of multi-modal lightning data;

[0121] To improve data quality reliability and ensure the spatiotemporal consistency of multi-modal lightning data, preprocessing of multi-modal lightning data is required before data processing, including:

[0122] S21: Data cleaning

[0123] For lightning location data: eliminate single monitoring site false alarm records (such as no synchronous lightning at multiple monitoring sites, marked as abnormal), fill in short-time missing values (interpolate with adjacent monitoring sites / time series);

[0124] For radar cloud data: suppress ground clutter (such as filtering low-altitude false echoes based on terrain mask), repair scanning blank area (spatial interpolation completion);

[0125] For atmospheric electric field data: filter electromagnetic interference spikes (sliding window mean filtering), correct baseline drift caused by sensor temperature drift (long-term trend fitting denoising).

[0126] S22: Spatio-temporal reference unification

[0127] The lightning positioning data (latitude and longitude + timestamp), radar cloud data (grid projection + time slice), and atmospheric electric field data (site coordinates + sampling time) are unified to the "WGS84 latitude and longitude + China Standard Time" reference, ensuring that multi-source data can be aligned and fused in the spatio-temporal dimension, forming spatio-temporally aligned original input, and laying the foundation for multi-modal fusion.

[0128] S23: Data standardization

[0129] Z-score normalization is performed on numerical features such as echo intensity and electric field amplitude to eliminate dimensional differences and provide "same scale, comparable" input for algorithm models.

[0130] S24: Quality enhancement verification

[0131] Lightning positioning data: cross-site consistency verification (multi-site joint positioning data reliability greater than a set threshold to be retained);

[0132] Radar cloud data: beam blocking correction (combined with digital elevation model DEM to correct echoes in mountainous areas);

[0133] Atmospheric electric field data: long-period trend calibration (eliminate baseline shift caused by sensor aging).

[0134] S3: Input the preprocessed multi-modal lightning data into the lightning warning model to obtain lightning multi-level warning results.

[0135] The lightning warning model includes a lightning movement trajectory prediction module, a thundercloud movement direction acquisition module, a lightning occurrence probability calculation module, a multi-modal dynamic fusion processing module, and a hierarchical warning module.

[0136] The lightning trajectory prediction module is used to predict the lightning trajectory based on lightning positioning data; the thundercloud movement direction acquisition module is used to process radar cloud data to obtain the thundercloud movement direction; the lightning occurrence probability calculation module is used to process and analyze atmospheric electric field data to obtain the lightning occurrence probability; the multi-modal dynamic fusion module is used to fuse lightning movement trajectory, thundercloud movement direction, and lightning occurrence probability to obtain lightning comprehensive risk probability and lightning predicted arrival time; the hierarchical warning module is used to analyze lightning comprehensive risk probability and lightning predicted arrival time to obtain lightning hierarchical warning results.

[0137] The processing process of the lightning trajectory prediction module includes:

[0138] S311: Graph structure construction

[0139] ① Spatial Griding: Divide the local lightning geography area into 1km x 1km grid (or adaptive grid), each grid contains the statistical characteristics of lightning events in a period of time (such as: lightning frequency, average current intensity).

[0140] ② Graph Structure Construction:

[0141] a. Node: Each spatial grid is a node.

[0142] b. Edge: The relationship between adjacent grids is constructed based on spatial proximity (such as Delaunay triangulation) or physical distance threshold (such as connecting grids within 5km), reflecting the spatial correlation of lightning conduction.

[0143] c. Dynamic graph structure update: When the grid lightning density > 5 times / km 2 , switch to 0.5km refined grid.

[0144] ③ Node Features: Each node contains time series features, such as:

[0145] a. Short-term features: Lightning frequency, current intensity average in the past 10 minutes.

[0146] b. Long-term features: Lightning frequency trend (calculated by moving average) in the past 1 hour.

[0147] S312: Spatial dependence relationship construction

[0148] Objective: Learn the spatial conduction relationship between nodes (such as lightning chain reaction in adjacent areas).

[0149] Model selection: Graph neural network GraphSAGE + GAT hybrid architecture

[0150] Step 1: GraphSAGE adjacent node aggregation

[0151] For each node v, sample a fixed number of adjacent nodes (such as 30) from its original adjacent node set N(v) (all nodes adjacent to node v), get the sampled adjacent node set

[0152] Step 2: GAT attention weighting

[0153] For each node, based on the adjacent node set N s (v), calculate the dynamic attention weight between nodes through the graph attention mechanism (GAT), and weightedly aggregate the adjacent node features to get the spatial embedding of node v in the current layer k The specific process is as follows:

[0154] (1) Attention coefficient calculation

[0155] For node v and its sampled neighbor node u (u∈N s (v)), the last layer embedding of both (v) and (u) are projected the feature of node v at k-1th layer; the feature of neighbor node u at k-1th layer) through a learnable weight matrix W k to the same dimension, and then concatenated and activated by LeakyReLU to get the original attention coefficient which reflects the spatial correlation strength between node v and neighbor node u:

[0156]

[0157] where, is the weight matrix of kth layer (projecting the d k-1 dimensional feature of last layer to the d k dimensional feature of current layer) ; is the learnable parameter vector of attention mechanism; || represents concatenation; LeakyReLU is the LeakyReLU activation function; superscript T represents the transpose of matrix; and are the k-1th layer features of node v and neighbor node u, respectively.

[0158] (2) Attention weight Normalization

[0159] The original attention coefficients of all neighbor nodes of node v are normalized by Softmax to get the normalized attention weight which represents the spatial influence degree of neighbor node pair, the greater the weight, the stronger the influence:

[0160]

[0161] (3) Weighted aggregation of neighbor node features

[0162] The projection feature of neighbor node u is weighted summed with the normalized attention weight , and then activated by an activation function (such as ReLU) to get the kth layer spatial embedding of node v

[0163]

[0164] where, σ() is the sigmoid function;

[0165] (4) Get the spatial embedding vector s v based on the spatial embedding of all layers of the node.

[0166] S313: Based on the spatial embedding vector s of each node v Obtain the predicted node state of each node at the future time step.

[0167] For each time step t, the spatial embedding vectors of all nodes Stacking forms the spatial feature matrix S of the current moment V is the total number of nodes, d s is the spatial embedding dimension.

[0168] First layer: Spatial self-attention (Spatial Attention)

[0169] Objective: For each time step t, model the global spatial dependency between all nodes at each time step. For each time step t, the spatial feature matrix S t Projection as query (Q t ), key (K t ), value (V t ) (for attention calculation):

[0170] Q t = S t W q

[0171] K t = S t W k

[0172] V t = S t W v

[0173] where W q , W k and W v are all learnable weight matrices; Q t , K t and V t are the query, key and value of the t-th time step, respectively.

[0174] Based on the query, key and value, the spatial attention weight between nodes is calculated by scaled dot-product attention, and the value V t is weighted and aggregated to obtain the spatial attention A t of the t-th time step:

[0175]

[0176] where Softmax() is row normalization; d v is the spatial embedding dimension.

[0177] Second layer: Temporal Attention

[0178] Objective: Model the global temporal dependency of node features evolution over time

[0179] Stack all the spatial attention outputs A1, A2,..., A t ,...,A T over time dimension to form a global spatio-temporal feature matrix:

[0180] H = [A1, A2,..., A t ,...,A T ] T

[0181] where A t is the spatial attention at time step t.

[0182] Input the spatial attention temporal matrix H into a Transformer Block (contains multi-head self-attention and feed-forward neural network) to model the dependency across time steps, output the temporal attention output matrix

[0183]

[0184] For each time step t, extract the slice of corresponding time step from , get the temporal attention T t , which represents the state of all nodes at time step t after integrating the global spatio-temporal dependency.

[0185] Third layer: Spatio-Temporal Joint Decoding

[0186] Fuse the spatial attention A t (space dimension: inter-node relationship) and the temporal attention T t (time dimension: node's own time evolution) through Cross-Attention to get the spatio-temporal joint feature F t (containing joint information of spatial relationship and time evolution), including:

[0187] Calculate the time-guided spatial fusion matrix C t at time t based on the spatial attention A t and the temporal attention T t :

[0188] C t = CrossAttention(Q = T t , K = A t , V = A t )

[0189] where Q comes from the time dimension, and K / V comes from the spatial dimension.

[0190] Based on spatial attention A t and temporal attention T t Compute the spatially guided temporal fusion matrix D t at time t:

[0191] D t = CrossAttention(Q=A t , K=T t , V=T t )

[0192] where Q comes from the time dimension, and K / V comes from the spatial dimension.

[0193] Concatenate the cross-attention outputs of both directions, and project to the target dimension through a fully connected layer to get the spatio-temporal joint feature F t at time t:

[0194] F t = [C t ||D t ]·W f

[0195] where W f is the fusion weight matrix.

[0196] Input the spatio-temporal joint feature matrix F t into the Transformer Decoder (decoder modeling future time steps) to predict the state of all nodes at the future 6 time steps (10 minutes per step):

[0197]

[0198] where S is the predicted state of all nodes at the future k-th time step.

[0199] S314: Obtain the lightning movement trajectory and velocity vector based on the predicted node state, including:

[0200] Input the predicted node state into a fully connected layer to generate the lightning occurrence probability p t+k at each grid at the future time step.

[0201] Use non-maximum suppression (NMS) to extract high-probability regions (p t+k > 0.7) on the spatial grid to form a lightning activity heat map.

[0202] Based on the Viterbi algorithm, the optimal path is searched on the heat map of the continuous time steps, and the lightning moving state T is output, the lightning moving state T includes a lightning moving track Track=(Tr T+1 ,Tr T+2 ,...,Tr T+6 ) and a speed vector w=(w T+1 ,w T +2 ,...,w T+6 ), as one of the input features of the multi-modal dynamic fusion module.

[0203] The processing process of the thundercloud moving direction acquisition module includes:

[0204] The multi-scale ViT architecture (MS-ViT) is used to process the radar cloud data (radar image), and the global structure features (such as thundercloud group shape intensity center) of a single frame of radar cloud data are captured.

[0205] S321: Based on the radar cloud data, spatial feature extraction is performed to obtain thundercloud spatial features f g , including:

[0206] 1. The radar cloud data with size HxW is divided into patches, each patch is flattened into a vector, and linearly projected into a D-dimensional embedding:

[0207] z0=[x 1 W;x 2 W;...;x i W;...;x N W]+E pos

[0208] Where x i is the pixel vector of the i-th patch, W is the projection matrix, and E pos is the position encoding matrix.

[0209] 2. Input z0 into the L-layer Transformer encoder for iterative processing, and finally obtain the thundercloud spatial feature f g .

[0210] S322: Based on the current frame of radar cloud data X t , the long short-term memory network LSTM is used for time series evolution prediction to obtain the time series evolution feature m T .

[0211] Objective: To learn the time series evolution law of thundercloud motion (such as moving speed splitting / merging).

[0212] The long short-term memory network LSTM maintains vertical (inter-layer) and horizontal (time) memory flow.

[0213] S323: based on the spatial features f g and the time evolution features m T feature fusion and decoding to obtain the moving direction D of the thundercloud;

[0214] Fusion strategy:

[0215] Gating fusion unit: based on the spatial features f g and the time evolution features m T to obtain the fused features R fusion :

[0216] R fusion = σ(W f ·[f g ; m T ])⊙f g +(1-σ(W f ·[f g ; m T ]))⊙m T

[0217] where σ is the sigmoid function, and W f is the learning weight.

[0218] Based on the fused features R fusion of the thundercloud, the moving direction D is obtained, including:

[0219] ①Predicting the cloud sequence: input the fused features R fusion of the thundercloud into the decoder, and output the future k frames of radar cloud images;

[0220] ②Thundercloud center positioning: binarize each frame of radar cloud image (threshold = 30 dBZ) → calculate the centroid of the connected region as the thundercloud center point;

[0221] ③Moving vector fitting: perform linear regression on the centroid sequence {(x1,y1),(x1,y1),...,(x k ,y k )} → obtain the moving direction D of the thundercloud (including the direction angle θ and the speed v radar ), which is used in the multi-modal dynamic fusion processing module for arrival time prediction.

[0222] The processing process of the lightning occurrence probability calculation module includes:

[0223] Through the Multi-scale Informer model, the multi-scale time sequence features (short-term fluctuations / long-term periods) of the atmospheric electric field data are extracted, the quantum kernel mapping of the quantum support vector machine QSVM is combined to enhance the small sample classification ability, and high-precision prediction of lightning probability is realized.

[0224] S331: Multi-scale Informer feature extraction is performed on the atmospheric electric field data E to obtain an atmospheric electric field feature matrix H, including:

[0225] First, different time scale features are extracted from the atmospheric electric field data E using multi-scale convolution kernels:

[0226] F k (E) = Conv 1D (E, kernel_size = k), k ∈ {3, 6, 12}

[0227] where k is the size of the convolution kernel;

[0228] Then, feature fusion is performed to output the atmospheric electric field feature matrix H:

[0229] H = Concat([F3(E), F6(E), F 12 (E)])

[0230] where F3(E) is a 30-minute data feature, F6(E) is a 60-minute data feature, and F 12 (E) is a 120-minute data feature.

[0231] S332: Obtain the lightning occurrence probability p based on the atmospheric electric field feature matrix H using a quantum support vector machine QSVM, including:

[0232] S3321: Map the atmospheric electric field feature matrix H to a quantum Hilbert space, and define a kernel matrix K ij :

[0233]

[0234] where k is the size of the convolution kernel; K ij represents the similarity between the i-th row and the j-th row; i and j represent the i-th row and the j-th row of the matrix, respectively; and represent the k-th element of the i-th row and the j-th row of the atmospheric electric field feature matrix H, respectively; θ k is a trainable angle parameter in the quantum kernel function; Π is a concatenation superposition of features extracted by all convolution kernels; |·| 2 is the modulus square.

[0235] S3322: Optimize the objective function to obtain the optimal quantum support vector machine QSVM classification model and the optimal weight coefficient, and the objective function is:

[0236]

[0237] where α i , αj weight coefficients of the i-th and j-th atmospheric electric field data samples, respectively i , β j class labels of the i-th and j-th atmospheric electric field data samples, respectively

[0238] S3323: Obtain a decision function f(x) based on the optimal weight coefficients, and the expression of the decision function is:

[0239]

[0240] wherein b is a bias term, which can be solved by support vector constraint;

[0241] S3324: Calculate the distance d of the atmospheric electric field data sample to the hyperplane based on the decision function, and the calculation formula is:

[0242]

[0243] wherein ||ω|| is the norm of the weight vector ω;

[0244] S3325: Obtain the lightning occurrence probability p based on the distance d of the sample to the hyperplane, and the calculation formula is:

[0245]

[0246] wherein r is a scaling parameter. When the positive and negative samples of the training data are unbalanced, r needs to be adjusted through cross-validation to optimize the early warning threshold;

[0247] The processing process of the multi-modal dynamic fusion module includes:

[0248] The fusion architecture is a Gated Attention Network (GAN)

[0249] S341: Introduce a dynamic weight distribution mechanism based on the lightning movement state T = [Track, w], the thundercloud movement direction D = [θ, v radar ] and the lightning occurrence probability p, and adaptively adjust the modal weight according to the thunderstorm evolution stage to obtain a multi-modal fusion feature F fusion , and the calculation formula is:

[0250] F fusion = α T ·T + α D ·D + α p ·p

[0251]

[0252] wherein q is a target modal identifier, taking values of T, D or p, and h is a denominator summation index, traversing T, D and p; Ffusion is the multi-modal fusion feature; a q is the attention mechanism weight; W q and W h are weight matrices; F q is the q-th modal feature, including lightning movement state T, thundercloud movement direction D or lightning occurrence probability p; g is a gating function; σ is a Sigmoid activation function; and || represents concatenation.

[0253] S342: Based on the multi-modal fusion feature F fusion , a double-task decoding layer is added to output lightning comprehensive risk probability P risk and lightning predicted arrival time T arrival .

[0254] ① Lightning comprehensive risk probability (Lightning Risk Probability)

[0255] Structure: fully connected layer + Sigmoid activation, and the calculation formula is:

[0256] P risk = σ(W risk · F fusion + b risk )

[0257] wherein, W risk is a risk weight vector, and b risk is a risk probability bias term.

[0258] ② The multi-modal fusion feature F fusion is input into a spatio-temporal regressor to obtain lightning predicted arrival time T arrival , and the spatio-temporal regressor contains a position-encoding GRU network (capturing movement trend), and the formula is:

[0259] T arrival = GRU(F fusion )

[0260] wherein, T arrival is in minutes.

[0261] Lightning comprehensive risk probability (P risk ) and lightning predicted arrival time (T arrival ) are directly input into a lightning grading early warning module to trigger a three-level early warning decision, and a lightning grading early warning result is obtained.

[0262] The processing process of the grading early warning module includes:

[0263] The output of the multi-modal dynamic fusion processing module is converted into actionable warning levels, which are combined with scene variables to execute graded responses.

[0264] S351: Dynamic adjustment of warning levels based on lightning comprehensive risk probability

[0265] Scene-adaptive threshold optimization:

[0266] P adjusted = P base × [1 + β × (P risk - 1)]

[0267] Where P base is the basic probability threshold (0.55 for level 1, 0.75 for level 2, and 0.90 for level 3), β is the scene sensitivity coefficient (0.15 for scenic areas, 0.20 for wind power plants, and 0.30 for oil storage areas), and P risk is the scene risk index (calculated from real-time monitoring data). The warning level judgment rules are shown in Table 1.

[0268] Table 1: Warning Level Judgment Table

[0269]

[0270] S352: Graded Response Strategy as shown in Table 2

[0271] Table 2: Graded Response Measures

[0272]

[0273] S4: Warning Cancellation

[0274] Cancellation Conditions:

[0275] Spatial Dimension: No lightning activity within 10 km of the local lightning geographical area (for 15 minutes)

[0276] Electric Field Dimension: Atmospheric Electric Field Intensity < 1 kV / m

[0277] Radar Dimension: CR Echo Intensity < 25 dBZ

[0278] S5: Backtracking Analysis Process:

[0279] False Alarm Analysis: Compare the warning trigger conditions with the actual lightning location (positioning error < 500 m is considered valid) Missed Alarm Trace: Check the sensor failure period (more than 5 minutes of electric field instrument disconnection is marked as abnormal)

[0280] Model Optimization: Update the attention mechanism weights (α T , α D , and α p ) of the multi-modal dynamic fusion module based on the backtracking data to achieve monthly iterative optimization.

[0281] The application discloses a multi-modal cooperative local lightning grading early warning system, which comprises a data acquisition module, a data preprocessing module and a lightning early warning model.

[0282] The application discloses a multi-modal cooperative local lightning grading early warning system, which comprises a data acquisition module, a data preprocessing module and a lightning early warning model.

[0283] Of course, the application is not limited to the details of the above exemplary embodiments, but also includes the same or similar structures which can be realized in other specific forms without departing from the spirit or essential characteristics of the application. Therefore, the embodiments should be regarded as exemplary and non-limiting, and the scope of the application is defined by the appended claims rather than the above description, and all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be included in the application. Any reference signs in the claims should not be regarded as limiting the claims involved.

[0284] In addition, it should be understood that, although the present specification is described in terms of embodiments, not every embodiment contains only one independent technical solution, and the description of the specification is only for the sake of clarity, and those skilled in the art should regard the specification as a whole, and the technical solutions in each embodiment can be combined appropriately to form other embodiments which can be understood by those skilled in the art.

[0285] The technical, shape, structure parts not described in detail in the application are well-known technologies.

Claims

1. A multi-modal collaborative local intelligent lightning graded warning method, characterized in that: include: S1: Collect multimodal lightning data in the local lightning geographical area. The multimodal lightning data includes lightning location data, radar cloud image data and atmospheric electric field data. S2: Preprocessing of multimodal lightning data; preprocessing includes data cleaning, temporal and spatial benchmark unification, normalization, and quality enhancement verification. S3: Input the pre-processed multi-modal lightning data into the lightning warning model to obtain the lightning graded warning results; The lightning warning model includes a lightning movement trajectory prediction module, a thundercloud movement direction acquisition module, a lightning occurrence probability calculation module, a multimodal dynamic fusion processing module, and a graded warning module. The lightning trajectory prediction module is used to predict the lightning trajectory based on the lightning location data to obtain the lightning movement state T; the thundercloud movement direction acquisition module is used to process the radar cloud image data to obtain the thundercloud movement direction D; the lightning occurrence probability calculation module is used to process and analyze the atmospheric electric field data to obtain the lightning occurrence probability p; the multimodal dynamic fusion module is used to fuse the lightning movement state T, the thundercloud movement direction D, and the lightning occurrence probability p to obtain the comprehensive lightning risk probability and the expected arrival time of lightning; the graded warning module is used to calculate the comprehensive lightning risk probability P risk and the lightning estimated arrival time T arrival Conduct analysis and obtain lightning graded warning results.

2. A multi-modal collaborative local intelligent lightning graded warning method according to claim 1, characterized in that: The processing of the lightning trajectory prediction module includes: S311: Divide the local lightning geographical area into grids, treat each grid as a node, and the relationships between adjacent grids as edges to construct a lightning graph structure. The characteristics of each node include lightning frequency and lightning current intensity. S312: Obtain the spatial embedding vector s of each node based on the lightning graph structure v ,include: S3121: For each node v and its neighboring nodes u, the original attention coefficient is obtained by projecting the learnable weight matrix The calculation formula is: Among them, W k is the learnable weight matrix of the kth layer; a k is the learnable parameter vector of the attention mechanism of the kth layer; || represents concatenation; LeakyReLU is the leaky ReLU activation function; the superscript T is the transpose of the matrix; and are the k-1th layer features of node v and neighboring node u respectively; S3122: Raw attention coefficients for all neighboring nodes of each node v Perform Softmax normalization to obtain normalized attention weights S3123: Using normalized attention weights Projected features of neighboring nodes u Perform weighted summation and then use the activation function to get the k-th spatial embedding of node v S3124: Get the spatial embedding vector s based on the spatial embedding of all layers of the node v ; S313: Obtain the predicted node states of all nodes in the future based on the spatial embedding vector of each node, including: S3131: For each time step t, the spatial embedding vectors of all nodes are stacked to form the spatial feature matrix of the current time step Where V is the total number of nodes; S3132: The spatial feature matrix S of each time step t t Projection is query Q t , key K t Sum V t , and calculate the spatial attention weights between nodes by scaling the dot product attention, and weighted aggregation value V t , get the spatial attention A at the t-th time step t , the calculation formula is: Q t =S t W q K t =S t W k V t =S t W v Among them, Softmax() is row normalization; d v is the spatial embedding dimension; W q 、W k and W v Both are learnable weight matrices; S3133: The spatial attention of all time steps is stacked along the time dimension and input into the feedforward network and the time steps are split to obtain the time attention T t ; S3134: Spatial Attention A t With temporal attention T t The spatiotemporal joint feature F is obtained by spatiotemporal cross attention fusion t ,include: Based on spatial attention t and temporal attention T t Calculate the time-guided spatial fusion matrix C t : C t =CrossAttention(Q=T t ,K=A t ,V=A t ) Among them, Q comes from the time dimension, K and V come from the space dimension; Based on spatial attention t and temporal attention T t Calculate the spatial guided temporal fusion matrix D t : D t =CrossAttention(Q=A t ,K=T t ,V=T t ) Among them, Q comes from the spatial dimension, and K / V comes from the time dimension; The time-guided spatial fusion matrix and the space-guided time fusion matrix are concatenated and projected to the target dimension through the fully connected layer to obtain the spatiotemporal joint feature F t : F t =[C t ||D t ]·W f Among them, W f is the fusion weight matrix, || represents splicing; S3135: The spatiotemporal joint feature matrix F t Input the Transformer decoder to get the predicted states of all nodes in the next K time steps: in, is the predicted state of all nodes in the kth time step in the future; S314: Based on the predicted node status Get the lightning trajectory Track and velocity vector w, including: S3141: Predicted node status Input the fully connected layer to generate the probability p of lightning occurrence for each grid in the future time step t+k ; S3142: Extract probability areas exceeding a set probability threshold on the spatial grid to form a lightning activity heat map; S3143: Based on the Viterbi algorithm, the optimal path is found on the heat map of the continuous time step to obtain the lightning movement state T; the lightning movement state T includes the lightning movement trajectory Track = (Tr T+1 ,Tr T+2 ,...,Tr T+K ) and velocity vector w=(w T+1 ,w T+2 ,...,w T+K ).

3. The multi-modal collaborative local intelligent lightning graded warning method according to claim 1, characterized in that: The processing of the thundercloud movement direction acquisition module includes: S321: Extract spatial features based on radar cloud image data to obtain thundercloud spatial features f g ; S322: Predict the time series evolution based on radar cloud image data and obtain the time series evolution feature m T ; S323: Based on thundercloud spatial characteristics f g and temporal evolution characteristics m T Perform feature fusion to obtain the thundercloud moving direction D.

4. The multi-modal collaborative local intelligent lightning graded warning method according to claim 3, characterized in that: In step S323, based on the thundercloud spatial feature f g and temporal evolution characteristics m T The feature fusion is performed to obtain the thundercloud movement direction D, including: S3231: Based on the spatial characteristics of thunderclouds g and temporal evolution characteristics m T Get the fusion feature R fusion , the calculation formula is: R fusion =σ(W f ·[f g ;m T ])⊙f g +(1-σ(W f ·[f g ;m T ]))⊙m T Where σ is the sigmoid function; W f To learn weights; S3232: Combine thundercloud features R fusion Input decoder, output future k frames of radar cloud images; S3233: Binarize each frame of radar cloud image and use the centroid of the connected area as the coordinate of the thundercloud center point; S3234: Based on the coordinates of the thundercloud center points of all frames of radar cloud images, a centroid sequence is formed, and then linear regression is performed to obtain the thundercloud movement direction D, which includes the direction angle θ and the speed v radar .

5. The multi-modal collaborative local intelligent lightning graded warning method according to claim 1, characterized in that: The processing of the lightning occurrence probability p calculation module includes: S331: Using a multi-scale convolution kernel to extract features of different time scales from the atmospheric electric field data E, and then performing feature fusion to obtain the atmospheric electric field feature matrix H; S332: Obtaining the lightning occurrence probability p using the quantum support vector machine QSVM based on the atmospheric electric field characteristic matrix H; including: S3321: Map the atmospheric electric field characteristic matrix H to quantum Hilbert space and define the kernel matrix K ij : Among them, k is the convolution kernel size; K ij Represents the similarity between the i-th row and the j-th row; i and j represent the i-th row and j-th row of the matrix respectively; and represent the kth element of the i-th row and j-th row of the atmospheric electric field characteristic matrix H respectively; θ k is the trainable angle parameter in the quantum kernel function; ∏ performs multiplication and superposition on the features extracted by all convolution kernels; |·| 2 is modulo square; S3322: Optimize the objective function to obtain the optimal quantum support vector machine QSVM classification model and the optimal weight coefficient. The objective function is: Among them, α i , α j Represent the weight coefficients of the i-th and j-th atmospheric electric field data samples, β i , β j Represent the category labels of the i-th and j-th atmospheric electric field data samples respectively; S3323: Obtain a decision function f(x) based on the optimal weight coefficient. The decision function expression is: Among them, b is the bias term; S3324: Calculate the distance d from the atmospheric electric field data sample to the hyperplane based on the decision function. The calculation formula is: Among them, ||ω|| is the norm of the weight vector ω; S3325: Based on the distance d between the sample and the hyperplane, obtain the probability p of lightning occurrence. The calculation formula is: Where r is the scaling parameter.

6. The multi-modal collaborative local intelligent lightning graded warning method according to claim 1, characterized in that: The processing of the multimodal dynamic fusion module includes: S341: Based on the lightning movement state T, thundercloud movement direction D and lightning occurrence probability p, a dynamic weight distribution mechanism is introduced to obtain the multimodal fusion feature F fusion , the calculation formula is: F fusion =a T ·T+a D ·D+a p ·p Where q is the target modality identifier, which takes the value of T, D or p, h is the denominator sum index, traversing T, D and p; F fusion is the multimodal fusion feature; α q is the attention mechanism weight; W q and W h is the weight matrix; F q is the qth modal feature, including the lightning movement state T, the thundercloud movement direction D or the lightning occurrence probability p; g is the gating function; σ is the Sigmoid activation function; || represents splicing; S342: Based on multimodal fusion feature F fusion Get the comprehensive lightning risk probability P risk and the lightning estimated arrival time T arrival ;include: S3421: Based on multimodal fusion feature F fusion Get the comprehensive lightning risk probability P risk , the calculation formula is: P risk =σ(W risk ·F fusion +b risk ) Among them, W risk is the risk weight vector; σ is the sigmoid function; b risk Risk probability bias term; S3422: Multimodal fusion feature F fusion Input the GRU network to obtain the estimated arrival time T of the lightning arrival .

7. The multi-modal collaborative local intelligent lightning graded warning method according to claim 1, characterized in that: The processing of the graded warning module includes: Based on the lightning comprehensive risk probability P risk and the lightning estimated arrival time T arrival Perform adaptive threshold dynamic adjustment to obtain lightning warning results; including: S3511: Based on the lightning comprehensive risk probability P risk Get the three-level adaptive threshold P adjusted , the calculation formula is: P adjusted =P base ×[1+β×(P risk -1)] Among them, P base Is the basic probability threshold, divided into the first level P base1 , Secondary P base2 and Level 3 P base3 , P base1 <P base2 <P base1 ;β is the scene sensitivity coefficient; P risk is the scenario risk index; S3512: Judge the comprehensive risk probability. When the comprehensive risk probability P risk When the threshold is less than the first level adaptive threshold, there is no warning; When the first-level adaptive threshold ≤ comprehensive risk probability P risk When the threshold is less than the second-level adaptive threshold, it is a first-level lightning warning; When the secondary adaptive threshold ≤ comprehensive risk probability P risk When the threshold is less than the third level adaptive threshold, it is the second level lightning warning; When the third-level adaptive threshold ≤ comprehensive risk probability P risk It is a level 3 thunderstorm warning.

8. The multi-modal collaborative local intelligent lightning graded warning method according to claim 1, characterized in that: Also includes: S4: Continuously monitor lightning activity, radar cloud data, and atmospheric electric field data in the local lightning geographic area, set a multimodal lightning data threshold, and cancel the warning response when the multimodal lightning data in the local lightning geographic area falls below the multimodal lightning data threshold; S5: Update the attention mechanism weights of the multimodal dynamic fusion module based on backtracking data to achieve iterative optimization of multimodal collaborative local lightning graded warning.

9. A multi-modal collaborative local lightning graded warning system, characterized by: It includes a data acquisition module, a data preprocessing module and a lightning warning model; the lightning warning model includes a lightning movement trajectory prediction module, a thundercloud movement direction acquisition module, a lightning occurrence probability calculation module, a cross-module dynamic fusion processing module and a graded warning module, which is used to implement a multi-modal collaborative local intelligent lightning graded warning method as described in any one of claims 1 to 8.

10. A storage medium, characterized in that: A computer program is stored thereon, and when the computer program is executed, it implements the multi-modal collaborative local intelligent lightning graded warning method according to any one of claims 1 to 8.

Images