A method and system for predicting lightning frequency and polarity

Abstract

This invention relates to the field of lightning prediction technology and discloses a method and system for predicting lightning frequency and polarity. The method involves acquiring lightning observation data and multi-source meteorological data within a target area, statistically classifying the lightning observation data to obtain lightning frequency (characterizing lightning intensity) and lightning polarity (characterizing charge distribution characteristics). The meteorological parameter sequences are input into a multi-channel feature extractor. Multiple parallel feature extraction channels in the multi-channel feature extractor extract the first feature vector from each meteorological parameter sequence. The lightning frequency probability vector is mapped to a frequency-guided feature vector. Based on dynamic weights, the first feature vector, the frequency-guided feature vector, and the polarity-guided feature vector are weighted and fused. The fused feature vector is input into a prediction model, outputting the lightning frequency probability distribution and lightning polarity probability distribution for the target area. This achieves high-precision, interpretable, collaborative prediction of lightning frequency and polarity.

Description

Technical Field

[0001] This invention relates to the field of lightning prediction technology, and in particular to a method and system for predicting lightning frequency and polarity. Background Technology

[0002] Lightning disasters are among the most severe meteorological disasters. Lightning frequency refers to the number of times lightning occurs within a specific time period, used to measure the severity and frequency of lightning occurrence. Lightning polarity refers to the positive or negative attribute of the charge released from the thundercloud to the ground during ground flash. Statistics on this are used to clarify the dominant type and intensity distribution of disasters. The proportion of positive and negative polarities in different regions reflects the differences in the charge structure of thunderstorms, providing basic data for regional lightning protection planning. Accurate prediction of lightning frequency and polarity is of great significance for power system operation, aerospace safety, and disaster prevention and mitigation.

[0003] Currently available lightning forecasting equipment has a single forecasting dimension and simply splices together data from multiple sources, resulting in coarse data fusion. Furthermore, it does not consider the inherent variation patterns of meteorological parameters during lightning, and the rationality and accuracy of the weighting of various meteorological parameters need to be improved. The data fusion effect and forecast accuracy are easily affected. Summary of the Invention

[0004] To overcome at least one of the shortcomings of the prior art described above, the present invention provides a method and system for predicting lightning frequency and polarity. This achieves high-precision, interpretable, coordinated prediction of lightning frequency and polarity.

[0005] The technical solution of this invention is implemented as follows: In a first aspect, the present invention provides a method for predicting lightning frequency and polarity, comprising: acquiring lightning observation data and multi-source meteorological data within a target area; statistically classifying the lightning observation data to obtain lightning frequency characterizing lightning intensity and lightning polarity characterizing charge distribution characteristics; mapping the lightning frequency to a lightning frequency probability vector and the lightning polarity to a lightning polarity probability vector; preprocessing the multi-source meteorological data into multiple meteorological parameter sequences; inputting the meteorological parameter sequences into a multi-channel feature extractor; extracting a first feature vector from each meteorological parameter sequence through multiple parallel feature extraction channels in the multi-channel feature extractor; and mapping the lightning frequency probability vector to a frequency-polarity probability vector. A secondary guiding feature vector is used to map the lightning polarity probability vector to a polarity guiding feature vector. The meteorological parameter sequence includes radar data sequence, satellite data sequence, and reanalysis field data sequence. A spatiotemporal dynamic fusion algorithm is used to calculate the dynamic weights of the first feature vector, frequency guiding feature vector, and polarity guiding feature vector. Based on these dynamic weights, the first feature vector, frequency guiding feature vector, and polarity guiding feature vector are weighted and fused to generate a fused feature vector. This fused feature vector is input into the prediction model, which outputs the lightning frequency probability distribution and lightning polarity probability distribution for the target area. Based on the lightning frequency probability distribution and lightning polarity probability distribution, the lightning prediction result for the target area is determined.

[0006] Based on the above technical solutions, preferably, the acquisition of lightning observation data and multi-source meteorological data within the target area, the statistical and tiered processing of the lightning observation data to obtain lightning frequency characterizing lightning intensity and lightning polarity characterizing charge distribution characteristics, mapping the lightning frequency to a lightning frequency probability vector, mapping the lightning polarity to a lightning polarity probability vector, and preprocessing the multi-source meteorological data into multiple meteorological parameter sequences, including: dividing the target area into grids of a preset size; statistically analyzing lightning observation data within the grids per unit time, counting the number of lightning occurrences, and mapping the number of lightning occurrences to a lightning frequency probability vector according to a preset lightning frequency level interval; statistically analyzing the number of positive polarity lightning occurrences within the grids, calculating the proportion of positive polarity lightning, and mapping the proportion of positive polarity lightning to a lightning polarity probability vector according to a preset lightning polarity frequency proportion interval.

[0007] Based on the above technical solution, preferably, the meteorological parameter sequence is input into a multi-channel feature extractor. Through multiple parallel feature extraction channels in the multi-channel feature extractor, the first feature vector in each meteorological parameter sequence is extracted. The lightning frequency probability vector is mapped to a frequency-guided feature vector, and the lightning polarity probability vector is mapped to a polarity-guided feature vector. The meteorological parameter sequence includes radar data sequence, satellite data sequence, and reanalysis field data sequence. The multiple feature extraction channels include a field strength evolution feature extractor, a field rheology feature extractor, and a field self-interaction feature extractor. The process includes: inputting the meteorological parameter sequence into a field strength evolution feature extractor to extract the physical field strength distribution and evolution dependence features, generating a field strength evolution feature vector; inputting the meteorological parameter sequence into a field rheological feature extractor to calculate the flow rate sequence and extract the fluctuation features, trend features, and extreme value features of the flow rate sequence, generating a field rheological feature vector; inputting the meteorological parameter sequence into a self-interaction feature extractor to calculate the autocorrelation features and frequency domain power spectrum features, generating a field self-interaction feature vector; and concatenating the field strength evolution feature vector, the field rheological feature vector, and the field self-interaction feature vector to obtain a first feature vector.

[0008] Based on the above technical solutions, preferably, the step of inputting the meteorological parameter sequence into a field strength evolution feature extractor to extract the physical field strength distribution and evolution dependency features, and generating a field strength evolution feature vector, includes: performing parallel convolution processing on the meteorological parameter sequence using multiple convolution kernels of different widths to extract the physical field strength distribution features; inputting the meteorological parameter sequence into a bidirectional long short-term memory network to extract the dependency features of the meteorological parameter sequence; and fusing the physical field strength distribution features and the dependency features to generate a field strength evolution feature vector.

[0009] Based on the above technical solutions, preferably, the meteorological parameter sequence is input into a field rheological feature extractor to calculate the flow rate sequence, and the fluctuation characteristics, trend characteristics, and extreme value characteristics of the flow rate sequence are extracted to generate a field rheological feature vector. This includes: performing first-order difference calculation on the meteorological parameter sequence to generate the flow rate sequence; performing statistical analysis on the flow rate sequence through multiple sliding windows of different dimensions to extract the fluctuation intensity characteristics and evolution trend characteristics of the flow rate sequence; extracting the extreme value characteristics of the flow rate sequence; and fusing the fluctuation intensity characteristics, evolution trend characteristics, and extreme value characteristics to generate a field rheological feature vector.

[0010] Based on the above technical solutions, preferably, the meteorological parameter sequence is input into a self-interaction feature extractor to calculate autocorrelation features and frequency domain power spectrum features, generating a field self-interaction feature vector. This includes: calculating the autocorrelation coefficient between the meteorological parameter sequence and its own lag sequence; performing a fast Fourier transform on the meteorological parameter sequence to calculate its power spectral density, and extracting the peak frequency and power value from the power spectral density; and weightedly fusing the autocorrelation coefficient, peak frequency, and power value to generate a field self-interaction feature vector.

[0011] Based on the above technical solutions, preferably, the dynamic weights of the first feature vector, frequency-guided feature vector, and polarity-guided feature vector are calculated using a spatiotemporal dynamic fusion algorithm. Based on these dynamic weights, the first feature vector, frequency-guided feature vector, and polarity-guided feature vector are weighted and fused to generate a fused feature vector. This fused feature vector is then input into a prediction model to output the lightning frequency probability distribution and lightning polarity probability distribution of the target region. The process includes: pre-setting initial static weights for the first feature vector, frequency-guided feature vector, and polarity-guided feature vector; normalizing these initial static weights to obtain basic weights; extracting spatial gradient features, spatial heterogeneity features, and center anomaly features from the first feature vector, frequency-guided feature vector, and polarity-guided feature vector; inputting these features into a Sigmoid activation function for mapping to obtain a spatial attention score; and extracting spatial gradient features, spatial heterogeneity features, and center anomaly features from the first feature vector, frequency-guided feature vector, and polarity-guided feature vector. Evolutionary trend features, volatility features, and mutation features are extracted and input into a Sigmoid activation function for mapping to obtain a temporal attention score. The basic weights, spatial attention scores, and temporal attention scores are weighted and fused, and the fusion result is normalized to obtain dynamic weights. Based on the calculated dynamic weights of the first feature vector, frequency-guided feature vector, and polarity-guided feature vector, the first feature vector, frequency-guided feature vector, and polarity-guided feature vector are weighted respectively. The weighted first feature vector, frequency-guided feature vector, and polarity-guided feature vector are summed to generate the fused feature vector. The fused feature vector is input into the prediction model, which includes a parallel frequency prediction branch and a polarity prediction branch. The frequency prediction branch processes the input and outputs a lightning frequency probability distribution via a Softmax activation function. The polarity prediction branch processes the input and outputs a lightning polarity probability distribution via a Sigmoid activation function.

[0012] Secondly, the present invention also discloses a lightning frequency and polarity prediction system, employing the lightning frequency and polarity prediction method described above. The system includes a data processing module, a feature extraction module, a cross-parameter correlation module, a weight calculation module, and a fusion prediction module, wherein... The data processing module is used to acquire lightning observation data and multi-source meteorological data within the target area, perform statistical classification processing on the lightning observation data to generate lightning frequency probability vector and lightning polarity probability vector; and preprocess the multi-source meteorological data to generate a meteorological parameter sequence set. The feature extraction module includes multiple feature extraction channels for generating a first feature vector, and for mapping the lightning frequency probability vector to a frequency-guided feature vector and the lightning polarity probability vector to a polarity-guided feature vector. The weight calculation module is used to calculate the dynamic weights of the first feature vector, the frequency-guided feature vector, and the polarity-guided feature vector; The fusion prediction module is used to perform weighted fusion of the first feature vector, frequency-guided feature vector and polarity-guided feature vector based on all dynamic weights to generate a fusion feature vector, and convert the fusion feature vector into the lightning frequency probability distribution and lightning polarity probability distribution of the target area.

[0013] Thirdly, the present invention also discloses an electronic device, including a processor and a memory; the memory stores a computer program, wherein the computer program, when executed by the processor, implements the above-described method for predicting the frequency and polarity of lightning.

[0014] Fourthly, the present invention also discloses a non-temporary computer-readable storage medium storing instructions that, when executed, perform the above-described method for predicting the frequency and polarity of lightning.

[0015] In summary, the lightning frequency and polarity prediction method and system provided by this invention have the following advantages over existing technologies: (1) The prediction tasks of lightning frequency and lightning polarity are placed in the same model for collaborative solution. By extracting multi-source meteorological features and introducing guiding feature vectors and independent prediction branches respectively, the complex patterns of two lightning activities can be learned simultaneously, improving the prediction accuracy. The spatiotemporal dynamic fusion algorithm calculates dynamic weights for each input sample in real time. These weights integrate basic weights, spatial attention weights, and temporal attention weights, enabling the model to intelligently adjust the degree of trust in different data sources and features according to the current specific meteorological conditions. This achieves the upgrade from a static model to a dynamic agent, significantly improving the prediction robustness and generalization ability under complex and changeable weather conditions. (2) By gridding lightning observation data and mapping it to lightning frequency probability vectors and lightning polarity probability vectors, discrete lightning events are transformed into standardized monitoring signals. This provides the model with a clear quantitative learning objective and allows the prediction results to be directly interpreted as the probability of occurrence of each level, greatly facilitating business integration and risk quantification assessment.

[0016] (3) By using three parallel channels for field strength evolution feature extraction, field rheology feature extraction and field self-interaction feature extraction, a first feature vector with clear physical meaning is constructed from the original meteorological parameter sequence. This makes the model learning process deeply aligned with the atmospheric physical process, greatly enhancing interpretability and operational credibility. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating a method for predicting lightning frequency and polarity according to an embodiment of the present invention; Figure 2 A structural block diagram of a spatiotemporal network prediction system for lightning frequency and polarity provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0020] See Figure 1 As shown, in a first aspect, the present invention provides a method for predicting the frequency and polarity of lightning, comprising: Acquire lightning observation data and multi-source meteorological data within the target area, statistically classify the lightning observation data to obtain lightning frequency, which characterizes the intensity of lightning occurrence, and lightning polarity, which characterizes the charge distribution characteristics. Map the lightning frequency to a lightning frequency probability vector and the lightning polarity to a lightning polarity probability vector. Preprocess the multi-source meteorological data and split it into multiple meteorological parameter sequences. The lightning observation data comes from lightning data detected by the ground lightning positioning network. For each lightning data point, the occurrence time t is specifically included. flash The latitude and longitude coordinates of the location where it occurred ( λ), lightning current amplitude I, and lightning polarity pol, where the occurrence time t flash Using Coordinated Universal Time (UTC), the polarity of lightning is either positive or negative.

[0021] The target area, such as a wind farm, is spatially divided into regular grids, for example, with a grid size of 1km × 1km. Each regular grid is assigned a unique identifier G. m，n This is represented as follows, where m is the row index and n is the column index. In terms of time, a spatiotemporal unit is formed using a preset duration, such as 1 hour. Statistical and hierarchical processing is then performed on the lightning observation data for each spatiotemporal unit, i.e., each regular grid, within that unit of time.

[0022] The lightning observation data is statistically categorized to obtain lightning frequency (characterizing lightning intensity) and lightning polarity (characterizing charge distribution). Lightning frequency is mapped to a lightning frequency probability vector, and lightning polarity is mapped to a lightning polarity probability vector. The multi-source meteorological data is preprocessed and split into multiple meteorological parameter sequences, including: dividing the target area into grids of preset size; statistically analyzing lightning observation data within each grid per unit time, counting the number of lightning occurrences, and mapping the number of occurrences to a lightning frequency probability vector according to a preset lightning frequency level interval; and statistically analyzing the number of positive polarity lightning occurrences within each grid, calculating the proportion of positive polarity lightning, and mapping the proportion of positive polarity lightning to a lightning polarity probability vector according to a preset lightning polarity frequency proportion interval.

[0023] Specifically, the lightning frequency grading method involves counting the total number of ground flashes occurring within the spatiotemporal unit, and mapping the total number of ground flashes to a discrete frequency gradation based on a preset lightning frequency level range. This application sets a total of 6 frequency gradations: 0-1 ground flashes for gradation 1, 2-3 ground flashes for gradation 2, 3-4 ground flashes for gradation 3, 4-5 ground flashes for gradation 4, 5-6 ground flashes for gradation 5, and more than 6 ground flashes for gradation 6. The lightning frequency is then mapped to a lightning frequency probability vector Y. freq Lightning frequency probability vector Y freq It is a K-dimensional vector, where K equals 6 in this application, and the lightning frequency probability vector Y freq =[y1, y2, y3, ..., y k ], where y k This represents the deterministic probability that the lightning frequency of this spatiotemporal unit belongs to the k-th level, where k is 1, 2, 3, 4, 5, or 6 and satisfies... During the training phase, for data with true labels, this vector is a one-hot vector, meaning that the element corresponding to the true gear is 1, and the rest are 0. For example, if the true gear is 3, then y3=1, and y1, y2, y4, y5 and y6 are all 0, providing a clear supervision signal for the model. During the prediction phase, this vector is the probability distribution of the model output.

[0024] The method for classifying lightning polarity involves counting the number of positive polarity lightning occurrences within a given spatiotemporal unit and calculating the proportion of positive polarity lightning to the total number of lightning occurrences. Based on a preset range of lightning polarity proportions, this proportion is mapped to a polarity level. This application sets two polarity levels: a positive polarity proportion ≥ 50% is encoded as positive polarity with a level value of 1; otherwise, it is encoded as negative polarity with a level value of 2. These polarity levels are then mapped to a lightning polarity probability vector Y. polar The lightning polarity probability vector is a two-dimensional vector, Y. polar =[p pos p neg ], where p pos +p neg Let p represent the probabilities of belonging to the positive and negative polarities, respectively, and satisfy p. pos +p neg =1.

[0025] The multi-source meteorological data is preprocessed to generate a sequence of meteorological parameters that can be directly input into the model. The multi-source meteorological data includes radar data, satellite data, and reanalysis field data. It should be noted that radar data includes combined reflectivity (CR), echo top height (ET), and vertical liquid water content (VIL); satellite data includes cloud top height (CTH) and cloud top temperature (CTT); and reanalysis field data includes temperature (T), relative humidity (RH), east-west wind vector (u), and north-south wind vector (v) at the three standard pressure layers of 500 hPa, 700 hPa, and 850 hPa. The combined reflectivity, measured in dBZ, represents the maximum reflectivity at different heights within the radar detection range, directly reflecting the concentration and size of water-forming substances such as raindrops or ice crystals within the cloud, and is a core parameter for determining convection intensity. The echo top height (ET), measured in km, represents the cloud top height where the radar reflectivity is ≥18 dBZ, reflecting the vertical development height of thunderstorm clouds; the higher the echo top height, the more vigorous the thunderstorm development. The vertical liquid water content (VIL), measured in kg / m³, is... 2The cloud top height (CTH) represents the total amount of liquid water within a vertical column per unit area, reflecting the degree of accumulation of water-forming substances within the cloud and serving as the material basis for the accumulation of charge during lightning activity. The cloud top height (CTH), measured in meters (m), represents the actual altitude of the cloud top, complementing the radar echo top height and reflecting the vertical development scale of thunderstorm clouds. The cloud top temperature (CTT), measured in Kelvin (K), represents the blackbody brightness temperature of the cloud top; lower temperatures indicate higher cloud tops and stronger convection, making it an important thermodynamic indicator for judging thunderstorm intensity. The reanalysis field data includes the temperature (T) at three standard pressure layers (500 hPa, 700 hPa, and 850 hPa), relative humidity (RH), east-west wind component (u), and north-south wind vector (v). Temperature (T), measured in Kelvin, reflects the thermodynamic state of the atmosphere; a large vertical temperature lapse rate favors convection development. Relative humidity (RH) reflects the atmospheric water vapor content; high humidity is a necessary condition for thunderstorm formation. The east-west wind component (u) and north-south wind component (v) can be combined to determine wind speed and direction, reflecting the dynamic environment of the atmosphere. The data is centered on the lightning occurrence time (t) within each spatiotemporal unit. flash Taking a reference time T0, we trace back N consecutive historical times, for example, N=8, denoted as T. -1 T -2 , ..., T -N This creates a unified timeline covering N+1 time points. For radar and satellite data, the most relevant spatial point is matched for each target location, and the time points (T) of that spatial point on the unified timeline are extracted. -1 T -2 , ..., T -N For reanalysis field data, the corresponding regular grid and data for the specified pressure layer at the corresponding time are extracted. After spatiotemporal alignment, the data are split according to independent physical observation variables. Each independent physical quantity, such as combined reflectance CR and 850 hPa temperature T, forms a separate meteorological parameter sequence. Meteorological parameter sequence X i The definition is as follows:

[0026]

[0027] Let represent the observed value of the i-th meteorological parameter at a relative time τ, where τ=0 corresponds to the reference time T0, and τ=-1, -2, ..., -N correspond to the historical time T0, respectively. -1 T -2 , ..., T -N N+1 is the length of the sequence, (.) T This indicates transpose, where the sequence is represented as a column vector.

[0028] The set of all M meteorological parameter sequences constitutes the input feature of a spatiotemporal unit, called the meteorological parameter sequence set, denoted as X = {X1, X2, ..., X...}.M For example, if the lightning combined reflectivity CR is selected as the first parameter, i.e., i=1, the observed values ​​at the reference time T0 and the eight time intervals prior to it are CR0, CR... -1 , ..., CR -8 Then its corresponding meteorological parameter sequence is: X CR =[CR0, CR -1 CR -2 , …, CR -8 ] T This fully depicts the dynamic evolution of reflectivity intensity during the critical period of thunderstorm development. Through the above preprocessing steps, the original multi-source, multi-dimensional meteorological data is effectively decomposed into multiple meteorological parameter sequences. Each meteorological parameter sequence is a univariate time series, which together constitute the standardized and structured input for subsequent multi-channel feature extraction of the model.

[0029] The meteorological parameter sequence is input into a multi-channel feature extractor. Through multiple parallel feature extraction channels in the multi-channel feature extractor, the first feature vector in each meteorological parameter sequence is extracted. The lightning frequency probability vector is mapped to a frequency-guided feature vector, and the lightning polarity probability vector is mapped to a polarity-guided feature vector. The meteorological parameter sequence includes radar data sequence, satellite data sequence, and reanalysis field data sequence.

[0030] The multiple feature extraction channels include a field strength evolution feature extractor, a field rheology feature extractor, and a field self-interaction feature extractor; the extraction process includes: inputting the meteorological parameter sequence into the field strength evolution feature extractor to extract the physical field strength distribution and evolution dependence features, generating a field strength evolution feature vector; inputting the meteorological parameter sequence into the field rheology feature extractor to calculate the flow rate sequence and extract the fluctuation features, trend features, and extreme value features of the flow rate sequence, generating a field rheology feature vector; inputting the meteorological parameter sequence into the self-interaction feature extractor to calculate the autocorrelation features and frequency domain power spectrum features, generating a field self-interaction feature vector; and concatenating the field strength evolution feature vector, the field rheology feature vector, and the field self-interaction feature vector to obtain a first feature vector.

[0031] Specifically, for each meteorological parameter sequence X i The process involves three parallel field strength evolution feature extractors, a field rheology feature extractor, and a field self-interaction feature extractor. First, the field strength fireworks feature extractor extracts the spatiotemporal distribution and overall evolution dependence features of the physical field strength.

[0032] The meteorological parameter sequence is input into a field strength evolution feature extractor to extract the physical field strength distribution and evolution dependency features, generating a field strength evolution feature vector. This includes: performing parallel convolution processing on the meteorological parameter sequence using multiple convolution kernels of different widths to extract physical field strength distribution features; inputting the meteorological parameter sequence into a bidirectional long short-term memory network to extract the dependency features of the meteorological parameter sequence; and fusing the physical field strength distribution features and the dependency features to generate a field strength evolution feature vector.

[0033] Specifically, the meteorological parameter sequence is processed in parallel using multiple convolution kernels of different widths to extract the physical field intensity distribution features. Specifically, J one-dimensional convolution kernels of different widths are used to process the meteorological parameter sequence X. i Parallel convolution is performed, where J can be 3, and the width of the one-dimensional convolution kernel can be 3, 5, or 7. The ReLU activation function is used to introduce nonlinearity to extract the physical field intensity distribution features. =ReLU(Conv1D(X i K j ), j=1,2,...,J. Where, K j This represents the lower j convolution kernels. This represents the corresponding physical field intensity distribution characteristics.

[0034] Specifically, the meteorological parameter sequence is input into a bidirectional long short-term memory network to extract the dependency features of the meteorological parameter sequence. i The input is fed into a bidirectional long short-term memory network to capture the long-term forward and backward dependencies of the meteorological parameter sequence. The bidirectional hidden state at the last time step is taken as the output: h N =Bi-LSTM(X i ) last Among them, Bi-LSTM (.) is a bidirectional long short-term memory network operation, which is a specific recurrent neural network structure. Bidirectional means that the network includes both a forward LSTM layer and a backward LSTM layer. This operation is applied to the meteorological parameter sequence X. i Encode the complex dependencies between the state at any given time in a meteorological parameter sequence and its entire historical and future states. last This represents the output state at the last time step, corresponding to the reference time T0, i.e., the output state when τ = 0; h NTo output the final context feature vector, this is the hidden state vector output by the Bi-LSTM network at the last time step, τ = 0. A bidirectional long short-term memory network is used to deeply encode the single-parameter time series and extract its final context fusion state as a feature. This step simulates how forecasters, when analyzing meteorological element curves, not only look at the current values ​​but also consider the entire historical evolution to understand the current stage and future trends. Global average pooling is applied to the physical field intensity distribution feature output by the j-th convolutional kernel, converting the physical field intensity distribution feature into a fixed-length global pooling vector. Global average pooling formula =GlobalAvgPool( ), and combine all pooling vectors with the hidden state h of the Bi-LSTM N The data is concatenated and then linearly transformed and dimensionality reduced through a fully connected layer to generate a field strength evolution feature vector. ;in, ,in, This is the weight matrix. For bias vectors, The dimension is d s The meteorological parameter sequence is transformed into a field strength evolution feature vector through a field strength evolution feature extractor. This provides input for subsequent processing.

[0035] The meteorological parameter sequence is input into a field strength evolution feature extractor to extract the physical field strength distribution and evolution dependency features, generating a field strength evolution feature vector. This includes: performing parallel convolution processing on the meteorological parameter sequence using multiple convolution kernels of different widths to extract physical field strength distribution features; inputting the meteorological parameter sequence into a bidirectional long short-term memory network to extract dependency features; and fusing the physical field strength distribution features and dependency features to generate a field strength evolution feature vector. The field strength evolution feature extractor is used to extract the rate of change of the physical field, including fluctuations, trends, and extreme values. First-order difference calculations are performed on the meteorological parameter sequence to generate a flow rate sequence representing the instantaneous rate of change. ; +1. This means that for the i-th meteorological parameter sequence, the difference between its adjacent time intervals is calculated to obtain a flow rate sequence of length N-1. Use a shorter sliding window, for example, with a length of w. s =3, calculate The moving standard deviation is used to obtain the average fluctuation intensity. ;in, ,in It is the standard deviation of the t-th shorter sliding window;

[0036] in, ; = It is the mean of a shorter sliding window; T s =Lw s +1 represents the total number of times the sliding window is shorter, and L is the flow rate sequence. The length of , where L=N.

[0037] Use a longer sliding window, for example, with a length of w. l =5, and w l >w s ,calculate The moving average is used to obtain the long-term moving average. ;in, , in, It is the mean of the t-th longer sliding window; T l =Lw l +1 represents the total number of times the longer sliding window is used.

[0038] calculate The maximum absolute value diff max and minimum value diff min ,in diff max This represents the maximum value of the entire difference sequence. The minimum value represents the entire difference sequence. The maximum and minimum values ​​together characterize the extreme range, which is particularly important for identifying thunderstorm outbreaks.

[0039] The above statistical features are concatenated and then passed through a fully connected layer to generate a field rheological feature vector. ;in W f Let b be the weight matrix. f For bias vectors, The dimension is d f .

[0040] The meteorological parameter sequence is input into a self-interaction feature extractor to calculate autocorrelation features and frequency domain power spectrum features, generating a field self-interaction feature vector. This includes: calculating the autocorrelation coefficient between the meteorological parameter sequence and its own lagged sequence; performing a fast Fourier transform on the meteorological parameter sequence to calculate its power spectral density, extracting the peak frequency and power value from the power spectral density; and weighted fusion of the autocorrelation coefficient, peak frequency, and power value to generate the field self-interaction feature vector. The meteorological parameter sequence X is calculated. i Autocorrelation coefficients at multiple lag orders l Where the lag order l = 1, 2, ..., L, where L = 5,

[0041]

[0042] in, yes The sequence mean.

[0043] For meteorological parameter series X i Perform a Fast Fourier Transform to obtain its frequency domain representation F. i =FFT(X) i ), calculate the power spectral density P i =|F i | 2 And extract the first K peak frequencies. and corresponding power values Where k = 1, 2, ..., K. The autocorrelation coefficients... By concatenating the features from the frequency domain with a fully connected layer, a field self-interaction feature vector is generated. ;

[0044] ; Among them, W c Let b be the weight matrix. c For bias vectors, The dimension is d c .

[0045] The field strength evolution feature vector, field rheology feature vector, and field self-interaction feature vector are concatenated to form the first feature vector H of the i-th meteorological parameter sequence. i ; Among them, H i =[ ], H i Total dimension d h =d s +d f +d c It integrates comprehensive information on the intensity evolution, rate of change, and periodicity of the meteorological parameter sequence.

[0046] The frequency-guided feature vector is mapped by a fully connected layer to the lightning frequency probability vector Y. freq Mapped to H i Feature vectors of the same dimension; Frequency-guided feature vector H freq =W emb_freq ×Y freq +b emb_freq Among them, W emb_freq Let b be the weight matrix. emb_freq This is the bias vector.

[0047] The polarity-guided feature vector is mapped through another fully connected layer, which is the Jiang Lei polarity probability vector Y. polar Mapped to feature vectors; Polarity-guided eigenvector H polar =W emb_polar× Y polar +b emb_polar Among them, W emb_polar Let b be the weight matrix. emb_polar This is the bias vector.

[0048] Among them, the frequency-guided feature vector H freq Frequency-guided feature vector H freq As global guiding information, it will participate in the fusion decision together with the first feature vector of each meteorological parameter sequence in subsequent steps.

[0049] The dynamic weights of the first feature vector, frequency-guided feature vector, and polarity-guided feature vector are calculated using a spatiotemporal dynamic fusion algorithm. Based on these dynamic weights, the first feature vector, frequency-guided feature vector, and polarity-guided feature vector are weighted and fused to generate a fused feature vector. This fused feature vector is then input into a prediction model to output the lightning frequency probability distribution and lightning polarity probability distribution of the target region. The process includes: pre-setting initial static weights for the first feature vector, frequency-guided feature vector, and polarity-guided feature vector; normalizing these initial static weights to obtain basic weights; extracting spatial gradient features, spatial heterogeneity features, and center anomaly features from the first feature vector, frequency-guided feature vector, and polarity-guided feature vector, and inputting them into a Sigmoid activation function for mapping to obtain spatial attention scores; and extracting evolution trends from the first feature vector, frequency-guided feature vector, and polarity-guided feature vector. Features, volatility features, and mutation features are input into a Sigmoid activation function for mapping to obtain a temporal attention score. The base weights, spatial attention scores, and temporal attention scores are weighted and fused, and the fusion result is normalized to obtain dynamic weights. Based on the calculated dynamic weights of the first feature vector, frequency-guided feature vector, and polarity-guided feature vector, the first feature vector, frequency-guided feature vector, and polarity-guided feature vector are weighted respectively. The weighted first feature vector, frequency-guided feature vector, and polarity-guided feature vector are summed to generate the fused feature vector. The fused feature vector is input into the prediction model, which includes a parallel frequency prediction branch and a polarity prediction branch. The frequency prediction branch processes the input and outputs a lightning frequency probability distribution via a Softmax activation function. The polarity prediction branch processes the input and outputs a lightning polarity probability distribution via a Sigmoid activation function.

[0050] Specifically, the spatiotemporal dynamic fusion algorithm is used to calculate the dynamic weights for each first feature vector, frequency-guided feature vector, and polarity-guided feature vector, including the basic weights. Spatial attention score Temporal attention score Each meteorological parameter sequence i is pre-set with an initial static weight. This initial static weight is predetermined based on its physical meaning or typical importance; for example, features directly related to strong convection (such as features corresponding to combined reflectance) may be assigned a higher initial weight. Let there be a total of M meteorological parameter sequences, and their initial static weights be denoted as... The basic weights are obtained through normalization. ;

[0051] After normalization, it satisfies This ensures that the weights sum to 1, facilitating subsequent fusion calculations.

[0052] Spatial attention is used to evaluate the importance of the first eigenvector, frequency-guided eigenvector, and polarity-guided eigenvector in the spatial environment surrounding the current target grid point. For each meteorological parameter sequence i, the spatial feature matrix of the target grid point and its surrounding K×K neighborhood at the current time T0 is extracted. Spatial gradient features are calculated using the Sobel operator. i The gradient magnitude is obtained by considering the gradient in the horizontal and vertical directions. And calculate its mean, mean =mean(G i ), where G i Let represent the gradient magnitude matrix of the i-th feature in the K×K neighborhood, calculated using the Sobel operator. mean() is the mean function, calculating the average of all elements in the matrix. The average gradient magnitude is a scalar value that reflects the degree of spatial variation of this feature. The spatial heterogeneity characteristic S is calculated as follows: i Standard deviation, Ω i =std(S i ), where S i This represents the spatial feature matrix of the i-th feature within a K×K neighborhood, including the eigenvalues ​​of the target point and its surrounding points. `std(.)` is the standard deviation function, used to calculate the standard deviation of all elements in the matrix. Ω i The spatial heterogeneity index is a scalar value that reflects the degree of non-uniformity in the spatial distribution of this feature; the central anomaly feature is calculated as the difference between the central point value and the neighborhood mean, Δ. i =x i (T0)-mean(S i This reflects the degree of anomaly of the target point relative to its surrounding environment. Where x... i (T0) represents the feature value of the i-th feature at the current time T0, mean(S i Let Δ be the mean of the i-th feature in the K×K neighborhood. i The central outlier is a scalar value that reflects the degree of anomaly of the target point relative to its surrounding environment. If Δ i A positive value indicates that it is higher than the surrounding area, Δ i A negative value indicates that the value is lower than the surrounding value.

[0053] The three features mentioned above are concatenated into a spatial feature vector. =[ Ω i Δ i Spatial attention scores are calculated using a fully connected layer and a sigmoid activation function. ,in, =[ Ω i Δ i ] represents the spatial feature vector of the i-th feature, with dimensions 3×1. This is a spatial attention weight matrix, which you can refer to for learning purposes; its dimensions are 1×3. This is the spatial attention bias matrix, which can be studied and referenced; it is a scalar value. Represents matrix penalty, (.) is the sigmoid function, which compresses the output to the interval (0,1). Let be the spatial attention score of the i-th feature, ranging from (0,1). A larger value indicates that the feature is more important in the current spatial environment.

[0054] Temporal attention is used to evaluate the criticality of the evolution of the first feature vector, frequency-guided feature vector, and polarity-guided feature vector in the period leading up to lightning occurrence for prediction. The input is the meteorological feature parameter X corresponding to each meteorological parameter sequence i. i It typically includes characteristic values ​​of T0 and several times prior, and the evolution trend characteristics affect the meteorological characteristic parameter X. i Perform linear fitting to obtain the slope s i This reflects the long-term changing trend of the characteristics; the volatility characteristic is calculated as X. i Standard deviation , =std(X i ), standard deviation Reflecting the degree of fluctuation; calculating meteorological characteristic parameter X based on abrupt change characteristics. i First-order difference Calculate the proportion of those whose absolute value is greater than a preset threshold β. ;

[0055] in, For X i The vector is a first-order difference vector, where β is the mutation threshold used to determine whether the difference values ​​constitute a significant mutation, count() is a counting function that counts the number of elements that meet the conditions, and length() is a length function that returns the number of elements in the vector. The mutation rate is in the range [0,1], reflecting the frequency of significant mutations in this feature.

[0056] The three features mentioned above are concatenated into a temporal feature vector. Temporal attention scores are calculated using a fully connected layer and a sigmoid activation function. ,in The larger the value, the more critical the temporal evolution pattern of the feature is to the current prediction.

[0057] The base weights, spatial attention scores, and temporal attention scores are fused using learnable weight parameters to generate the final dynamic weights. ;

[0058] Where, 0≤ ≤1, and + ≤1, and To balance the contribution ratios of base weights, spatial attention, and temporal attention, learnable parameters are used. and The base weights, spatial attention, and temporal attention are combined proportionally to obtain unnormalized dynamic weights. ; For a temporal attention system, the range is also between [0,1], which controls the contribution ratio of the temporal attention score.

[0059] The Softmax function is used to normalize the dynamic weights, ensuring that the sum of all weights is 1, and the weights are dynamically fused. ; Here, exp(.) is the exponential function, ensuring that all inputs are positive; Sum the exponents of all unnormalized weights and use the sum as the normalized denominator. For dynamic fusion weights.

[0060] Based on the calculated dynamic weights, all feature vectors are weighted and summed to generate a fused feature vector. Let the i-th feature vector be h. i It includes a first feature vector, a frequency-guided feature vector, and a polarity-guided feature vector, with a dynamic weight of w. i Then the feature vectors are fused. ; where h i Let be the i-th eigenvector.

[0061] Fuse feature vector H fuse The input is fed into a prediction model, which includes two parallel branches: a frequency prediction branch and a polarity prediction branch. The frequency prediction branch integrates the fused feature vectors. After processing, the lightning frequency probability distribution P is output through the Softmax activation function. freq ∈R Cfreq C freq Number of lightning frequency levels; frequency prediction ; Among them, Z freq This is the original output vector of the frequency prediction distribution, with dimension C. freq ×1, Cfreq Let exp(.) be the number of lightning frequency levels, and let exp(.) be the exponential function. Let be the probability of the c-th level in the probability distribution of lightning frequency, with a range of (0,1).

[0062] Polarity prediction branch pairs fused feature vectors After processing, the polarity probability distribution P is output through the Sigmoid activation function. polar ∈R 2 Let P and denot be the probabilities of positive and negative polarity, respectively; the polarity prediction formula is P. polar [1]= (Z) polar ), P polar [2]= P polar [1], where Z polar The original output scalar for the polarity prediction branch. (.) represents the Sigmoid function, P polar [1] represents the probability of positive polarity lightning, ranging from (0,1), P polar [2] represents the probability of negative polarity lightning.

[0063] The decision is made based on the probability distribution of the output. In the lightning frequency prediction, the category with the highest probability is selected as the predicted lightning frequency level. The prediction frequency = argmax(P) freq Predicted polarity = argmax(P) polar The algorithm, argmax(), is a maximum index function that returns the category index corresponding to the maximum probability value in the probability vector. The output is an integer category label, representing the predicted lightning frequency level and lightning polarity, respectively. This invention introduces spatial and temporal attention mechanisms, enabling the system to dynamically adjust the weights of each feature based on the spatial anomaly degree and temporal evolution pattern of the current sample. This achieves scenario-adaptive fusion prediction, significantly improving the accuracy and interpretability of lightning frequency and polarity prediction.

[0064] Based on the lightning frequency probability distribution and the lightning polarity probability distribution, the lightning prediction result for the target area is determined.

[0065] This invention achieves a lightning frequency prediction accuracy of 87.6% and a polarity prediction accuracy of 88.1% on the test set, a significant improvement over the traditional fixed-weight fusion method's 75%. Furthermore, through dynamic attention weight visualization, forecasters can clearly understand which meteorological parameters each prediction is based on, greatly enhancing operational trust. From feature extraction to fusion, the model structure design is inspired by physical processes, making its predictions more consistent with atmospheric science laws and exhibiting stronger generalization ability for rare or extreme events. The dynamic fusion mechanism allows the model to flexibly respond to different types of convective weather without requiring different models for different weather conditions. Moreover, the co-evolutionary reinforcement learning strategy can more efficiently search for high-performance, low-complexity model parameter combinations, reducing the burden of manual parameter tuning.

[0066] Secondly, please refer to Figure 2 As shown, the present invention also provides a lightning frequency and polarity spatiotemporal network prediction system, employing the lightning frequency and polarity spatiotemporal network prediction method described above. The system includes a data processing module 101, a feature extraction module 102, a weight calculation module 103, and a fusion prediction module 104, wherein... The data processing module 101 is used to acquire lightning observation data and multi-source meteorological data within the target area, perform statistical classification processing on the lightning observation data to generate lightning frequency probability vector and lightning polarity probability vector; and preprocess the multi-source meteorological data to generate a meteorological parameter sequence set. The feature extraction module 102 includes multiple feature extraction channels for generating a first feature vector, and for mapping the lightning frequency probability vector to a frequency-guided feature vector and the lightning polarity probability vector to a polarity-guided feature vector. The weight calculation module 103 is used to calculate the dynamic weights of the first feature vector, the frequency-guided feature vector, and the polarity-guided feature vector; The fusion prediction module 104 is used to perform weighted fusion of the first feature vector, frequency-guided feature vector and polarity-guided feature vector based on all dynamic weights to generate a fusion feature vector, and convert the fusion feature vector into the lightning frequency probability distribution and lightning polarity probability distribution of the target area.

[0067] It should be noted that this system corresponds to the aforementioned method for predicting the frequency and polarity of lightning. All implementation methods in the above method embodiments are applicable to the embodiments of this system and can achieve the same technical effect.

[0068] Thirdly, please refer to Figure 3 As shown, Figure 3This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device 200 provided in this application includes a processor 210 and a memory 220; the memory 220 stores a computer program, wherein the computer program, when executed by the processor, implements the aforementioned method for predicting the frequency and polarity of lightning through a spatiotemporal network.

[0069] Specifically, processor 210 may include, for example, a general-purpose microprocessor, an instruction set processor and / or an associated chipset and / or a special-purpose microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. Processor 210 may also include onboard memory for caching purposes. Processor 210 may be a single processing unit or multiple processing units for performing different actions of the method flow according to embodiments of this application.

[0070] Memory 220 may be any medium capable of containing, storing, transmitting, propagating, or transmitting instructions. For example, memory 220 may include, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, instruments, or propagation media. Specific examples of memory 220 include: magnetic storage devices such as magnetic tape or hard disk drives (HDDs); optical storage devices such as optical discs (CD-ROMs); and may also be random access memory (RAM) or flash memory; and / or wired / wireless communication links.

[0071] This application also provides a computer-readable medium storing a computer program that, when executed by a processor, implements the above-described method for predicting the spatiotemporal network of lightning frequency and polarity. This computer-readable medium may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into that device / apparatus / system. The aforementioned computer-readable medium carries one or more programs, which, when executed, implement the method according to the embodiments of this application.

[0072] According to embodiments of this application, a computer-readable medium may be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this application, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media can also be any computer-readable medium other than computer-readable storage media, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wireless, wired, optical fiber, radio frequency signals, etc., or any suitable combination thereof.

[0073] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting the frequency and polarity of lightning, characterized in that, include: Acquire lightning observation data and multi-source meteorological data within the target area, statistically classify the lightning observation data to obtain lightning frequency, which characterizes the intensity of lightning occurrence, and lightning polarity, which characterizes the charge distribution characteristics. Map the lightning frequency to a lightning frequency probability vector and the lightning polarity to a lightning polarity probability vector. Preprocess the multi-source meteorological data and split it into multiple meteorological parameter sequences. The meteorological parameter sequence is input into a multi-channel feature extractor. Through multiple parallel feature extraction channels in the multi-channel feature extractor, the first feature vector in each meteorological parameter sequence is extracted. The lightning frequency probability vector is mapped to a frequency-guided feature vector, and the lightning polarity probability vector is mapped to a polarity-guided feature vector. The meteorological parameter sequence includes radar data sequence, satellite data sequence, and reanalysis field data sequence. The dynamic weights of the first feature vector, frequency-guided feature vector, and polarity-guided feature vector are calculated using a spatiotemporal dynamic fusion algorithm. Based on the dynamic weights, the first feature vector, frequency-guided feature vector, and polarity-guided feature vector are weighted and fused to generate a fused feature vector. The fused feature vector is then input into the prediction model to output the lightning frequency probability distribution and lightning polarity probability distribution of the target area. Based on the lightning frequency probability distribution and the lightning polarity probability distribution, the lightning prediction result for the target area is determined.

2. The method for predicting lightning frequency and polarity according to claim 1, characterized in that, The process involves acquiring lightning observation data and multi-source meteorological data within the target area, statistically classifying the lightning observation data to obtain lightning frequency (characterizing lightning intensity) and lightning polarity (characterizing charge distribution characteristics), mapping lightning frequency to a lightning frequency probability vector, and mapping lightning polarity to a lightning polarity probability vector. The multi-source meteorological data is then preprocessed and split into multiple meteorological parameter sequences, including: Divide the target area into a grid of preset size; The system collects lightning observation data from the grid within a unit of time, counts the number of lightning occurrences, and maps the number of lightning occurrences into a lightning frequency probability vector according to a preset lightning frequency level range. The number of positive polarity lightning occurrences within the grid is counted, the proportion of positive polarity lightning is calculated, and the proportion of positive polarity lightning is mapped to a lightning polarity probability vector according to the preset lightning polarity frequency ratio range.

3. The method for predicting lightning frequency and polarity according to claim 2, characterized in that, The meteorological parameter sequence is input into a multi-channel feature extractor. Through multiple parallel feature extraction channels in the multi-channel feature extractor, the first feature vector of each meteorological parameter sequence is extracted. The lightning frequency probability vector is mapped to a frequency-guided feature vector, and the lightning polarity probability vector is mapped to a polarity-guided feature vector. The meteorological parameter sequence includes radar data sequences, satellite data sequences, and reanalysis field data sequences. The multiple feature extraction channels include a field intensity evolution feature extractor, a field rheology feature extractor, and a field self-interaction feature extractor; including: The meteorological parameter sequence is input into the field strength evolution feature extractor to extract the physical field strength distribution and evolution dependence features, and generate the field strength evolution feature vector. The meteorological parameter sequence is input into the field rheological feature extractor to calculate the flow rate sequence, and the fluctuation characteristics, trend characteristics and extreme value characteristics of the flow rate sequence are extracted to generate the field rheological feature vector. The meteorological parameter sequence is input into the self-interaction feature extractor to calculate the autocorrelation feature and frequency domain power spectrum feature, and generate the field self-interaction feature vector. The first feature vector is obtained by concatenating the field strength evolution feature vector, the field rheology feature vector, and the field self-interaction feature vector.

4. The method for predicting lightning frequency and polarity according to claim 3, characterized in that, The step of inputting the meteorological parameter sequence into the field strength evolution feature extractor to extract the physical field strength distribution and evolution dependency features, and generating a field strength evolution feature vector, includes: The meteorological parameter sequence is processed by parallel convolution using multiple convolution kernels of different widths to extract the physical field intensity distribution features; The meteorological parameter sequence is input into a bidirectional long short-term memory network to extract the dependency features of the meteorological parameter sequence; The physical field intensity distribution characteristics and the dependency characteristics are fused to generate a field intensity evolution feature vector.

5. The method for predicting lightning frequency and polarity according to claim 3, characterized in that, The meteorological parameter sequence is input into the field rheological feature extractor to calculate the flow rate sequence, and the fluctuation characteristics, trend characteristics, and extreme value characteristics of the flow rate sequence are extracted to generate a field rheological feature vector, including: The meteorological parameter sequence is subjected to first-order difference calculation to generate a flow rate sequence; The flow rate sequence is statistically analyzed using sliding windows of multiple different dimensions to extract the characteristics of the fluctuation intensity and evolution trend of the flow rate sequence. The extreme value features of the flow rate sequence are extracted, and the fluctuation intensity features, evolution trend features and extreme value features are fused to generate a field rheological feature vector.

6. The method for predicting lightning frequency and polarity according to claim 3, characterized in that, The meteorological parameter sequence is input into a self-interaction feature extractor to calculate autocorrelation features and frequency domain power spectrum features, generating a field self-interaction feature vector, including: Calculate the autocorrelation coefficient between the meteorological parameter series and its own lagged series; Perform a fast Fourier transform on the meteorological parameter sequence to calculate its power spectral density, and extract the peak frequency and power value from the power spectral density; The autocorrelation coefficient, peak frequency, and power value are weighted and fused to generate a field self-interaction feature vector.

7. The method for predicting lightning frequency and polarity according to claim 1, characterized in that, The dynamic weights of the first feature vector, frequency-guided feature vector, and polarity-guided feature vector are calculated using a spatiotemporal dynamic fusion algorithm. Based on these dynamic weights, the first feature vector, frequency-guided feature vector, and polarity-guided feature vector are weighted and fused to generate a fused feature vector. This fused feature vector is then input into a prediction model, which outputs the lightning frequency probability distribution and lightning polarity probability distribution for the target area, including: Initial static weights are preset for the first feature vector, frequency-guided feature vector, and polarity-guided feature vector. The initial static weights of the first feature vector, frequency-guided feature vector, and polarity-guided feature vector are normalized to obtain the basic weights. Spatial gradient features, spatial heterogeneity features, and center anomaly features are extracted from the first feature vector, frequency-guided feature vector, and polarity-guided feature vector, and then input into the Sigmoid activation function for mapping to obtain the spatial attention score. Evolution trend features, volatility features, and mutation features are extracted from the first feature vector, frequency-guided feature vector, and polarity-guided feature vector, and then input into the Sigmoid activation function for mapping to obtain the temporal attention score. The basic weights, spatial attention scores, and temporal attention scores are weighted and fused, and the fusion result is normalized to obtain dynamic weights. Based on the calculated dynamic weights of the first feature vector, frequency-guided feature vector, and polarity-guided feature vector, the first feature vector, frequency-guided feature vector, and polarity-guided feature vector are weighted respectively. The weighted first feature vector, frequency-guided feature vector, and polarity-guided feature vector are summed to generate the fused feature vector; The fused feature vector is input into the prediction model, which includes a parallel frequency prediction branch and a polarity prediction branch. The frequency prediction branch processes the input and outputs a lightning frequency probability distribution via a Softmax activation function. The polarity prediction branch processes the input and outputs a lightning polarity probability distribution via a Sigmoid activation function.

8. A lightning frequency and polarity prediction system, employing the lightning frequency and polarity prediction method as described in any one of claims 1 to 7, characterized in that, The system includes a data processing module, a feature extraction module, a weight calculation module, and a fusion prediction module, wherein... The data processing module is used to acquire lightning observation data and multi-source meteorological data within the target area, perform statistical classification processing on the lightning observation data to generate lightning frequency probability vector and lightning polarity probability vector; and preprocess the multi-source meteorological data to generate a meteorological parameter sequence set. The feature extraction module includes multiple feature extraction channels for generating a first feature vector, and for mapping the lightning frequency probability vector to a frequency-guided feature vector and the lightning polarity probability vector to a polarity-guided feature vector. The weight calculation module is used to calculate the dynamic weights of the first feature vector, the frequency-guided feature vector, and the polarity-guided feature vector; The fusion prediction module is used to perform weighted fusion of the first feature vector, frequency-guided feature vector and polarity-guided feature vector based on all dynamic weights to generate a fusion feature vector, and convert the fusion feature vector into the lightning frequency probability distribution and lightning polarity probability distribution of the target area.

9. An electronic device, characterized in that, It includes a processor and a memory; the memory stores a computer program, wherein the computer program, when executed by the processor, implements the lightning frequency and polarity prediction method as described in any one of claims 1 to 7.

10. A non-temporary computer-readable storage medium, characterized in that, The non-temporary computer-readable storage medium stores instructions that, when executed, perform the lightning frequency and polarity prediction method as described in any one of claims 1 to 7.

Images