Lightning Prediction Method, System and Device Based on GCN and ConvLSTM

Abstract

This application relates to the field of meteorological prediction technology, and discloses a lightning prediction method, system and device based on GCN and ConvLSTM. The method includes: Step S1, using multi-source meteorological data as multi-source data input and performing preprocessing operations on it; Step S2, constructing an ST-GLNet prediction model based on GCN and ConvLSTM and training it, processing the input data and outputting the lightning occurrence probability to achieve lightning prediction; the input layer is provided with multiple input channels corresponding to input data of different dimensions, and each input channel is provided with a single encoder for extracting the spatio-temporal features of the input data; the GCN layer is used to extract the spatial features of the input data, and the ConvLSTM layer performs time series modeling on the output data of the GCN layer; the output layer outputs one-dimensional rasterized data representing the lightning occurrence probability. This application introduces GCN to extract the spatial features between grid points and neighbor nodes, combines ConvLSTM to model the time features of the data, and effectively solves the problems existing in traditional methods through the synergistic effect of ST-GLNet.

Description

Technical Field

[0001] This application relates to the field of meteorological prediction technology, and particularly relates to a lightning prediction method, system and device based on GCN and ConvLSTM. Background Art

[0002] At present, lightning prediction, as an important meteorological service, has been widely applied in industries such as aviation, energy, and communication. However, the existing prediction technologies still have significant limitations. The mainstream methods mainly rely on numerical weather prediction models and traditional statistical methods based on historical data, which predict the probability of lightning occurrence by analyzing the historical trends and patterns of meteorological data. Nevertheless, due to limitations in spatial resolution and temporal resolution, their prediction results often fail to meet the precise requirements of local lightning warnings.

[0003] The occurrence of lightning has strong local and non-linear characteristics, and it is difficult for traditional methods to fully capture these complex characteristics, resulting in limited prediction accuracy. For example, although the early warning model based on radar meteorological data can provide effective early warnings in some cases, due to the limitations of radar data (such as insufficient radar cloud information), the early warning success rate is relatively low. In addition, traditional numerical weather prediction models also face challenges in dealing with severe storm weather, because the strong interaction of different physical and microphysical processes at different scales makes accurate prediction difficult.

[0004] The rapid accumulation of meteorological observation data provides rich data support for lightning prediction, including data from automatic weather stations (AWS), radar observation data, and actual lightning occurrence records. However, how to effectively integrate these multi-source heterogeneous data remains a challenge. Traditional models have weak capabilities in extracting spatio-temporal features of weather station data and radar data, and it is difficult to utilize their inherent spatial topological relationships and temporal dynamic characteristics.

[0005] In recent years, deep learning technologies have shown significant advantages in the field of spatio-temporal data processing. In particular, models such as graph convolutional networks (GCN) and long short-term memory networks (LSTM) have great potential in dealing with complex spatio-temporal relationships. However, there are still certain limitations in the application of existing deep learning models in lightning prediction: on the one hand, many models only focus on a single data source and it is difficult to achieve joint modeling of multi-source data; on the other hand, the lack of temporal modeling capabilities and the inefficiency of spatial feature extraction make it impossible for the models to fully exploit the potential of meteorological data. For example, although LSTM has advantages in processing time series data, its efficiency in spatial feature extraction is relatively low, which limits its application effect in complex meteorological data. Summary of the Invention

[0006] The technical problem to be solved by this application is to overcome the deficiencies of the prior art and provide a lightning prediction method, system and device based on GCN and ConvLSTM, combining the advantages of graph convolution (GCN) in extracting data spatial topological features and modeling the spatial dependence between grid points and neighbor nodes, and the powerful ability of convolutional long short-term memory network (ConvLSTM) in capturing time dynamic characteristics and non-linear time series relationships. Through the synergistic effect of ST-GLNet, effectively solve the problems of insufficient spatial and temporal resolution, limited spatio-temporal feature extraction ability, insufficient multi-source data fusion and insufficient non-linear modeling ability in existing methods.

[0007] To achieve the above object, the first aspect of this application provides a lightning prediction method based on GCN and ConvLSTM, including the following steps:

[0008] Step S1, using multi-source meteorological data as multi-source data input and performing preprocessing operations on it. The multi-source meteorological data includes automatic weather station data, radar data, and lightning observation data. Among them, the automatic weather station data and radar data are used to extract the spatio-temporal features of lightning occurrence, and the lightning observation data is used as the true label for supervising model training.

[0009] Step S2, constructing and training an ST-GLNet prediction model based on GCN and ConvLSTM. The architecture of the ST-GLNet prediction model includes an input layer, an encoder, a fusion module, a decoder, and an output layer, processes the input data and outputs the lightning prediction result to achieve lightning prediction.

[0010] The input layer is provided with multiple input channels corresponding to input data of different dimensions, and each input channel is provided with a single encoder to extract the spatio-temporal features of the input data.

[0011] The encoder includes a GCN layer and a ConvLSTM layer. The GCN layer is used to extract the spatial features of the input data, including introducing spectral graph regularization, defining the energy propagation rule between nodes as the topological connection of horizontal and vertical adjacent units through a physically constrained adjacency matrix construction method, and combining self-loop connection and symmetric normalization processing; it also includes modeling spatio-temporal features at multiple levels by stacking the GCN layer and the ConvLSTM layer through a topological-aware spatio-temporal hierarchical structure; the ConvLSTM layer performs time series modeling on the output data of the GCN layer. Among them, the hidden state update mechanism of the ConvLSTM layer is given physical topological constraints, and the convolutional kernel of the ConvLSTM layer shares the same spatial relationship with the adjacency matrix; and

[0012] The fusion module fuses the outputs of the encoder in all input channels through a correlation-aware dynamic weight allocation mechanism and a competitive gradient learning strategy as the input to the decoder.

[0013] Optionally, the input layer includes three independent input channels. The first channel receives radar data with a dimension of , where , are the number of grid divisions in the longitude and latitude directions of the geographical grid data respectively, is the number of features of the radar data; the second channel receives single-channel rasterized lightning event data with a dimension of ; the third channel receives multi-channel rasterized data of automatic weather stations with a dimension of , where is the number of features of the rasterized data of the automatic weather station. The three input channels are processed by independent encoders respectively.

[0014] Optionally, the spectral graph regularization includes an operation of pre-building a fixed adjacency matrix based on the spatial adjacency relationship of grid cells. The pre-building means capturing the spatial dependence relationship between grid points by pre-building a fixed adjacency matrix based on the spatial adjacency relationship of grid cells, specifically including:

[0015] Each grid cell is mapped to a node in the graph, and the adjacency matrix is established through four-way connection , where is the number of grid cells, expressed as:

[0016] ;

[0017] Among them, represents the grid spacing, and are the coordinates of the corresponding grid nodes and node in the two-dimensional space;

[0018] After that, the adjacency matrix is symmetrically normalized by the node degree matrix to obtain the normalized adjacency matrix , where is the identity matrix to ensure that each node can be connected to itself;

[0019] The spectral graph regularization is further enhanced by Laplacian smoothing, expressed as:

[0020] ;

[0021] Among them, represents the node feature of the th layer, Denote the node features of the layer as the normalized adjacency matrix as the spectral graph convolution operation is the smoothing scale parameter denote the Frobenius norm, the auxiliary loss term penalize the node features of the layer and the smoothing result of the layer for the deviation.

[0022] Optionally, the encoder adopts cascaded GCN layers and ConvLSTM layers, including adopting the topological-aware spatio-temporal hierarchical structure, and modeling spatio-temporal features by stacking three GCN layers and ConvLSTM layers; wherein, the GCN layer extracts the spatial features of the input data as the input of the ConvLSTM layer, including receiving three types of inputs: radar, lightning observation and AWS data, and respectively extracting the spatio-temporal features of each data. Taking radar data as an example, it is expressed as:

[0023] ;

[0024] wherein, is the time step, is the number of grid nodes, is the feature dimension of radar data;

[0025] The GCN layer extracts the spatial features of grid nodes through two convolutional operations. In each layer of GCN, the feature information of adjacent nodes is aggregated, which is expressed as:

[0026] ;

[0027] wherein, denote the node features of the layer as is the activation function, is the normalized adjacency matrix, is the node degree matrix, is the output feature after the GCN operation of this layer.

[0028] Optionally, the hidden state and memory state of each time step are generated and updated through the ConvLSTM layer, including after obtaining the spatial features, inputting them into the ConvLSTM layer for temporal modeling, and the ConvLSTM layer updates the hidden state and memory state , expressed as:

[0029] ;

[0030] ;

[0031] ;

[0032] ;

[0033] ;

[0034] ;

[0035] Among them, is the activation function, is the input gate, is the forget gate, is the output gate, , , and are bias terms, represents the calculation process of the ConvLSTM cell, is the feature output from the last layer of GCN, and respectively represent the memory state and hidden state at the current time step, is the adjacency matrix explicitly constrains the local receptive field of the convolutional kernel through the Hadamard product. The convolutional kernel and 's sliding window pattern is aligned with the topological structure of the adjacency matrix . In the time dimension, it inherits the state persistence of the physical system through the memory state at the previous time step. In the spatial dimension, it inherits the topological dependence of the GCN layer through the operation.

[0036] Optionally, the fusion module reduces the dimensions of the hidden state and memory state output by each encoder through a 1×1 convolutional kernel and then concatenates them along the channel dimension to generate the initial state of the decoder; the operations performed by the fusion module include:

[0037] First, for the hidden state , and output by each encoder, reduce the dimensions through a 1×1 convolutional layer and map them to a unified feature space, expressed as:

[0038] ;

[0039] ;

[0040] ;

[0041] Among them, and represent the spatial dimensions respectively, is the feature dimension after dimensionality reduction, corresponds to radar data, corresponds to lightning data, corresponds to AWS data;

[0042] The hidden state and memory state after dimensionality reduction are fused through a concatenation operation along the channel dimension to generate a fused feature , which is expressed as:

[0043] ;

[0044] Among them, represents the concatenation operation on the channel dimension;

[0045] After that, through the correlation-aware dynamic weight allocation mechanism, including weighted summation of the concatenated features by a learnable weight coefficient to generate the final fused feature , which is expressed as:

[0046] ;

[0047] ;

[0048] ;

[0049] Among them, are the weights corresponding to radar, lightning, and AWS data respectively;

[0050] During the training process, cross-modal adaptive feature enhancement is achieved through a competitive gradient learning strategy, including: updating the weights by optimizing the gradient formula , which is expressed as:

[0051] ;

[0052] Among them, is the loss function, is the weight corresponding to the modality, is the hidden state of the th modality of the th sample, is the number of samples, is the regularization coefficient.

[0053] Optionally, in step S2, it further includes training the model using binary cross-entropy and the Adam optimizer, using binary cross-entropy as the loss function to quantify the gap between the predicted value and the true observed value, and using the Adam optimizer for parameter update, and adjusting the hyperparameters of the model according to the experimental results. The hyperparameters at least include the learning rate, the convolutional kernel size, and the number of filters.

[0054] Optionally, the lightning prediction result includes the spatial distribution and the time distribution of lightning occurrence. The lightning prediction result is output as one-dimensional rasterized data, and the value of each grid cell represents the probability of lightning occurrence in that cell. The output result is organized according to the time dimension and stored as a structured file.

[0055] To achieve the above object, the second aspect of the present application provides a lightning prediction system based on GCN and ConvLSTM. The system includes:

[0056] A data preprocessing module that uses multi-source meteorological data as multi-source data input and performs preprocessing operations on it. The multi-source meteorological data includes automatic weather station data, radar data, and lightning observation data. Among them, the automatic weather station data and radar data are used to extract the spatio-temporal characteristics of lightning occurrence, and the lightning observation data is used as the true label for supervising model training.

[0057] An ST-GLNet prediction module. The ST-GLNet prediction module includes an ST-GLNet prediction model. The architecture of the prediction model includes an input layer, an encoder, a fusion module, a decoder, and an output layer. It processes the input data and outputs the lightning prediction result to achieve lightning prediction.

[0058] The input layer is provided with multiple input channels corresponding to input data of different dimensions, and each input channel is provided with a single encoder to extract the spatio-temporal characteristics of the input data.

[0059] The input layer is provided with multiple input channels corresponding to input data of different dimensions, and each input channel is provided with a single encoder to extract the spatio-temporal characteristics of the input data.

[0060] The encoder includes a GCN layer and a ConvLSTM layer. The GCN layer is used to extract the spatial features of the input data, including introducing spectral graph regularization, constructing an adjacency matrix through a physically constrained method, defining the energy propagation rule between nodes as the topological connection of horizontally and vertically adjacent units, and combining self-loop connection and symmetric normalization processing. It also includes stacking the GCN layer and the ConvLSTM layer through a topological-aware spatio-temporal hierarchical structure to model spatio-temporal features at multiple levels. The ConvLSTM layer performs time series modeling on the output data of the GCN layer. Among them, the hidden state update mechanism of the ConvLSTM layer is given physical topological constraints, and the convolutional kernel of the ConvLSTM layer shares the same spatial relationship with the adjacency matrix; and

[0061] The fusion module fuses the outputs of the encoder in all input channels through a correlation-aware dynamic weight allocation mechanism and a competitive gradient learning strategy as the input of the decoder.

[0062] To achieve the above object, the third aspect of the present application provides a lightning prediction device based on GCN and ConvLSTM, including a processor and a memory. A computer program is stored on the memory. When the computer program is executed by the processor, the method described above is implemented.

[0063] After adopting the above technical solutions, the present application has the following beneficial effects compared with the prior art:

[0064] In the present application, based on the adjacency matrix construction mechanism of the graph convolutional network (GCN), the spatial dependence relationship between grid points can be accurately quantified. Through multi-layer graph convolutional operations, the model can not only effectively extract complex spatial features, but also show significant advantages in high-resolution grid prediction. Compared with traditional methods, the present application effectively reduces the problem of over-wide coverage of the alarm area caused by insufficient spatial resolution in traditional prediction by finely identifying high / low probability distribution areas of lightning, significantly improves the spatial positioning accuracy, and the ST-GLNet model shows stronger time consistency and dynamic adaptability, and can handle lightning events with a longer time span. By introducing the convolutional long short-term memory network (ConvLSTM) to construct a time series memory network and maintaining the hidden state and cell state of the time step, the time-varying features and non-linear evolution laws of lightning events can be effectively captured. Compared with traditional static prediction methods, the present application can dynamically track the three-dimensional evolution process of the lightning system. In the scenario of severe convective weather, the ST-GLNet shows better time continuity modeling ability and can accurately capture the evolution trajectory of lightning events with a longer time span.

[0065] In this application, radar data, Automatic Weather Station (AWS) data, and lightning observation data are effectively fused. Radar data provides powerful spatial background information, AWS data provides multi-dimensional meteorological parameters, and lightning observation data is used as the true label for training guidance. By fusing multi-modal data, the key factors of lightning occurrence can be mined from multiple dimensions, effectively improving the prediction accuracy of lightning occurrence areas. Compared with traditional methods based on a single data source, this application significantly reduces the false alarm rate and missed alarm rate, and improves the reliability.

[0066] In this application, combined with the architectures of GCN and ConvLSTM, the ST-GLNet model can efficiently process different types and dimensions of meteorological data, and has stronger adaptability when facing different spatial distributions, time spans, and data sources. This application can flexibly meet different early warning requirements, support large-scale data processing and real-time prediction, and demonstrates excellent scalability under various meteorological conditions.

[0067] The following further describes the specific embodiments of this application in detail with reference to the accompanying drawings. Description of the Drawings

[0068] The accompanying drawings, as part of this application, are used to provide a further understanding of this application. The schematic embodiments and descriptions of this application are used to explain this application, but do not constitute an improper limitation of this application. Obviously, the accompanying drawings in the following description are only some embodiments, and those of ordinary skill in the art can obtain other drawings based on these drawings without creative efforts.

[0069] The accompanying drawings are:

[0070] Figure 1 It is the overall logic schematic diagram of the lightning prediction method in this specific embodiment;

[0071] Figure 2 It is the logic schematic diagram of the prediction method based on ST-GLNet in this specific embodiment;

[0072] Figure 3 It is the logic schematic diagram of the ST-GLNet prediction model in this specific embodiment;

[0073] Figure 4 It is the result schematic diagram predicted by the ST-GLNet prediction model at time t in this specific embodiment;

[0074] Figure 5 It is the record schematic diagram of the actual occurrence of lightning at time t in this specific embodiment;

[0075] Figure 6It is a schematic comparison diagram of the average metrics of ST-GLNet and ConvLSTM using meteorological observation data (AWS) and lightning observation data (LIG) in this specific embodiment;

[0076] Figure 7 It is a schematic comparison diagram of the average metrics of ST-GLNet and ConvLSTM using meteorological observation data (AWS), lightning observation data (LIG), and radar data (Radar) in this specific embodiment. Specific embodiment

[0077] To make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the following will clearly and completely describe the technical solutions in the embodiments in conjunction with the accompanying drawings in the embodiments of the present application. The following embodiments are used to illustrate the present application but are not used to limit the scope of the present application.

[0078] In view of the unique spatial energy propagation characteristics and multi-source data heterogeneity of the lightning prediction task, a systematic innovation has been made to the combined architecture of traditional GCN and ConvLSTM. Please refer to Figure 1 This application provides a lightning prediction method based on GCN and ConvLSTM, including the following steps:

[0079] Step S1: Use multi-source meteorological data as multi-source data input and perform preprocessing operations on it. The multi-source meteorological data includes automatic weather station data, radar data, and lightning observation data. Among them, the automatic weather station data and radar data are used to extract the spatio-temporal characteristics of lightning occurrence, and the lightning observation data is used to supervise the true labels of model training;

[0080] Step S2: Construct and train an ST-GLNet prediction model based on GCN and ConvLSTM. The architecture of the ST-GLNet prediction model includes an input layer, an encoder, a fusion module, a decoder, and an output layer. Process the input data and output the lightning prediction result to achieve lightning prediction;

[0081] The input layer is provided with multiple input channels corresponding to input data of different dimensions, and each input channel is provided with a single encoder to extract the spatio-temporal characteristics of the input data; the encoder includes a GCN layer and a ConvLSTM layer. The GCN layer is used to extract the spatial characteristics of the input data, and the ConvLSTM layer performs time series modeling on the output data of the GCN layer; the fusion module is used to fuse the outputs of the encoder modules in all input channels as the input of the decoder; the decoder is used to restore the spatial resolution and generate the predicted lightning distribution map; the output layer splices the outputs of the decoder through the time dimension to generate the final prediction sequence and outputs one-dimensional rasterized data representing the lightning occurrence probability.

[0082] Specifically, in terms of spatial feature modeling, through the construction method of the physical constraint adjacency matrix, the energy propagation rule between nodes is defined as the topological connection of horizontally and vertically adjacent units. Combining self-loop connection and symmetric normalization processing, the conservation characteristics of atmospheric electrostatic field energy are explicitly simulated. In time dynamics modeling, the hidden state update mechanism of the ConvLSTM layer is given physical topological constraints, and its convolution kernel shares the same spatial relationship with the adjacency matrix to ensure that the spatio-temporal evolution conforms to the dynamic law of charge accumulation and release in severe convective weather. In the multi-source data fusion link, a correlation-aware dynamic weight allocation mechanism is proposed, and an adaptive feature enhancement between modalities is achieved through a competitive gradient learning strategy, effectively suppressing noise interference while enhancing high-value signals related to lightning.

[0083] Specifically, in the model training stage, spectral graph regularization constraints are introduced. By aligning the energy distribution of the feature propagation process with the physical topology, the model's ability to express lightning energy transfer features is significantly improved. The decoder adopts a hybrid upsampling strategy that combines deconvolution kernels and interpolation algorithms, maintaining the multi-modal long-range dependence relationship during the spatial resolution reconstruction process. The finally output lightning probability field fully depicts the charge transport process through temporal stacking, forming an interpretable prediction result that conforms to the characteristics of atmospheric electricity.

[0084] It should be noted that the execution subject of the lightning prediction method in this embodiment is the ST-GLNet prediction device based on the graph convolutional network (GCN) and the convolutional long short-term memory network (ConvLSTM). This device can be an electronic device, a component in an electronic device, an integrated circuit, or a chip. The electronic device can be a mobile electronic device or a non-mobile electronic device. Exemplarily, the mobile electronic device can be a mobile phone, a tablet computer, a laptop computer, a handheld computer, a vehicle-mounted electronic device, a wearable device, etc., and the non-mobile electronic device can be a server and a personal computer, etc. This application does not make specific limitations. Hereinafter, taking the execution subject as a server as an example, the ST-GLNet prediction method based on the graph convolutional network (GCN) and the convolutional long short-term memory network (ConvLSTM) in this embodiment will be described.

[0085] In addition, the terms "first" and "second" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, the features defined with "first" and "second" may explicitly or implicitly include one or more of such features. In the description of the embodiments of this application, the meaning of "a plurality" is two or more, unless otherwise specifically defined.

[0086] In an implementable embodiment, the input layer includes three independent input channels. The first channel receives radar data with a dimension of where is the number of features of radar data; the second channel receives rasterized data of lightning observation events of a single channel with a dimension of (i.e., the lightning observation data has only one feature); the third channel receives rasterized data of an Automatic Weather Station (AWS) with a multi-channel dimension of (i.e., the automatic weather station contains multiple features), where is the number of features of the rasterized data of the automatic weather station, and the three input channels are processed by independent encoders respectively.

[0087] In an implementable embodiment, the first channel is a radar feature channel: it receives a normalized radar reflectivity data set with an input dimension of where and are respectively the number of grid divisions in the longitude and latitude directions of the geographical grid data, and represents the radar data feature dimension; the second channel is a lightning observation channel: it inputs rasterized observation data of single-channel lightning events (LIG) with a dimension of and the value of each grid cell represents a binary label of the lightning event; the third channel is a meteorological data channel: it processes rasterized fusion data of multi-source meteorological station parameters with an input dimension of where is the number of meteorological features of the Automatic Weather Station (AWS).

[0088] In this embodiment, the radar data, AWS data, and lightning observation data are effectively fused. The radar data provides powerful spatial background information, the AWS data provides multi-dimensional meteorological parameters, and the lightning observation data is used as the true label for training guidance. By fusing multi-modal data, the model can mine the key factors of lightning occurrence from multiple dimensions and effectively improve the prediction accuracy of lightning occurrence areas.

[0089] In another implementable embodiment, for AWS data, each meteorological station has only one fixed location, and the number is far from enough to fill each grid in the grid map. Therefore, aiming at the complexity and diversity of AWS data, an efficient data preprocessing method is proposed, covering multiple links from data file management, reading, cleaning, multi-source data merging, grid point and site data association, interpolation calculation, to multi-thread optimization, result output and file lock mechanism, and finally format conversion.

[0090] Specifically, the preprocessing of AWS data in step S1 includes the following steps:

[0091] Step S10, executing a data reading operation, reading a CSV data file that complies with the "YYYY-MM-DD-HH" naming format from a specified directory, parsing the file name to extract timestamp information and recording an abnormal log to ensure data integrity and reliability of preprocessing;

[0092] Step S11, perform data cleaning operations, and remove abnormal records according to preset rules. Abnormal records include abnormal temperature, humidity and wind speed parameters, such as temperature values exceeding the minimum and maximum ranges, relative humidity not between 0 and 100, and maximum wind speed exceeding 35 m / s, so as to improve data quality;

[0093] Step S12, using a multi-source data fusion strategy to merge multiple site data files with the same timestamp to generate a unified data set containing latitude and longitude coordinates and meteorological parameters;

[0094] Step S13, constructing a grid center point coordinate matrix, and matching the two nearest meteorological stations for each grid point using a nearest neighbor algorithm;

[0095] Step S14, using the Kriging interpolation method with disturbance optimization to perform spatial interpolation calculations on the temperature, humidity and wind speed parameters, wherein a random disturbance of ±0.0001 degrees / unit is added to the station coordinates and the observed values;

[0096] Step S15, using a multi-threaded parallel computing architecture, outputting the interpolation results in a CSV format with a file lock mechanism, and storing them according to a timestamp naming rule;

[0097] Step S16, reshape the CSV data into a two-dimensional array structure that matches the deep learning model input, and store it in a .npy format file according to meteorological parameter classification.

[0098] In practical applications, AWS data preprocessing includes file management, data cleaning, interpolation calculation, and format conversion. Data files are named as "YYYY-MM-DD-HH" to represent year, month, day, and hour. The time information is extracted by parsing the file name and the log is recorded to ensure data integrity. Abnormal temperature, humidity, and wind speed values are removed during the cleaning process to improve data quality. Subsequently, the Kriging interpolation method is used to map the site observation data to the spatial grid, and key meteorological parameter values are generated for each grid point. To improve efficiency, the interpolation process uses multi-threaded parallel computing, and the results are stored in .npy format, providing efficient input for subsequent model processing.

[0099] It should be noted that this embodiment realizes efficient processing of AWS data from reading to interpolation calculation, providing high-quality input data for the operation of lightning prediction models. At the same time, it adopts multi-source data fusion strategy and Kriging interpolation algorithm to improve data spatial continuity and temporal resolution, laying the foundation for accurate lightning warning.

[0100] In an implementable embodiment, the preprocessing format conversion step converts the cleaned and interpolated CSV format data into the.npy format that can be directly read by the model, improving the model running efficiency and reducing the data reading overhead.

[0101] Specifically, the format conversion module converts the CSV format data into the directly readable.npy format, which specifically includes the following steps:

[0102] Step S160, read the timestamp-marked CSV files in chronological order from the directory of the preprocessed CSV files, and the file names follow the format of "YYYY-MM-DD-HH";

[0103] Step S161, extract the key meteorological parameter data fields of temperature, humidity, wind speed, and air pressure in the CSV files, and at the same time parse the file names to obtain the date and time information and rename them;

[0104] Step S162, reshape the one-dimensional data sequences of each meteorological parameter into a two-dimensional matrix according to the number of longitude dimensions and latitude dimensions of the preset grid map, generating the length of the grid map and width formed two-dimensional array;

[0105] Step S163, create independent storage directories for each meteorological parameter, and store the reshaped two-dimensional arrays as.npy format files according to the parameter categories respectively;

[0106] Step S164, organize the output data using a hierarchical path structure.

[0107] Specifically, during the conversion process, read the site data files one by one from the CSV folder storing the preprocessing results, extract and process the key meteorological parameters, extract the date and time information from the file names and rename them. Reshape the meteorological parameter data in the one-dimensional array format into a two-dimensional array with the length of and width of formed to match the spatial rasterized structure required by the model input. The reshaped data are saved as.npy files respectively, and each meteorological parameter is stored in an independent folder. The file naming adopts a unified time identification format to ensure clear and orderly data. The storage path of the output data is designed hierarchically, and the data of different meteorological parameters are stored in independent folders, improving the automation degree of data conversion, ensuring clear data management, and avoiding the risks of file chaos or overwriting.

[0108] In another realizable embodiment, to solve the problem of insufficient radar data processing in existing lightning prediction methods, this embodiment realizes the efficient combination of radar data and other data sources through steps such as data reading, integration, cleaning, correlation, interpolation, and fusion, providing high-quality input data for the model.

[0109] Specifically, the preprocessing of radar data in step S1 includes the following steps:

[0110] Step S100, read the radar data through the data integration module, extract the longitude and latitude grid, meteorological variables such as CREF, ET, VIL, and scan time, where CREF is the composite reflectivity, ET is the echo top height, VIL is the vertically integrated liquid water content, and construct a time-series radar data cube;

[0111] Step S110, sort the multi-radar data within the same hour according to the timestamp, perform grid-level numerical accumulation and file number statistics on each meteorological variable, and generate a spatio-temporal average data array;

[0112] Step S120, perform data cleaning operations, use neighborhood interpolation to fill in missing grid points, identify and correct outliers beyond the preset range through the dynamic threshold method, compare the data of multiple files within the same hour, detect data acquisition or transmission errors, and mark and remove a file data as abnormal when it is significantly different from other files;

[0113] Step S130, establish a spatial correlation model, map the longitude and latitude of the automatic weather station to the center point of the radar grid, and construct a site-grid matching relationship matrix;

[0114] Step S140, use the adaptive Kriging interpolation algorithm to interpolate meteorological parameters such as wind speed, temperature, and relative humidity of the automatic weather station to the radar grid, perform downsampling matching when the radar resolution is higher than that of the automatic weather station to ensure the smooth progress of fusion, and perform three-dimensional interpolation repair based on the continuity of the meteorological field for the missing area of the radar data;

[0115] Step S150, output the fused multi-dimensional data tensor, including radar reflectivity, AWS interpolation parameters, and time-coded features.

[0116] Specifically, the preprocessing of radar data starts with file reading and time integration, and extracts longitude and latitude, data variables (such as CREF, ET, VIL), and time attributes through tools such as cinrad. To ensure the time consistency of the data, the files within the same hour are sorted in chronological order and processed one by one. By accumulating the file variable values and calculating the average, representative data for each hour is generated, which not only maintains the integrity of the time dimension but also reduces the impact of single-file fluctuations. In the cleaning stage, invalid values and outliers of the spring steel plate are repaired, the data quality is improved by interpolation or removing missing data, and the files with abnormal data are removed to ensure consistency.

[0117] After the spatio-temporal features are extracted from AWS, LIG, and Radar, the hidden states and memory states output by each encoder are subjected to 1×1 convolution for dimensionality reduction, and the dimensionality-reduced states are concatenated along the last dimension to form a fused feature. In the multi-data fusion stage, in this embodiment, the longitude and latitude information of the grid center points of the radar data is used to match the AWS site locations, and the AWS data is interpolated onto the grid points of the radar data. When the resolution of the radar data is significantly higher than that of the AWS data, the resolution is aligned through downsampling to ensure the fusion accuracy. At the same time, for the locally missing areas in the radar data, interpolation filling is used for repair to ensure the continuity and integrity of the spatial data.

[0118] In an implementable embodiment, the preprocessing of the lightning observation data in step S1 includes the following steps:

[0119] Step S1000, receiving the lightning observation data file and the grid data file. The lightning observation data contains the longitude and latitude coordinates and timestamps of each lightning event, and the grid data contains the longitude and latitude of each grid center point and the unique number;

[0120] Step S1100, using the CUDA acceleration technology parallel computing framework to batch-convert the longitude and latitude of the grid center points into radian tensor data, and loading the longitude and latitude data in the lightning observation data at the same time;

[0121] Step S1200, using the haversine formula to perform parallel spherical distance calculations for each lightning event and all grid center points. The calculation formula is:

[0122] ;

[0123] represents the distance between two points; represents the radius of the earth (about 6371 km on average); and are the latitudes of the two points, expressed in radians; represents the latitude difference between the two points; and represents the longitudes of two points, in radians; represents the longitude difference between two points;

[0124] Step S1300, set the matching threshold radius, mark the grid number closest in distance through CUDA atomic operations, and record a special identifier for unmatched events;

[0125] Step S1400, output the lightning event dataset with grid encoding, and sort and store it in CSV format according to the occurrence time;

[0126] Step S1500, construct a timestamp sequence, group and aggregate lightning events by "year, month, day, hour" to generate a binary grid matrix, set the grid positions where events occur to 1, and the rest to 0;

[0127] Step S1600, store the converted binary grid in independent files in hours, with the file names in standard time format, and the content is the binary values stored row by row, representing the position status of the grid.

[0128] Specifically, the lightning observation data calculates the spherical distance from each event to the center point of the grid through the CUDA-accelerated Haversine formula, and matches the event to the closest grid number. Finally, a binary grid file stored in hours is generated, representing the occurrence status of lightning events in each grid cell (1 means occurrence, 0 means non-occurrence). All data is organized in a time series and stored in.npy format, providing structured and high-quality input support.

[0129] It should be noted that in this embodiment, the lightning observation data is converted from CSV format to binary grid format, providing high-quality input data for the efficient operation of the lightning prediction model. This format has a clear structure and efficient storage, and can accurately reflect the spatial distribution of lightning events per hour, providing a solid foundation for the time and space modeling of the model. The preprocessing method of this embodiment successfully maps the lightning observation data from a single event form to spatial grid cells, improves the calculation efficiency with the help of CUDA acceleration technology, ensures the spatial accuracy of the data, and lays a solid foundation for the spatio-temporal modeling of large-scale lightning events.

[0130] To address the essential mismatch problem between grid isolation processing and physically-driven energy propagation dynamics, we propose an innovative integration of spectral graph regularization and topological-aware spatio-temporal hierarchies. Specifically: Spectral graph regularization includes performing an operation of pre-building a fixed adjacency matrix based on the spatial adjacency relationship of grid cells, where pre-building means capturing the spatial dependence relationship between grid points by pre-building a fixed adjacency matrix based on the spatial adjacency relationship of grid cells, specifically including:

[0131] Each grid cell is mapped to a node in the graph, and an adjacency matrix is established through four-way connections , where is the number of grid cells, expressed as:

[0132] ;

[0133] Among them, represents the grid spacing, and are the coordinates of the corresponding grid nodes and node in the two-dimensional space;

[0134] After that, in order to maintain the stability and effectiveness in graph convolution operations, the adjacency matrix is symmetrically normalized by the node degree matrix to obtain the normalized adjacency matrix , where is the identity matrix, ensuring that each node can be connected to itself;

[0135] The spectral graph regularization is further enhanced through Laplacian smoothing, expressed as:

[0136] ;

[0137] Among them, represents the node features of the th layer, represents the node features of the th layer, is the trainable parameter of the current layer, is the normalized adjacency matrix, represents the spectral graph convolution operation, is the smoothing scale parameter, represents the Frobenius norm, and the regularization term is the auxiliary loss term penalizes the deviation between the node features of the th layer and the smoothed result (i.e., the spectral graph convolution mode) of the th layer .

[0138] Perform the operation of pre-building a fixed adjacency matrix based on the spatial adjacency relationship of grid cells. The pre-building includes mapping each grid cell to a graph node and establishing the initial structure of the fixed adjacency matrix with four-way connections between nodes; symmetrically normalizing the fixed adjacency matrix by the node degree matrix to form the stable neighborhood propagation constraint condition for graph convolution;

[0139] Perform two-layer graph convolution operations to capture the spatial dependence relationship between grid points and neighbor nodes and achieve local and global feature extraction.

[0140] After the data preprocessing is completed, the model models the topological structure of the spatial data through a fixed adjacency matrix, and the construction of the fixed adjacency matrix is based on the rasterized spatial relationship of the input data. Each raster cell is regarded as a node in the graph, and connections are only established with its neighbors above, below, left, and right, thus forming a regular grid graph. The fixed adjacency matrix refers to the adjacency matrix pre-constructed according to the spatial adjacency relationship of the raster cells in the preprocessing stage, and this matrix remains unchanged in subsequent graph convolution operations. To avoid instability caused by uneven connection structures in graph convolution operations, this embodiment further normalizes the fixed adjacency matrix. Specifically, by calculating the degree matrix of each node, an inverse square root normalization operation is performed on the fixed adjacency matrix, thereby improving the stability of graph convolution operations and the accuracy of numerical calculations. Using this fixed adjacency matrix, GCN can efficiently capture the spatial dependence and local correlation between raster points and neighbor nodes, providing strong support for the high-resolution prediction of the model.

[0141] In an implementable embodiment, the encoder adopts cascaded GCN layers and ConvLSTM layers, including adopting a topology-aware spatio-temporal hierarchical structure, and modeling spatio-temporal features by stacking three GCN layers and ConvLSTM layers; the goal is to integrate spatial dependence relationships through a graph convolutional network (GCN) and capture temporal dependence relationships through ConvLSTM to achieve the prediction of lightning events evolving over time. Among them, the GCN layer extracts the spatial features of the input data as the input of the ConvLSTM layer, including receiving three types of inputs: radar, lightning observation, and AWS data, and extracting the spatio-temporal features of each data respectively. Taking radar data as an example, it is expressed as:

[0142] ;

[0143] Among them, is the time step, is the number of grid nodes, is the radar data feature dimension;

[0144] The GCN layer extracts the spatial features of raster nodes through two-layer convolution operations. In each layer of GCN, the feature information of adjacent nodes is aggregated, which is expressed as:

[0145] ;

[0146] Among them, represents the node features of the th layer, is the trainable parameter of the current layer, is the activation function, is the normalized adjacency matrix, is the node degree matrix, It is the output feature after the GCN operation of this layer.

[0147] In an implementable embodiment, the hidden state and memory state at each time step are generated and updated through the ConvLSTM layer, including after obtaining the spatial features, inputting them into the ConvLSTM layer for temporal modeling, and the ConvLSTM layer updates the hidden state at each time step and the memory state , ConvLSTM can capture the dynamic evolution in time and establish the temporal features of grid points changing over time, expressed as:

[0148] ;

[0149] ;

[0150] ;

[0151] ;

[0152] ;

[0153] ;

[0154] where is the activation function, is the input gate, is the forget gate, is the output gate, , , and are the bias terms, represents the calculation process of the ConvLSTM cell, is the feature output from the last layer of GCN, and represent the memory state and hidden state at the current time step respectively, is the adjacency matrix explicitly constrains the local receptive field of the convolutional kernel through the Hadamard product, and the convolutional kernels and have a sliding window pattern aligned with the topological structure of the adjacency matrix , inherit the state persistence of the physical system through the memory state at the previous time step in the time dimension, and inherit the topological dependence of the GCN layer through the operation in the spatial dimension.

[0155] In an implementable embodiment, the fusion module reduces the dimensions of the hidden state and memory state output by each encoder through a 1×1 convolutional kernel and then concatenates them along the channel dimension to generate the initial state of the decoder. The operations performed by the fusion module include:

[0156] First, for the hidden state output by each encoder , and , the dimensions are reduced through a 1×1 convolutional layer and mapped to a unified feature space, expressed as:

[0157] ;

[0158] ;

[0159] ;

[0160] Among them, and respectively represent the spatial dimensions, is the feature dimension after dimension reduction, corresponds to radar data, corresponds to lightning data, corresponds to AWS data; through this dimension reduction operation, it is ensured that features from different modalities can be effectively fused in the same spatial dimension;

[0161] The dimension-reduced hidden state and memory state are fused through a concatenation operation along the channel dimension to generate a fused feature , expressed as:

[0162] ;

[0163] Among them, represents the concatenation operation on the channel dimension; the concatenated fused feature is a tensor, containing the joint information of data sources from different modalities;

[0164] After that, through a correlation-aware dynamic weight assignment mechanism, including weighted summation of the concatenated features with learnable weight coefficients , the final fused feature is generated, and each modality's feature is assigned a dynamically adjusted weight for adaptive fusion according to the correlation of each modality, expressed as:

[0165] ;

[0166] ;

[0167] ;

[0168] Among them, are the weights corresponding to radar, lightning, and AWS data respectively; this weighted fusion strategy ensures that the mutual influence between different data sources is effectively fused, and at the same time, it also ensures that in different scenarios, the dynamic adjustment of weights can adapt to the diversity of data;

[0169] During the training process, inter-modal adaptive feature reinforcement is achieved through a competitive gradient learning strategy, including: updating the weights by optimizing the gradient formula , expressed as:

[0170] ;

[0171] Among them, is the loss function, is the weight corresponding to the modality, is the th hidden state of the th modality of the th sample, is the number of samples, is the regularization coefficient; this gradient update formula encourages competitive learning between different modalities - when the features of a certain modality have a strong correlation with the features of other modalities, the weight of this modality will increase. At the same time, the regularization term

[0172] prevents the situation where a single modality dominates the entire model. In addition, the memory state C is responsible for storing long-term memories, mainly capturing the slowly changing features in the time series, such as the environmental background state. These states are more stable, and the memories of different patterns are more complementary rather than competitive. Therefore, for the memory state

[0173] .

[0174] In an achievable implementation, the decoder is designed as a resolution restoration network, specifically including:

[0175] Temporal convolutional unit: 32 3×3 convolutional kernels are used to compress the spatio-temporal feature dimension to ;

[0176] Dynamic reconstruction layer: The ConvLSTM layer maintains 32 layers to iteratively update the state of the temporal features and maintain the temporal coherence of the results;

[0177] Spatial resolution restoration component: Upsample spatially through a deconvolution layer with a stride of 2×2, gradually restoring the raster dimension to the original input resolution.

[0178] In an achievable implementation, the fusion module reduces the dimensions of the hidden states ( ) and memory states ( ) output by each encoder through a 1×1 convolutional kernel and then concatenates them along the channel dimension to generate the initial state of the decoder. Here, the superscript R represents the hidden or memory state from the radar channel, the superscript L represents the hidden or memory state from the lightning observation channel, and the superscript A represents the hidden or memory state from the automatic weather station (AWS) channel. The fusion module performs operations as follows:

[0179] Perform 1×1 convolutional dimensionality reduction on the hidden states and memory states output by each encoder;

[0180] Concatenate the dimension-reduced states along the last dimension to form a fused feature;

[0181] Use downsampling convolution and fully connected layers to generate the initial state of the decoder.

[0182] In another achievable implementation, the decoder includes a cascaded structure of a temporally distributed convolutional layer, a ConvLSTM layer, and a deconvolution layer. The decoder is configured to gradually restore the spatial resolution and generate a predicted lightning distribution map. The operations performed by the decoder include the following steps:

[0183] The temporally distributed convolutional layer applies a convolutional operation to the input features in the time dimension to extract spatial features in the time series;

[0184] The ConvLSTM layer captures the dynamic change features in the time series;

[0185] Gradually upsample the feature map through multi-layer deconvolution operations to restore high-resolution spatial information.

[0186] It should be noted that Figure 2 in , , …, represents the time series input data, generally referring to the meteorological observation data from the 1st frame to the t-1th frame (or from the 1st moment to the t-1th moment). The hidden layer means that for the raster data of a certain moment or a certain input channel (denoted as X1), after the input data undergoes certain preprocessing or convolutional operations, it will be sent to the next step (hidden layer / encoder) for calculation. Y represents the final "lightning probability distribution", and σ represents using the Sigmoid function to map the data to the range (0, 1).

[0187] Figure 3 in Denote the spatio-temporal data input to the model at time t; and are respectively the cell state and memory state representing the output of the X data encoder. Denote respectively the hidden states output by the encoders of radar data, lightning observation data, and AWS observation data; Denote respectively the memory states output by the encoders of radar data, lightning observation data, and AWS observation data; C and H denote the fused memory state and hidden state. Denote the predicted output result.

[0188] Please refer to Figure 3 , in an implementable embodiment, the decoder part supports an initialization optimization strategy. By performing a power transformation and Sigmoid function normalization on the input data, the distribution characteristics of the input data are optimized. The power transformation is used to adjust the dynamic range of the data to make it closer to a normal distribution; the Sigmoid function maps the data to the range (0, 1) to ensure that the input data has a consistent scale. This strategy significantly improves the convergence speed and prediction accuracy of the model, while enhancing the stability of training.

[0189] Please refer to Figure 3 , Figure 4 and Figure 5 , in practical applications, the output layer performs a Cropping3D operation to ensure that the output size matches. The final output feature map is cropped to ensure that the output size is consistent with the input data; the predicted results output by the decoder are concatenated in the time dimension to generate a complete prediction sequence, and the output is one-dimensional rasterized data representing the probability distribution of lightning occurrence. Figure 4 The numbers in Figure 4 and Figure 5 represent probability values, and the abscissa in

[0190] Specifically, the output layer performs result optimization operations, specifically including:

[0191] Boundary alignment unit: Use a Cropping3D layer to crop redundant feature boundaries to ensure that the output geographical raster size is consistent with the input data;

[0192] Probability conversion module: After activation by the Sigmoid function, generate a one-dimensional lightning probability raster data matrix, and each element represents the lightning occurrence probability value of the corresponding geographical unit within the time series window.

[0193] In an implementable embodiment, in step S2, it further includes prediction model training and optimization;

[0194] The model is trained using binary cross - entropy and the Adam optimizer. Binary cross - entropy is used as the loss function to quantify the gap between the predicted values and the true observed values, and the Adam optimizer is used for parameter updates. The hyperparameters of the model are adjusted according to the experimental results, and the hyperparameters at least include the learning rate, the convolutional kernel size, and the number of filters.

[0195] In another feasible implementation, lightning prediction result output and application;

[0196] The lightning prediction results include the spatial distribution and temporal distribution of lightning occurrences. The lightning prediction results are output as one - dimensional rasterized data, and the value of each raster cell represents the probability of lightning occurrence in that cell. The output results are organized according to the time dimension and stored as a structured file.

[0197] To verify the effectiveness of the ST - GLNet model, the model performance was evaluated using two different sets of data under the same cross - entropy weight and the same learning rate, and was compared with the classical model ConvLSTM. The evaluation metrics include threat score (TS), equal - weight threat score (ETS), probability of detection (POD), false alarm rate (FAR), mean square error ratio (MAR), bias score (BS), and accuracy (AC). For the experimental results, please refer specifically to Figure 6 and Figure 7 .

[0198] Please refer to Figure 6 , using meteorological observation data (AWS) and lightning observation data (LIG). The experimental results show that ST - GLNet is superior to ConvLSTM in most evaluation metrics. Specifically, ST - GLNet achieved a threat score (TS) of 0.3043, which is a 15.7% improvement compared to 0.2629 of ConvLSTM. The probability of detection (POD) is 0.7504, significantly better than 0.7066 of ConvLSTM. In terms of accuracy (AC), ST - GLNet is slightly higher than ConvLSTM, being 0.9717 and 0.9653 respectively. In addition, ST - GLNet performs better in the equal - weight threat score (ETS), and is slightly better in the false alarm rate (FAR) and mean square error ratio (MAR) metrics, but the difference from ConvLSTM is small.

[0199] Please refer to Figure 7, meteorological observation data (AWS), lightning observation data (LIG), and radar data (Radar) are used. The experimental results show that ST-GLNet also exhibits excellent performance. Its threat score (TS) is 0.2531, showing an 8.5% improvement compared to 0.2332 of ConvLSTM. The probability of detection (POD) reaches 0.6611, higher than 0.5724 of ConvLSTM, indicating that ST-GLNet has strong detection ability. In terms of accuracy (AC), ST-GLNet and ConvLSTM perform similarly, being 0.9755 and 0.9778 respectively, but ST-GLNet has a more excellent comprehensive performance in other metrics.

[0200] The ST-GLNet model outperforms the ConvLSTM model on different datasets, especially in key metrics such as threat score and probability of detection. This indicates that ST-GLNet can more effectively capture spatio-temporal features, improve prediction accuracy, and maintain a high generalization ability, providing a better solution for modeling complex scenarios.

[0201] Based on the same inventive concept, this application also provides a lightning prediction system based on GCN and ConvLSTM. The system includes:

[0202] A data preprocessing module that uses multi-source meteorological data as multi-source data input and performs preprocessing operations on it. The multi-source meteorological data includes automatic weather station data, radar data, and lightning observation data. Among them, the automatic weather station data and radar data are used to extract the spatio-temporal features of lightning occurrence, and the lightning observation data is used as the true label for supervising model training;

[0203] An ST-GLNet prediction module. The ST-GLNet prediction module includes an ST-GLNet prediction model. The architecture of the prediction model includes an input layer, an encoder, a fusion module, a decoder, and an output layer. It processes the input data and outputs the lightning prediction result to achieve lightning prediction;

[0204] The input layer is provided with multiple input channels corresponding to input data of different dimensions, and each input channel is provided with a single encoder to extract the spatio-temporal features of the input data;

[0205] The encoder includes a GCN layer and a ConvLSTM layer. The GCN layer is used to extract the spatial features of the input data, including introducing spectral graph regularization, constructing an adjacency matrix through a physically constrained method, defining the energy propagation rule between nodes as the topological connection of horizontally and vertically adjacent units, and combining self-loop connection and symmetric normalization processing. It also includes stacking the GCN layer and the ConvLSTM layer through a topology-aware spatio-temporal hierarchical structure to model spatio-temporal features at multiple levels. The ConvLSTM layer performs time series modeling on the output data of the GCN layer. Among them, the hidden state update mechanism of the ConvLSTM layer is given physical topology constraints, and the convolutional kernels of the ConvLSTM layer share the same spatial relationship with the adjacency matrix; and

[0206] The fusion module fuses the outputs of the encoders in all input channels through a correlation-aware dynamic weight allocation mechanism and a competitive gradient learning strategy as the input to the decoder.

[0207] Based on the same inventive concept, the present application also provides a lightning prediction device based on GCN and ConvLSTM, including a processor and a memory. When a computer program stored on the memory is executed by the processor, the method as described above is implemented.

[0208] The program product for implementing the above method in the present application can adopt a portable compact disc read-only memory and includes program code, and can run on a terminal device, such as a personal computer. However, the program product of the present application is not limited thereto. In the present application, a readable storage medium can be any tangible medium that contains or stores a program, and this program can be used by or in combination with an instruction execution system, device, or component.

[0209] It should be noted that a computer-readable storage medium can include a data signal propagated in a baseband or as part of a carrier wave, which carries the readable program code. Such a propagated data signal can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. A readable storage medium can also be any readable medium other than a readable storage medium, and this readable medium can send, propagate, or transmit a program for use by or in combination with an instruction execution system, device, or component. The program code contained on the readable storage medium can be transmitted by any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination of the above.

[0210] The above are only the preferred embodiments of the present application, and do not impose any form of limitation on the present application. Although the present application has been disclosed above with the preferred embodiments, it is not intended to limit the present application. Any person skilled in the art of the present application can make some changes or modifications to equivalent embodiments of equivalent changes by using the technical content prompted above within the scope of the technical solution of the present application. The implementation schemes in the above embodiments can also be further combined or replaced. However, as long as the content does not deviate from the technical solution of the present application, any simple modification, equivalent change and modification made to the above embodiments according to the technical essence of the present application still belong to the scope of the present application scheme.

Claims

1. A lightning prediction method based on GCN and ConvLSTM, characterized in that, Including the following steps: Step S1: Use multi-source meteorological data as multi-source data input and perform preprocessing operations on it. The multi-source meteorological data includes automatic weather station data, radar data, and lightning observation data. Among them, the automatic weather station data and radar data are used to extract the spatio-temporal characteristics of lightning occurrence, and the lightning observation data is used as the true label for supervising model training; Step S2: Construct and train an ST-GLNet prediction model based on GCN and ConvLSTM. The architecture of the ST-GLNet prediction model includes an input layer, an encoder, a fusion module, a decoder, and an output layer. Process the input data and output the lightning prediction result to achieve lightning prediction; The input layer is provided with multiple input channels corresponding to input data of different dimensions, and each input channel is provided with a single encoder to extract the spatio-temporal characteristics of the input data; The encoder includes a GCN layer and a ConvLSTM layer. The GCN layer is used to extract the spatial characteristics of the input data, including introducing spectral graph regularization, constructing an adjacency matrix through a physically constrained method, defining the energy propagation rule between nodes as the topological connection of horizontally and vertically adjacent units, combining self-loop connection and symmetric normalization processing; It also includes stacking the GCN layer and the ConvLSTM layer through a topology-aware spatio-temporal hierarchical structure to model spatio-temporal characteristics at multiple levels; The ConvLSTM layer performs time series modeling on the output data of the GCN layer. Among them, the hidden state update mechanism of the ConvLSTM layer is given a physical topology constraint, and the convolutional kernel of the ConvLSTM layer shares the same spatial relationship with the adjacency matrix; And The fusion module fuses the outputs of the encoders in all input channels through a correlation-aware dynamic weight allocation mechanism and a competitive gradient learning strategy as the input of the decoder.

2. The method according to claim 1, wherein The input layer includes three independent input channels. The first channel receives radar data with a dimension of , where and are respectively the grid division numbers in the longitude and latitude directions of the geographical grid data, and is the number of features of the radar data. The second channel receives single-channel lightning event rasterized data with a dimension of ; the third channel receives multi-channel automatic weather station rasterized data with a dimension of , where is the number of features of the automatic weather station rasterized data, and the three input channels are respectively processed by the independent encoders.

3. The method according to claim 1, characterized in that, The spectral graph regularization includes performing an operation of pre-constructing a fixed adjacency matrix based on the spatial adjacent relationship of grid cells. The pre-construction means capturing the spatial dependence relationship between grid points by pre-constructing a fixed adjacency matrix based on the spatial adjacent relationship of grid cells, specifically including: Each grid cell is mapped to a node in the graph, and an adjacency matrix is established through four-way connection , where is the number of grid cells, expressed as: ； Among them, represents the grid spacing, and are the coordinates of the corresponding grid node and node in the two-dimensional space; After that, the adjacency matrix is symmetrically normalized by the node degree matrix to obtain the normalized adjacency matrix , where is the identity matrix, ensuring that each node can be connected to itself; Further enhancing the spectral graph regularization through Laplacian smoothing, expressed as: ； Among them, represents the layer node features, represents the layer node features, is the trainable parameter of the current layer, is the normalized adjacency matrix, represents the spectral graph convolution operation, is the smoothing scale parameter, represents the Frobenius norm, auxiliary loss term penalizes the layer node features and the layer smoothing result deviation.

4. The method according to claim 1, wherein The encoder adopts a cascaded GCN layer and ConvLSTM layer, including adopting the topology-aware spatio-temporal hierarchical structure, and modeling spatio-temporal characteristics by stacking three layers of GCN layer and ConvLSTM layer; Among them, the GCN layer extracts the spatial characteristics of the input data as the input of the ConvLSTM layer, including receiving three types of inputs: radar, lightning observation, and AWS data, and respectively extracting the spatio-temporal characteristics of each data. Taking radar data as an example, it is expressed as: ； wherein, is the time step, is the number of grid nodes, is the radar data feature dimension; The GCN layer extracts the spatial characteristics of grid nodes through two convolutional operations. In each layer of GCN, aggregate the feature information of adjacent nodes, expressed as: ； Among them, represents the layer node features, is the trainable parameter of the current layer, is the activation function, is the normalized adjacency matrix, is the node degree matrix, is the output feature after the GCN operation of this layer.

5. The method according to claim 4, characterized in that Generate and update the hidden state and memory state at each time step through the ConvLSTM layer, including after obtaining the spatial features, inputting them into the ConvLSTM layer for temporal modeling, and the ConvLSTM layer updates the hidden state at each time step and the memory state , which is expressed as: ； ； ； ； ； ； Among them, is the activation function, is the input gate, is the forget gate, is the output gate, , , and are the bias terms, represents the calculation process of the ConvLSTM cell, is the feature output from the last GCN layer, and respectively represent the memory state and the hidden state at the current time step, is the adjacency matrix explicitly constrains the local receptive field of the convolutional kernel through the Hadamard product. The convolutional kernel and have a sliding window pattern aligned with the topological structure of the adjacency matrix . In the time dimension, it inherits the state persistence of the physical system through the memory state at the previous time step, and in the space dimension, it inherits the topological dependency of the GCN layer through the operation.

6. The method according to claim 5, wherein The fusion module reduces the dimensions of the hidden state and memory state output by each encoder through a 1×1 convolutional kernel and then concatenates them along the channel dimension to generate the initial state of the decoder; The operations performed by the fusion module include: First, for the hidden states output by each encoder , and , dimensionality reduction is performed through a 1×1 convolutional layer to map them to a unified feature space, expressed as: ； ； ； Among them, and respectively represent the spatial dimensions, is the feature dimension after dimensionality reduction, corresponds to radar data, corresponds to lightning data, corresponds to AWS data; The hidden state and memory state after dimensionality reduction are fused through a concatenation operation along the channel dimension to generate a fused feature , which is expressed as: ； Among them, represents the concatenation operation in the channel dimension; Afterwards, through the correlation-aware dynamic weight allocation mechanism, including passing the concatenated features through learnable weight coefficients to perform weighted summation to generate the final fused feature , which is expressed as: ； ； ； Among them, are the weights corresponding to radar, lightning, and AWS data respectively; During the training process, cross-modal adaptive feature enhancement is achieved through a competitive gradient learning strategy, including updating the weights by optimizing the gradient formula , which is expressed as: ； in, is the loss function, is the weight of the corresponding mode, It is The sample The hidden state of each modality, is the sample size, is the regularization coefficient.

7. According to the method described in claim 1, wherein, In step S2, it also includes training the model using binary cross-entropy and the Adam optimizer, using binary cross-entropy as the loss function to quantify the gap between the predicted value and the true observed value, and using the Adam optimizer for parameter update. The hyperparameters of the model are adjusted according to the experimental results. The hyperparameters at least include the learning rate, the convolutional kernel size, and the number of filters.

8. The method according to claim 1, wherein The lightning prediction result includes the spatial distribution and the time distribution of lightning occurrences. The lightning prediction result is output as one-dimensional rasterized data, and the value of each raster cell represents the probability of lightning occurrence in that cell. The output result is organized according to the time dimension and stored as a structured file.

9. A lightning prediction system based on GCN and ConvLSTM, characterized in that, The system includes: A data preprocessing module that uses multi-source meteorological data as multi-source data input and performs preprocessing operations on it. The multi-source meteorological data includes automatic weather station data, radar data, and lightning observation data. Among them, the automatic weather station data and radar data are used to extract the spatio-temporal characteristics of lightning occurrences, and the lightning observation data is used as the true label for supervising model training. An ST-GLNet prediction module. The ST-GLNet prediction module includes an ST-GLNet prediction model. The architecture of the prediction model includes an input layer, an encoder, a fusion module, a decoder, and an output layer. It processes the input data and outputs the lightning prediction result to achieve lightning prediction. The input layer is provided with multiple input channels corresponding to input data of different dimensions, and each input channel is provided with a single encoder to extract the spatio-temporal characteristics of the input data. The encoder includes a GCN layer and a ConvLSTM layer. The GCN layer is used to extract the spatial characteristics of the input data, including introducing spectral graph regularization, defining the energy propagation rule between nodes as the topological connection of horizontally and vertically adjacent units through a physically constrained adjacency matrix construction method, and combining self-loop connection and symmetric normalization processing. It also includes modeling spatio-temporal characteristics at multiple levels by stacking the GCN layer and the ConvLSTM layer through a topology-aware spatio-temporal hierarchical structure. The ConvLSTM layer performs time series modeling on the output data of the GCN layer. Among them, the hidden state update mechanism of the ConvLSTM layer is given physical topology constraints, and the convolutional kernel of the ConvLSTM layer shares the same spatial relationship with the adjacency matrix. And The fusion module fuses the outputs of the encoders in all input channels through a correlation-aware dynamic weight allocation mechanism and a competitive gradient learning strategy as the input of the decoder.

10. A lightning prediction device based on GCN and ConvLSTM, characterized in that, It includes a processor and a memory. A computer program is stored on the memory. When the computer program is executed by the processor, the method according to any one of claims 1-8 is implemented.

Images