Lightning prediction method, system and device based on GCN and ConvLSTM

Abstract

The invention relates to the technical field of meteorological prediction, and discloses a lightning prediction method, system and device based on GCN and ConvLSTM, and the method comprises the steps: S1, employing multi-source meteorological data as multi-source data input, and carrying out the preprocessing of the multi-source data; s2, constructing an ST-GLNet prediction model based on GCN and ConvLSTM, training the ST-GLNet prediction model, processing input data, and outputting a lightning occurrence probability to realize lightning prediction; the input layer is provided with a plurality of input channels respectively corresponding to input data of different dimensions, and each input channel is internally provided with a single encoder for extracting spatial-temporal characteristics of the input data; the GCN layer is used for extracting spatial features of input data, and the ConvLSTM layer performs time sequence modeling on output data of the GCN layer; the output layer outputs one-dimensional rasterized data representing the probability of lightning occurrence. According to the method, the GCN is introduced to extract the spatial features between the grid points and the neighbor nodes, modeling is carried out on the time features of the data in combination with ConvLSTM, and the problems existing in a traditional method are effectively solved through the synergistic effect of ST-GLNet.

Description

Technical Field

[0001] The present application relates to the technical field of meteorological forecasting, and in particular to a lightning forecasting method, system and device based on GCN and ConvLSTM. Background Art

[0002] At present, lightning prediction, as an important meteorological service, has been widely used in aviation, energy, communications and other industries. However, existing prediction technologies still have significant limitations. Mainstream methods mainly rely on numerical weather forecast models and traditional statistical methods based on historical data. These methods predict the probability of lightning by analyzing the historical trends and laws of meteorological data. Nevertheless, due to the limitations in spatial and temporal resolution, the prediction results are often difficult to meet the precise needs of local lightning warnings.

[0003] The occurrence of lightning has strong local and nonlinear characteristics. Traditional methods are difficult to fully capture these complex characteristics, resulting in limited prediction accuracy. For example, although the warning model based on radar meteorological data can provide effective warnings in some cases, the warning success rate is low due to the limitations of radar data (such as insufficient radar cloud information). In addition, traditional numerical weather forecast 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 predictions difficult.

[0004] The rapid accumulation of meteorological observation data provides rich data support for lightning prediction, including automatic weather station (AWS) data, 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 ability to extract spatiotemporal features of meteorological station data and radar data, and it is difficult to utilize their inherent spatial topological relationships and time series dynamic characteristics.

[0005] In recent years, deep learning technology has shown significant advantages in the field of spatiotemporal data processing, especially models such as graph convolutional networks (GCN) and long short-term memory networks (LSTM), which have great potential in processing complex spatiotemporal relationships. However, existing deep learning models still have certain limitations in lightning prediction applications: on the one hand, many models only focus on a single data source, making it difficult to achieve joint modeling of multi-source data; on the other hand, the lack of time series modeling capabilities and the inefficiency of spatial feature extraction make it impossible for the model to fully tap the potential of meteorological data. For example, although LSTM has advantages in processing time series data, its efficiency in spatial feature extraction is 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 spatial dependencies between grid points and neighboring nodes, and the powerful ability of convolutional long short-term memory network (ConvLSTM) in capturing temporal dynamic characteristics and nonlinear temporal relationships. Through the synergy of ST-GLNet, the existing methods can effectively solve the problems of insufficient spatial and temporal resolution, limited spatiotemporal feature extraction capabilities, insufficient multi-source data fusion, and insufficient nonlinear modeling capabilities.

[0007] To achieve the above objectives, the first aspect of the present application provides a lightning prediction method based on GCN and ConvLSTM, comprising the following steps: 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, wherein the automatic weather station data and radar data are used to extract the spatiotemporal characteristics of lightning occurrence, and the lightning observation data is used to supervise the real labels of model training; Step S2, constructing and training a ST-GLNet prediction model based on GCN and ConvLSTM, wherein 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 a lightning prediction result to realize lightning prediction; The input layer is provided with a plurality of input channels corresponding to input data of different dimensions respectively, and each input channel is provided with a single encoder for extracting the spatiotemporal features of the input data; The encoder includes a GCN layer and a ConvLSTM layer, wherein the GCN layer is used to extract spatial features of input data, including introducing spectral 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-circular connection with symmetric normalization processing; and also includes stacking the GCN layer and the ConvLSTM layer through topological perception of the spatiotemporal hierarchy structure to model spatiotemporal features at multiple levels; the ConvLSTM layer performs time series modeling on the output data of the GCN layer, wherein the hidden state update mechanism of the ConvLSTM layer is given physical topological constraints, and the convolution 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 as inputs of the decoder through a correlation-aware dynamic weight allocation mechanism and a competitive gradient learning strategy.

[0008] Optionally, the input layer includes three independent input channels, and the first channel receives a dimension of Radar data, where , are the number of grid divisions in the longitude and latitude directions of the geographic grid data, is the number of features of radar data; the receiving dimension of the second channel is The single-channel lightning event rasterized data of the third channel receives the dimension of Multi-channel automatic weather station rasterized data, The number of features of the automatic weather station rasterized data, the three input channels are processed by independent encoders.

[0009] Optionally, the spectral graph regularization includes performing an operation of pre-constructing a fixed adjacency matrix based on the spatial neighbor relationship of the grid cells, wherein the pre-construction refers to capturing the spatial dependency relationship between grid points by pre-constructing a fixed adjacency matrix based on the spatial neighbor relationship of the grid cells, specifically including: Each grid cell is mapped to a node in the graph, and an adjacency matrix is ​​established through four-way connections. ,in is the number of grid cells, expressed as: ; in, represents the grid spacing, and The corresponding grid node and nodes Coordinates in two-dimensional space; Afterwards, the adjacency matrix is ​​constructed by the node degree matrix Perform symmetric normalization to obtain the standardized adjacency matrix ,in is the identity matrix, ensuring that each node can be connected to itself; Spectral regularization is further enhanced by Laplace smoothing, expressed as: ; in, Indicates Layer node features, Indicates 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 Punishment Layer Node Features With Layer smoothing results deviation.

[0010] Optionally, the encoder uses a cascaded GCN layer and a ConvLSTM layer, including using the topology-aware spatiotemporal hierarchical structure to model spatiotemporal features by stacking three GCN layers and ConvLSTM layers; wherein the GCN layer extracts spatial features of input data as input to the ConvLSTM layer, including receiving three types of inputs, namely radar, lightning observation and AWS data, and extracting the spatiotemporal features of each data respectively. Taking radar data as an example, it is expressed as: ; in, is the time step, is the number of grid nodes, is the characteristic dimension of radar data; The GCN layer extracts the spatial features of grid nodes through two layers of convolution operations. In each layer of GCN, the feature information of adjacent nodes is aggregated, which is expressed as: ; in, Indicates 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, It is the output feature after the GCN operation of this layer.

[0011] Optionally, the hidden state and memory state of each time step are generated and updated through the ConvLSTM layer, including inputting the ConvLSTM layer for time series modeling after obtaining the spatial features, and the ConvLSTM layer updating the hidden state of each time step and memory status , expressed as: ; ; ; ; ; ; in, is the activation function, is the input gate, For the Gate of Forgetfulness, is the output gate, , , and is the bias term, Represents the calculation process of the ConvLSTM unit, is the feature output from the last layer of GCN, and Respectively represent the memory state and hidden state of the current time step, is the adjacency matrix The local receptive field of the convolution kernel is explicitly constrained by the Hadamard product. and Sliding window mode and adjacency matrix The topological structure of Inheriting the state persistence of the physical system, in the spatial dimension through Operations inherit the topological dependencies of the GCN layers.

[0012] Optionally, the fusion module reduces the dimension of the hidden state and memory state output by each encoder through a 1×1 convolution kernel and then splices them along the channel dimension to generate an initial state of the decoder; the fusion module performs the following operations: First, for each encoder output hidden state , and , the dimension is reduced through a 1×1 convolutional layer and mapped to a unified feature space, which is expressed as: ; ; ; in, and Represent the spatial dimensions, is the feature dimension after dimensionality reduction, Corresponding radar data, Corresponding lightning data, Corresponding to AWS data; The hidden state and memory state after dimensionality reduction are fused through the splicing operation along the channel dimension to generate fused features , expressed as: ; in, Represents the concatenation operation on the channel dimension; Afterwards, the relevance-aware dynamic weight allocation mechanism is used to allocate the concatenated features to learnable weight coefficients. Perform weighted summation to generate the final fusion feature , expressed as: ; ; ; in, These are the weights corresponding to radar, lightning and AWS data; During the training process, the inter-modality adaptive feature enhancement is achieved through a competitive gradient learning strategy, including: updating the weights by optimizing the gradient formula , 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.

[0013] Optionally, step S2 also includes using binary cross entropy and Adam optimizer for model training, using binary cross entropy as a loss function to quantify the gap between the predicted value and the true observed value, and using Adam optimizer for parameter update, and adjusting the hyperparameters of the model according to experimental results, wherein the hyperparameters include at least learning rate, convolution kernel size, and number of filters.

[0014] Optionally, the lightning prediction result includes the spatial distribution and temporal distribution of lightning occurrence, and the lightning prediction result is output as one-dimensional rasterized data, the value of each grid cell represents the probability of lightning occurring in the cell, and the output result is organized according to the time dimension and stored as a structured file.

[0015] To achieve the above objectives, the second aspect of the present application provides a lightning prediction system based on GCN and ConvLSTM, the system comprising: A data preprocessing module, which uses multi-source meteorological data as multi-source data input and performs preprocessing operations on it, wherein the multi-source meteorological data includes automatic weather station data, radar data, and lightning observation data, wherein the automatic weather station data and radar data are used to extract the spatiotemporal characteristics of lightning occurrence, and the lightning observation data is used to supervise the real labels of model training; 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, processes the input data and outputs a lightning prediction result to realize lightning prediction; The input layer is provided with a plurality of input channels corresponding to input data of different dimensions respectively, and each input channel is provided with a single encoder for extracting the spatiotemporal features of the input data; The input layer is provided with a plurality of input channels corresponding to input data of different dimensions respectively, and each input channel is provided with a single encoder for extracting the spatiotemporal features of the input data; The encoder includes a GCN layer and a ConvLSTM layer, wherein the GCN layer is used to extract spatial features of input data, including introducing spectral 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-circular connection with symmetric normalization processing; and also includes stacking the GCN layer and the ConvLSTM layer through topological perception of the spatiotemporal hierarchy structure to model spatiotemporal features at multiple levels; the ConvLSTM layer performs time series modeling on the output data of the GCN layer, wherein the hidden state update mechanism of the ConvLSTM layer is given physical topological constraints, and the convolution 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 as inputs of the decoder through a correlation-aware dynamic weight allocation mechanism and a competitive gradient learning strategy.

[0016] To achieve the above-mentioned purpose, the third aspect of the present application provides a lightning prediction device based on GCN and ConvLSTM, including a processor and a memory, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the method described above is implemented.

[0017] After adopting the above technical solution, the present application has the following beneficial effects compared with the prior art: In this application, the adjacency matrix construction mechanism based on the graph convolutional network (GCN) can accurately quantify the spatial dependencies between grid points. Through multi-layer graph convolution 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, this application can effectively reduce the problem of excessive coverage of the alarm area caused by insufficient spatial resolution in traditional predictions by finely identifying the high / low probability distribution areas of lightning, and significantly improve the spatial positioning accuracy. The ST-GLNet model shows stronger time consistency and dynamic adaptability, and can handle lightning events with a long time span. By introducing the convolutional long short-term memory network (ConvLSTM) to construct a temporal memory network, the hidden state and cell state of the time step are maintained, and the time-varying characteristics and nonlinear evolution laws of lightning events are effectively captured. Compared with traditional static prediction methods, this application can dynamically track the three-dimensional evolution process of the lightning system. In the severe convective weather scenario, ST-GLNet shows better time continuity modeling capabilities, and can accurately capture the evolution trajectory of lightning events with a long time span.

[0018] In this application, radar data, automatic weather station (AWS) data and lightning observation data are effectively integrated. Radar data provides powerful spatial background information, automatic weather station (AWS) data provides multi-dimensional meteorological parameters, and lightning observation data is used as real labels for training guidance. By integrating multimodal data, it is possible to explore the key factors of lightning occurrence 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 reliability.

[0019] In this application, the ST-GLNet model, combined with the architecture of GCN and ConvLSTM, can efficiently process meteorological data of different types and dimensions, and has stronger adaptability when facing different spatial distributions, time spans and data sources. This application can flexibly respond to different warning needs, support large-scale data processing and real-time prediction, and show excellent scalability under various meteorological conditions.

[0020] The specific implementation methods of the present application are further described in detail below in conjunction with the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The accompanying drawings are part of this application and are used to provide a further understanding of this application. The illustrative embodiments of this application and their descriptions are used to explain this application, but do not constitute an improper limitation on this application. Obviously, the drawings described below are only some embodiments. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0022] The attached pictures are: Figure 1 It is an overall logic diagram of the lightning prediction method in this specific implementation mode; Figure 2 is a logical schematic diagram of the prediction method based on ST-GLNet in this specific implementation mode; Figure 3 It is a logical schematic diagram of the ST-GLNet prediction model in this specific implementation mode; Figure 4 Schematic diagram of the prediction result of the ST-GLNet prediction model at time t in this specific implementation method; Figure 5 This is a schematic diagram of recording the actual occurrence of lightning at time t in this specific implementation mode; Figure 6 is a schematic diagram comparing the average indicators of ST-GLNet and ConvLSTM using meteorological observation data (AWS) and lightning observation data (LIG) in this specific implementation; Figure 7 It is a schematic diagram comparing the average indicators of meteorological observation data (AWS), lightning observation data (LIG) and radar data (Radar) used by ST-GLNet and ConvLSTM in this specific implementation. DETAILED DESCRIPTION

[0023] In order to make the purpose, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments will be clearly and completely described below in conjunction with the 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.

[0024] In view of the unique spatial energy propagation characteristics and multi-source data heterogeneity of lightning prediction tasks, a systematic innovation has been made to the combined architecture of traditional GCN and ConvLSTM. Figure 1 , the present application provides a lightning prediction method based on GCN and ConvLSTM, comprising the following steps: 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, wherein the automatic weather station data and radar data are used to extract the spatiotemporal characteristics of lightning occurrence, and the lightning observation data is used to supervise the real label of model training; Step S2, constructing and training a 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. The input data is processed and the lightning prediction result is output to realize lightning prediction. The input layer is provided with multiple input channels corresponding to input data of different dimensions, and a single encoder is provided in each input channel to extract the spatiotemporal features of the input data; the encoder includes a GCN layer and a ConvLSTM layer, 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 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 a predicted lightning distribution map; the output layer splices the output of the decoder through the time dimension to generate the final prediction sequence, and outputs one-dimensional rasterized data representing the probability of lightning occurrence.

[0025] Specifically, in terms of spatial feature modeling, the energy propagation rules between nodes are defined as topological connections of horizontally and vertically adjacent units through the construction method of physically constrained adjacency matrix. The conservation characteristics of atmospheric electrostatic field energy are explicitly simulated by combining self-circular connections with symmetric normalization processing. In temporal dynamic 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 spatiotemporal evolution conforms to the dynamic laws 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 adaptive feature enhancement between modalities is achieved through a competitive gradient learning strategy, which effectively suppresses noise interference while enhancing lightning-related high-value signals.

[0026] Specifically, the spectral regularization constraint is introduced in the model training stage, and the energy distribution of the physical topology is aligned through the feature propagation process, which significantly improves the model's ability to express the characteristics of lightning energy transfer. The decoder adopts a hybrid upsampling strategy that combines the deconvolution kernel with the interpolation algorithm to maintain multimodal long-range dependencies during the spatial resolution reconstruction process. The final output lightning probability field fully describes the charge transport process through time series stacking, forming an interpretable prediction result that conforms to the electrical characteristics of the atmosphere.

[0027] It should be noted that the execution subject of the lightning prediction method in this embodiment is an ST-GLNet prediction device based on a graph convolutional network (GCN) and a convolutional long short-term memory network (ConvLSTM), which 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 PDA, a vehicle-mounted electronic device, a wearable device, etc., and the non-mobile electronic device can be a server and a personal computer, etc., which is not specifically limited in this application. The following takes the execution subject as an example of a server to describe the ST-GLNet prediction method based on a graph convolutional network (GCN) and a convolutional long short-term memory network (ConvLSTM) in this embodiment.

[0028] In addition, the terms "first" and "second" are used for descriptive purposes only and should not be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of the embodiments of the present application, the meaning of "multiple" is two or more, unless otherwise clearly and specifically defined.

[0029] In one possible implementation, the input layer includes three independent input channels, the first channel receiving dimension is Radar data, where is the number of features of radar data; the receiving dimension of the second channel is The rasterized data of lightning observation events with a single channel (i.e., the lightning observation data has only one feature); the third channel receives the dimension The multi-channel (i.e., the automatic weather station contains multiple features) automatic weather station (AWS) rasterized data, where The number of features of the automatic weather station rasterized data, the three input channels are processed by independent encoders.

[0030] In one feasible implementation, the first channel is a radar feature channel: receiving a normalized radar reflectivity data set, the input dimension is ,in , are the number of grid divisions in the longitude and latitude directions of the geographic grid data, Characterize the characteristic dimension of radar data; the second channel is the lightning observation channel: input single-channel lightning event rasterized observation data (LIG), the dimension is , each grid cell value represents the binary mark of lightning events; the third channel is the meteorological data channel: it processes the rasterized fusion data of multi-source meteorological station parameters, and the input dimension is ,in is the number of meteorological features of the Automatic Weather Station (AWS).

[0031] In this embodiment, radar data, AWS data and lightning observation data are effectively integrated. Radar data provides powerful spatial background information, AWS data provides multi-dimensional meteorological parameters, and lightning observation data is used as a real label for training guidance. By integrating multimodal data, the model can mine the key factors of lightning occurrence from multiple dimensions and effectively improve the prediction accuracy of lightning occurrence areas.

[0032] In another feasible implementation, for AWS data, each weather station has only one fixed location, and the number is far from enough to fill each grid in the grid map. Therefore, in view of 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 station data association, interpolation calculation, to multi-threaded optimization, result output and file locking mechanism, and final format conversion.

[0033] Specifically, the preprocessing of AWS data in step S1 includes the following steps: 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; 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; 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; 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; 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; 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; 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.

[0034] 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.

[0035] 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.

[0036] In one feasible implementation, the preprocessing format conversion step converts the cleaned and interpolated CSV format data into a .npy format that can be directly read by the model, thereby improving the model operation efficiency and reducing the data reading overhead.

[0037] Specifically, the CSV format data is converted into a .npy format that can be directly read through the format conversion module, which specifically includes the following steps: Step S160, reading the CSV files with timestamps from the pre-processed CSV file directory in chronological order, and the file names follow the "YYYY-MM-DD-HH" format; Step S161, extracting the key meteorological parameter data fields of temperature, humidity, wind speed, and air pressure in the CSV file, parsing the file name to obtain date and time information and renaming it; Step S162: reshape the one-dimensional data sequence of each meteorological parameter into a two-dimensional matrix according to the longitude dimension and latitude dimension of the preset grid map to generate a longitude matrix of the grid map. He Kuan The formed two-dimensional array; Step S163, creating an independent storage directory for each meteorological parameter, and storing the reshaped two-dimensional array into .npy format files according to parameter categories; Step S164, using a hierarchical path structure to organize the output data.

[0038] Specifically, during the conversion process, the station data files are read one by one from the CSV folder storing the preprocessing results, the key meteorological parameters are extracted and processed, and the date and time information is extracted and renamed according to the file name. The meteorological parameter data in the one-dimensional array format is reshaped into a raster map with a length of and width The two-dimensional array formed matches the spatial grid structure required by the model input. The reshaped data are saved as .npy files, and each meteorological parameter is stored in a separate folder. The file naming uses a unified time mark format to ensure that the data is clear and orderly. The storage path of the output data is designed in a hierarchical manner, and different meteorological parameter data are stored in separate folders to improve the automation of data conversion, ensure clear data management, and avoid file confusion or overwriting risks.

[0039] In another feasible implementation, in order to solve the problem of insufficient radar data processing in existing lightning prediction methods, this embodiment realizes the efficient combination of radar data with other data sources through steps such as data reading and integration, cleaning, association, interpolation and fusion, thereby providing high-quality input data for the model.

[0040] Specifically, the preprocessing of radar data in step S1 includes the following steps: Step S100, reading radar data through a data integration module, extracting latitude and longitude grids, meteorological variables such as CREF, ET, VIL and scanning time, where CREF is the combined reflectivity, ET is the echo top height, and VIL is the vertically integrated liquid water content, and constructing a radar data cube of the time series; Step S110, sorting multiple radar data within the same hour by timestamp, performing grid-level value accumulation and file number statistics on each meteorological variable, and generating a spatiotemporal average data array; Step S120, perform data cleaning operations, use neighborhood interpolation to fill in missing grid points, identify and correct abnormal values ​​that exceed the preset range through the dynamic threshold method, compare multiple file data within the same hour, detect data collection or transmission errors, and when a file data is significantly different from other files, mark it as abnormal and remove it; Step S130, establishing a spatial correlation model, mapping the latitude and longitude of the automatic weather station site to the center point of the radar grid, and constructing a site-grid matching relationship matrix; Step S140, using an adaptive Kriging interpolation algorithm to interpolate the meteorological parameters of the automatic weather station, such as wind speed, temperature, relative humidity, etc., to the radar grid, and when the radar resolution is higher than that of the automatic weather station, downsampling matching is performed to ensure smooth fusion, and three-dimensional interpolation repair based on the continuity of the meteorological field is performed on the area where radar data is missing; Step S150, outputting the fused multi-dimensional data tensor, including radar reflectivity, AWS interpolation parameters and time coding features.

[0041] Specifically, the preprocessing of radar data starts with file reading and time integration. The latitude and longitude, data variables (such as CREF, ET, VIL) and time attributes are extracted 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 value, 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 spring steel plates are repaired, data quality is improved by interpolation or elimination of missing data, and files with abnormal data are eliminated to ensure consistency.

[0042] After AWS, LIG, and Radar finish extracting spatiotemporal features, the hidden states of each encoder output and memory status Perform 1×1 convolution dimensionality reduction, and splice the reduced state along the last dimension into fusion features. In the multi-data fusion stage, this embodiment uses the latitude and longitude information of the grid center point of the radar data to match the AWS site location, and interpolates the AWS data to 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 fusion accuracy. At the same time, for the locally missing areas in the radar data, interpolation filling is used to repair them to ensure the continuity and integrity of the spatial data.

[0043] In one feasible implementation, the preprocessing of lightning observation data in step S1 includes the following steps: Step S1000, receiving a lightning observation data file and a grid data file, wherein the lightning observation data includes the latitude and longitude coordinates and timestamp of each lightning event, and the grid data includes the latitude and longitude of each grid center point and a unique number; Step S1100, using the CUDA acceleration technology parallel computing framework to batch convert the longitude and latitude of the center points of the grid into radian tensor data, and at the same time load the longitude and latitude data in the lightning observation data; Step S1200, using the haversine formula to perform parallel spherical distance calculation between each lightning event and all grid center points, the calculation formula is: ; 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; Indicates the latitude difference between two points; and Indicates the longitude of two points, expressed in radians; Represents the longitude difference between two points; Step S1300, setting a matching threshold radius, marking the nearest grid number through CUDA atomic operations, and recording a special mark for unmatched events; Step S1400, outputting a lightning event dataset with raster coding, and storing it in CSV format in order of occurrence time; Step S1500, constructing a timestamp sequence, grouping and aggregating lightning events by "year, month, day, hour" to generate a binary grid matrix, setting the grid position where the event occurred to 1, and the rest to 0; Step S1600, the converted binary grid is stored as an independent file in units of hours. The file name adopts a standard time format and the content is a binary value stored line by line, indicating the position status of the grid.

[0044] Specifically, the lightning observation data is calculated using the CUDA-accelerated Haversine formula to calculate the spherical distance from each event to the center of the grid, and the event is matched to the nearest grid number. Finally, a binary grid file stored hourly is generated to indicate the occurrence status of the lightning event in each grid unit (1 for occurrence, 0 for non-occurrence). All data are organized in time series and stored in .npy format, providing structured and high-quality input support.

[0045] It should be noted that this embodiment converts lightning observation data from CSV format to binary raster 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 every hour, providing a solid foundation for the temporal and spatial modeling of the model. The preprocessing method of this embodiment successfully maps lightning observation data from a single event form to a spatial grid unit, improves computing efficiency with the help of CUDA acceleration technology, ensures the spatial accuracy of the data, and lays a solid foundation for the spatiotemporal modeling of large-scale lightning events.

[0046] To address the essential mismatch between grid isolation processing and physically driven energy propagation dynamics, we propose an innovative fusion of spectral regularization and topology-aware spatiotemporal hierarchy. Specifically, spectral regularization involves performing operations to pre-construct a fixed adjacency matrix based on the spatial adjacency of grid cells, where pre-construction means capturing the spatial dependencies between grid points by pre-constructing a fixed adjacency matrix based on the spatial adjacency of grid cells, including: Each grid cell is mapped to a node in the graph, and an adjacency matrix is ​​established through four-way connections. ,in is the number of grid cells, expressed as: ; in, represents the grid spacing, and The corresponding grid node and nodes Coordinates in two-dimensional space; Afterwards, in order to maintain stability and effectiveness in the graph convolution operation, the adjacency matrix is ​​transformed into a node degree matrix Perform symmetric normalization to obtain the standardized adjacency matrix ,in is the identity matrix, ensuring that each node can be connected to itself; Spectral regularization is further enhanced by Laplace smoothing, expressed as: ; in, Indicates Layer node features, Indicates 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, and the regularization term is the auxiliary loss term Punishment Layer Node Features With Layer smoothing result (i.e. spectral convolution mode) deviation.

[0047] Execute 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 a fixed adjacency matrix initialization structure with four-way connections between nodes; symmetric normalization of the fixed adjacency matrix is ​​performed through the node degree matrix to form a stable neighborhood propagation constraint condition for graph convolution; Two layers of graph convolution operations are performed to capture the spatial dependencies between grid points and neighboring nodes, thus achieving local and global feature extraction.

[0048] 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 gridded spatial relationship of the input data. Each grid cell is regarded as a node in the graph, and only establishes connections with its upper, lower, left and right neighbors to form a regular grid graph. The fixed adjacency matrix refers to the adjacency matrix pre-constructed according to the spatial adjacency relationship of the grid cells in the preprocessing stage, and this matrix remains unchanged in subsequent graph convolution operations. In order to avoid the instability caused by the uneven connection structure in the graph convolution operation, this embodiment further normalizes the fixed adjacency matrix. Specifically, by calculating the degree matrix of each node, the fixed adjacency matrix is ​​normalized by an inverse square root operation, thereby improving the stability of the graph convolution operation and the accuracy of the numerical calculation. Using this fixed adjacency matrix, GCN can efficiently capture the spatial dependency and local correlation between grid points and neighboring nodes, providing strong support for the high-resolution prediction of the model.

[0049] In an achievable implementation, the encoder uses a cascaded GCN layer and a ConvLSTM layer, including a topology-aware spatiotemporal hierarchy, and models spatiotemporal features by stacking three GCN layers and ConvLSTM layers; the goal is to integrate spatial dependencies through a graph convolutional network (GCN), and to capture temporal dependencies through ConvLSTM, so as to achieve lightning event prediction that evolves 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 spatiotemporal features of each data respectively. Taking radar data as an example, it is expressed as: ; in, is the time step, is the number of grid nodes, is the characteristic dimension of radar data; The GCN layer extracts the spatial features of grid nodes through two layers of convolution operations. In each layer of GCN, the feature information of adjacent nodes is aggregated, which is expressed as: ; in, Indicates 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, It is the output feature after the GCN operation of this layer.

[0050] In one achievable implementation, 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 the ConvLSTM layer for time series modeling, and the ConvLSTM layer updating the hidden state of each time step and memory status , ConvLSTM can capture the dynamic evolution over time and establish the temporal characteristics of the grid points changing over time, which can be expressed as: ; ; ; ; ; ; in, is the activation function, is the input gate, For the Gate of Forgetfulness, is the output gate, , , and is the bias term, Represents the calculation process of the ConvLSTM unit, is the feature output from the last layer of GCN, and Respectively represent the memory state and hidden state of the current time step, is the adjacency matrix The local receptive field of the convolution kernel is explicitly constrained by the Hadamard product. and Sliding window mode and adjacency matrix The topological structure of Inheriting the state persistence of the physical system, in the spatial dimension through Operations inherit the topological dependencies of the GCN layers.

[0051] In an achievable implementation, the fusion module reduces the dimension of the hidden state and memory state output by each encoder through a 1×1 convolution kernel and then splices them along the channel dimension to generate the initial state of the decoder; the fusion module performs the following operations: First, for each encoder output hidden state , and , the dimension is reduced through a 1×1 convolutional layer and mapped to a unified feature space, which is expressed as: ; ; ; in, and Represent the spatial dimensions, is the feature dimension after dimensionality reduction, Corresponding radar data, Corresponding lightning data, Corresponding to AWS data; through this dimensionality reduction operation, it is ensured that features from different modalities can be effectively fused in the same spatial dimension; The hidden state and memory state after dimensionality reduction are fused through the splicing operation along the channel dimension to generate fused features , expressed as: ; in, Represents the splicing operation on the channel dimension; the fusion feature after splicing For one A tensor of containing joint information from data sources of different modalities; Afterwards, a dynamic weight allocation mechanism based on correlation perception is used to allocate the concatenated features through learnable weight coefficients. Perform weighted summation to generate the final fusion feature , the features of each modality are assigned a dynamically adjusted weight so that they can be adaptively fused according to the relevance of each modality, expressed as: ; ; ; in, These are the weights for radar, lightning, and AWS data respectively. This weighted fusion strategy ensures that the mutual influence between different data sources is effectively integrated, and also ensures that the dynamic adjustment of weights can adapt to the diversity of data in different scenarios. During the training process, the inter-modality adaptive feature enhancement is achieved through a competitive gradient learning strategy, including: updating the weights by optimizing the gradient formula , 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; this gradient update formula encourages competitive learning between different modalities - when the features of a certain modality When the features of other modalities are highly correlated, the weight of this modality will increase. At the same time, Regularization term This prevents a single mode from dominating the entire model.

[0052] In addition, the memory state C is responsible for preserving long-term memory, mainly capturing slowly changing features in the time series, such as the state of the environmental background. These states are more stable, and the memories of different modes are more complementary rather than competitive. Therefore, for the memory state , which are integrated via a connection-preserving projection, expressed as: .

[0053] In an achievable implementation, the decoder is designed as a resolution recovery network, specifically comprising: Temporal convolution unit: 32 3×3 convolution kernels are used to compress the spatiotemporal feature dimensions to ; Dynamic reconstruction layer: The ConvLSTM layer maintains 32 layers to iteratively update the state of the time series features to maintain the temporal coherence of the results; Spatial resolution restoration component: Perform spatial upsampling through a deconvolution layer with a stride of 2×2 to gradually restore the grid dimensions to the original input resolution.

[0054] In one feasible implementation, the fusion module combines the hidden states ( ) and memory state ( ) is concatenated along the channel dimension after dimensionality reduction through a 1×1 convolution kernel to generate the initial state of the decoder. 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 the following operations: Perform 1×1 convolution dimensionality reduction on the hidden state and memory state output by each encoder; Concatenate the reduced-dimensional state along the last dimension into fusion features; Downsampling convolution and fully connected layers are used to generate the initial state of the decoder.

[0055] In another achievable embodiment, the decoder includes a cascade structure of a time-distributed convolution layer, a ConvLSTM layer, and a deconvolution layer, and the decoder is configured to gradually restore the spatial resolution and generate a predicted lightning distribution map; the decoder performs the following steps: The time-distributed convolution layer applies convolution operations to the input features in the time dimension to extract spatial features in the time series; The ConvLSTM layer captures the dynamic changes in the time series; Through multi-layer deconvolution operations, the feature maps are gradually upsampled to restore high-resolution spatial information.

[0056] It should be noted that Figure 2 middle , , …, Represents time series input data, generally refers to the meteorological observations from frame 1 to frame t-1 (or from moment 1 to moment t-1). The hidden layer refers to the grid data (denoted as X) at a certain moment or input channel. 1 ), after a certain preprocessing or convolution operation, the input data will be sent to the next step (hidden layer / encoder) for calculation. Y represents the final "lightning probability distribution", and σ represents the use of Sigmoid function to map the data to the range of (0, 1).

[0057] Figure 3 middle represents the spatiotemporal data input into the model at time t; and are the cell state and memory state output by the X data encoder respectively. They represent the hidden states of the encoder outputs of radar data, lightning observation data, and AWS observation data respectively; They represent the memory states output by the encoders of radar data, lightning observation data, and AWS observation data respectively; C and H represent the memory state and hidden state after fusion. Represents the predicted output.

[0058] See also Figure 3 In an achievable implementation, the decoder part supports an initialization optimization strategy, which optimizes the distribution characteristics of the input data by performing a power transformation and Sigmoid function normalization on the input data. 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 of (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 the training.

[0059] See also Figure 3 , Figure 4 and Figure 5 ,In actual applications, the output layer performs the Cropping3D operation to ensure the ,output size matches, and crops the feature map of the final output to ,ensure that the output size is consistent with the input data; the prediction results output by the ,decoder are concatenated in the time dimension to generate a complete prediction sequence, which is ,output as one-dimensional rasterized data representing the probability distribution of lightning ,occurrence. Figure 4 The numbers in represent the probability values. Figure 4 and Figure 5 The horizontal axis represents east longitude and the vertical axis represents north latitude, and the unit is degree.

[0060] Specifically, the output layer performs result optimization operations, including: Boundary alignment unit: Cropping3D layer is used to crop redundant feature boundaries to ensure that the output geographic raster size is consistent with the input data; Probability conversion module: After being activated by the Sigmoid function, a one-dimensional lightning probability grid data matrix is ​​generated, and each element represents the probability value of lightning occurrence in the corresponding geographic unit within the time series window.

[0061] In one practicable implementation, step S2 further includes prediction model training and optimization; Binary cross entropy and Adam optimizer are used for model training. Binary cross entropy is used as the loss function to quantify the gap between the predicted value and the true observed value. Adam optimizer is used for parameter update. The hyperparameters of the model are adjusted according to the experimental results. The hyperparameters include at least the learning rate, convolution kernel size and the number of filters.

[0062] In another achievable implementation, the lightning prediction result is output and applied; The lightning prediction results include the spatial distribution and temporal distribution of lightning occurrence. The lightning prediction results are output as one-dimensional rasterized data. The value of each grid cell represents the probability of lightning occurring in that cell. The output results are organized according to the time dimension and stored as structured files.

[0063] In order 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 compared with the classic model ConvLSTM. The evaluation indicators include threat score (TS), equal-weighted threat score (ETS), detection rate (POD), false alarm rate (FAR), mean square error ratio (MAR), bias score (BS) and accuracy (AC). For specific experimental results, please refer to Figure 6 and Figure 7 .

[0064] See also Figure 6, using meteorological observation data (AWS) and lightning observation data (LIG). Experimental results show that ST-GLNet outperforms ConvLSTM in most evaluation indicators. Specifically, ST-GLNet achieved 0.3043 in threat score (TS), which is 15.7% higher than 0.2629 of ConvLSTM. The detection rate (POD) was 0.7504, which is significantly better than 0.7066 of ConvLSTM. In terms of accuracy (AC), ST-GLNet is slightly higher than ConvLSTM, which are 0.9717 and 0.9653 respectively. In addition, ST-GLNet performs better in equal-weighted threat score (ETS), and slightly better in false alarm rate (FAR) and mean square error ratio (MAR) indicators, but the difference with ConvLSTM is small.

[0065] See also Figure 7 , using meteorological observation data (AWS) and lightning observation data (LIG) and radar data (Radar). Experimental results show that ST-GLNet also demonstrates excellent performance. Its threat score (TS) is 0.2531, which is 8.5% higher than ConvLSTM's 0.2332. The detection rate (POD) reaches 0.6611, which is higher than ConvLSTM's 0.5724, indicating that ST-GLNet has strong detection capabilities. In terms of accuracy (AC), ST-GLNet and ConvLSTM perform similarly, at 0.9755 and 0.9778 respectively, but ST-GLNet has better comprehensive performance in other indicators.

[0066] The ST-GLNet model outperforms the ConvLSTM model on different data sets, especially in key indicators such as threat score and detection rate. This shows that ST-GLNet can more effectively capture spatiotemporal features, improve prediction accuracy, while maintaining high generalization ability, and provide a better solution for modeling complex scenarios.

[0067] Based on the same inventive concept, the present application also provides a lightning prediction system based on GCN and ConvLSTM, the system comprising: The data preprocessing module 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. The automatic weather station data and radar data are used to extract the spatiotemporal characteristics of lightning occurrence, and the lightning observation data is used to supervise the real labels of model training; ST-GLNet prediction module, the ST-GLNet prediction module includes the 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 results to realize 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 spatiotemporal features of the input data; 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 the introduction of spectral regularization, the physical constraint adjacency matrix construction method, the definition of the energy propagation rule between nodes as the topological connection of horizontally and vertically adjacent units, and the combination of self-circular connection and symmetric normalization processing; it also includes topological perception of the spatiotemporal hierarchy, stacking the GCN layer and the ConvLSTM layer, and modeling the spatiotemporal features at multiple levels; the ConvLSTM layer performs time series modeling on the output data of the GCN layer, wherein the hidden state update mechanism of the ConvLSTM layer is given physical topological constraints, and the convolution 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 as the input of the decoder through a relevance-aware dynamic weight assignment mechanism and a competitive gradient learning strategy.

[0068] 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, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the above method is implemented.

[0069] The program product of the present application for implementing the above method may adopt a portable compact disk read-only memory and include program code, and may be 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 may be any tangible medium containing or storing a program, which may be used by or in combination with an instruction execution system, device or device.

[0070] It should be noted that a computer-readable storage medium may include a data signal propagated in a baseband or as part of a carrier wave, which carries a readable program code. Such propagated data signals may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. The readable storage medium may also be any readable medium other than a readable storage medium, which may send, propagate, or transmit a program for use by or in conjunction with an instruction execution system, device, or device. The program code contained on the readable storage medium may be transmitted using any appropriate medium, including but not limited to wireless, wired, optical cable, RF, etc., or any suitable combination of the above.

[0071] The above are only preferred embodiments of the present application, and are not intended to limit the present application in any form. Although the present application has been disclosed as above with preferred embodiments, it is not intended to limit the present application. Any technician familiar with the present application can make some changes or modifications to equivalent embodiments of equivalent changes using the above-mentioned technical contents without departing from 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, any simple modifications, equivalent changes and modifications made to the above embodiments based on the technical essence of the present application without departing from the content of the technical solution of the present application are still within the scope of the solution of the present application.

Claims

1. A lightning prediction method based on GCN and ConvLSTM, characterized in that: The following steps are involved: 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, wherein the automatic weather station data and radar data are used to extract the spatiotemporal characteristics of lightning occurrence, and the lightning observation data is used to supervise the real labels of model training; Step S2, constructing and training a ST-GLNet prediction model based on GCN and ConvLSTM, wherein 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 a lightning prediction result to realize lightning prediction; The input layer is provided with a plurality of input channels corresponding to input data of different dimensions respectively, and each input channel is provided with a single encoder for extracting the spatiotemporal features of the input data; The encoder includes a GCN layer and a ConvLSTM layer, wherein the GCN layer is used to extract spatial features of input data, including introducing spectral 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-circular connection with symmetric normalization processing; and also includes stacking the GCN layer and the ConvLSTM layer through topological perception of the spatiotemporal hierarchical structure to model spatiotemporal features at multiple levels; The ConvLSTM layer performs time series modeling on the output data of the GCN layer, wherein the hidden state update mechanism of the ConvLSTM layer is given a physical topology constraint, and the convolution kernel of the ConvLSTM layer shares the same spatial relationship with the adjacency matrix; as well as The fusion module fuses the outputs of the encoders in all input channels as inputs of the decoder through a correlation-aware dynamic weight allocation mechanism and a competitive gradient learning strategy.

2. The method according to claim 1, characterized in that The input layer includes three independent input channels. The first channel receives the dimension Radar data, where , are the number of grid divisions in the longitude and latitude directions of the geographic grid data, is the number of features of radar data; The receiving dimension of the second channel is The single-channel lightning event rasterized data of the third channel receives the dimension of Multi-channel automatic weather station rasterized data, The number of features of the automatic weather station rasterized data, the three input channels are processed by 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 neighbor relationship of the grid cells, wherein the pre-construction refers to capturing the spatial dependency relationship between grid points by pre-constructing a fixed adjacency matrix based on the spatial neighbor relationship of the grid cells, specifically including: Each grid cell is mapped to a node in the graph, and an adjacency matrix is ​​established through four-way connections. ,in is the number of grid cells, expressed as: ； in, represents the grid spacing, and Corresponding grid nodes and nodes Coordinates in two-dimensional space; Afterwards, the adjacency matrix is ​​constructed by the node degree matrix Perform symmetric normalization to obtain the standardized adjacency matrix ,in is the identity matrix, ensuring that each node can be connected to itself; Spectral regularization is further enhanced by Laplace smoothing, expressed as: ； in, Indicates Layer node features, Indicates 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 Punishment Layer Node Features With Layer smoothing results deviation.

4. The method according to claim 1, characterized in that The encoder uses a cascaded GCN layer and a ConvLSTM layer, including using the topologically aware spatiotemporal hierarchical structure to model spatiotemporal 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, namely radar, lightning observation and AWS data, and extracting the spatiotemporal features of each data respectively. Taking radar data as an example, it is expressed as: ； in, is the time step, is the number of grid nodes, is the characteristic dimension of radar data; The GCN layer extracts the spatial features of grid nodes through two layers of convolution operations. In each layer of GCN, the feature information of adjacent nodes is aggregated, which is expressed as: ； in, Indicates 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, It is the output feature after the GCN operation of this layer.

5. The method according to claim 4, characterized in that The hidden state and memory state of each time step are generated and updated through the ConvLSTM layer, including inputting the ConvLSTM layer for time series modeling after obtaining the spatial features, and the ConvLSTM layer updating the hidden state of each time step and memory status , expressed as: ； ； ； ； ； ； in, is the activation function, is the input gate, For the Gate of Forgetfulness, is the output gate, , , and is the bias term, Represents the calculation process of the ConvLSTM unit, is the feature output from the last layer of GCN, and Respectively represent the memory state and hidden state of the current time step, is the adjacency matrix The local receptive field of the convolution kernel is explicitly constrained by the Hadamard product. and Sliding window mode and adjacency matrix The topological structure of Inheriting the state persistence of the physical system, in the spatial dimension through Operations inherit the topological dependencies of the GCN layers.

6. The method according to claim 5, characterized in that The fusion module reduces the dimension of the hidden state and memory state output by each encoder through a 1×1 convolution kernel and then concatenates them along the channel dimension to generate the initial state of the decoder; The fusion module performs the following operations: First, for each encoder output hidden state , and , the dimension is reduced through a 1×1 convolutional layer and mapped to a unified feature space, which is expressed as: ； ； ； in, and Represent the spatial dimensions, is the feature dimension after dimensionality reduction, Corresponding radar data, Corresponding lightning data, Corresponding to AWS data; The hidden state and memory state after dimensionality reduction are fused through the splicing operation along the channel dimension to generate fused features , expressed as: ； in, Represents the concatenation operation on the channel dimension; Afterwards, the relevance-aware dynamic weight allocation mechanism is used to allocate the concatenated features to learnable weight coefficients. Perform weighted summation to generate the final fusion feature , expressed as: ； ； ； in, These are the weights corresponding to radar, lightning and AWS data; During the training process, the inter-modality adaptive feature enhancement is achieved through a competitive gradient learning strategy, including: updating the weights by optimizing the gradient formula , 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. The method according to claim 1, characterized in that Step S2 also includes using binary cross entropy and Adam optimizer for model training, using binary cross entropy as a loss function to quantify the gap between the predicted value and the true observed value, and using Adam optimizer for parameter update, and adjusting the hyperparameters of the model according to the experimental results, wherein the hyperparameters include at least the learning rate, the convolution kernel size, and the number of filters.

8. The method according to claim 1, characterized in that The lightning prediction result includes the spatial distribution and time distribution of lightning occurrence. The lightning prediction result is output as one-dimensional grid data. The value of each grid cell represents the probability of lightning occurrence in the 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 comprises: A data preprocessing module, which uses multi-source meteorological data as multi-source data input and performs preprocessing operations on it, wherein the multi-source meteorological data includes automatic weather station data, radar data, and lightning observation data, wherein the automatic weather station data and radar data are used to extract the spatiotemporal characteristics of lightning occurrence, and the lightning observation data is used to supervise the real labels of model training; 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, processes the input data and outputs a lightning prediction result to realize lightning prediction; The input layer is provided with a plurality of input channels corresponding to input data of different dimensions respectively, and each input channel is provided with a single encoder for extracting the spatiotemporal features of the input data; The encoder includes a GCN layer and a ConvLSTM layer, wherein the GCN layer is used to extract spatial features of input data, including introducing spectral 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-circular connection with symmetric normalization processing; and also includes stacking the GCN layer and the ConvLSTM layer through topological perception of the spatiotemporal hierarchy structure to model spatiotemporal features at multiple levels; the ConvLSTM layer performs time series modeling on the output data of the GCN layer, wherein the hidden state update mechanism of the ConvLSTM layer is given physical topological constraints, and the convolution 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 as inputs of the decoder through a correlation-aware dynamic weight allocation mechanism and a competitive gradient learning strategy.

10. A lightning prediction device based on GCN and ConvLSTM, characterized in that: The method comprises a processor and a memory, wherein a computer program is stored in the memory, and when the computer program is executed by the processor, the method according to any one of claims 1 to 8 is implemented.

Images