A method and system for predicting lightning approaching trends

Abstract

This invention discloses a method and system for near-term lightning trend forecasting. The method includes: acquiring the spatial distribution and time series of cloud cover; extracting a first cloud cover texture feature vector; generating a layered cloud cover feature matrix; reducing the dimensionality of the first cloud cover texture feature vector to generate a second cloud cover texture feature vector; combining the layered cloud cover feature matrix with clustering to generate a regionalized cloud cover distribution feature set; extracting humidity, temperature, and wind field data to generate a lightning triggering factor dataset; analyzing the regionalized cloud cover distribution feature set using time series analysis and optical flow methods to form a dynamic cloud cover feature set; and generating a lightning intensity level classification feature set based on the dynamic cloud cover feature set and the layered cloud cover feature matrix, combined with pre-established lightning-related cloud type labels. This invention can significantly improve the accuracy and timeliness of lightning prediction through deep fusion and dynamic optimization of spatiotemporal features, providing efficient technical support for meteorological disaster early warning.

Description

Technical Field

[0001] This invention relates to the field of lightning now-near trend forecasting technology, and in particular to a lightning now-near trend forecasting method and system. Background Technology

[0002] Lightning nowcasting is a crucial research area in meteorological forecasting, playing a vital role in ensuring the safe operation of industries such as aviation, power, and transportation. Accurate lightning activity prediction can effectively reduce disaster losses and improve socio-economic benefits. However, current lightning forecasting methods have significant limitations when dealing with complex meteorological scenarios. Existing technologies often struggle to adapt to the diversity of cloud cover distribution across different regions, especially when handling coastal high-humidity cloud clusters and sparse inland cloud systems. The lack of precise capture of dynamic changes in cloud cover leads to insufficient timeliness and reliability of forecast results. Furthermore, existing methods suffer from bottlenecks in processing massive amounts of meteorological data, making it difficult to quickly extract features directly related to lightning activity, thus limiting the level of forecast precision.

[0003] The core challenge lies in extracting key dynamic features from the complex and ever-changing cloud cover distribution to achieve accurate predictions of lightning activity trends. First, the spatiotemporal characteristics of cloud cover distribution are complex. Coastal areas have dense and rapidly evolving cloud clusters, while inland areas have sparse and uneven cloud systems. This difference makes single feature extraction methods difficult to apply. For example, in coastal areas, a lightning event might occur due to the rapid accumulation of high-humidity clouds, but existing systems struggle to identify such rapidly changing patterns from massive amounts of cloud cover data in a short time. Second, due to the complexity of cloud cover dynamic characteristics, existing methods lack universality under different meteorological conditions, making it difficult to accurately distinguish the triggering conditions of lightning activity using a unified framework. This often leads to classification bias or inaccurate prediction time windows when facing regional differences in forecasting systems. For example, a coastal city might miss the optimal warning window due to rapid cloud movement, while inland areas experience frequent false alarms due to sparse cloud systems.

[0004] Therefore, how to efficiently analyze the spatiotemporal characteristics of cloud cover changes and construct adaptive feature extraction and classification methods for various cloud scenarios has become a key issue in lightning near-term trend forecasting. Summary of the Invention

[0005] In view of this, the purpose of this invention is to propose a method for predicting the imminent trend of lightning, which can significantly improve the accuracy and timeliness of lightning prediction through deep fusion and dynamic optimization of spatiotemporal features, and provide efficient technical support for meteorological disaster early warning.

[0006] According to one aspect of the present invention, a method for forecasting the imminent trend of lightning is provided, the method comprising:

[0007] Obtain the spatial distribution and time series of cloud cover, extract the first cloud cover texture feature vector, and generate a cloud cover hierarchical feature matrix; reduce the dimensionality of the first cloud cover texture feature vector to generate the second cloud cover texture feature vector, and combine the cloud cover hierarchical feature matrix to generate a regionalized cloud cover distribution feature set through clustering.

[0008] Humidity, temperature, and wind field data are extracted to generate a lightning triggering factor dataset. Time series analysis and optical flow methods are used to analyze the regional cloud cover distribution characteristics set and form a dynamic cloud cover characteristic set.

[0009] Based on the dynamic cloud cover feature set and the cloud cover hierarchical feature matrix, combined with the pre-established lightning-related cloud type labels, a lightning intensity level classification feature set is generated.

[0010] Based on the lightning intensity level classification feature set and the cloud cover stratification feature matrix, calculate the coupling weight coefficient between cloud cover vertical stratification and lightning intensity classification, and generate the cloud cover-lightning vertical stratification coupling feature set.

[0011] Based on the lightning intensity level classification feature set, cloud cover dynamic feature set, and cloud cover-lightning vertical layer coupling feature set, combined with lightning monitoring data, the lightning occurrence probability distribution and lightning spatial location coordinate sequence are generated by time series analysis, forming a lightning probability distribution feature set.

[0012] Based on the lightning probability distribution feature set and the cloud cover periodic feature vector, the time series analysis method is used to generate the lightning occurrence probability time series and the lightning time window range series, forming a lightning time window prediction feature set;

[0013] Based on the periodic feature vector of cloud cover and the time window range sequence of lightning, the rhythm matching degree between the periodic fluctuation of cloud cover and the time window of lightning is calculated, and a set of cloud cover-lightning time rhythm coupling features is generated.

[0014] Based on the lightning time window prediction feature set and the lightning triggering factor dataset, the lightning movement path sequence and lightning duration sequence are generated by time series analysis, forming a lightning trend prediction feature set.

[0015] Based on the lightning trend prediction feature set, the cloud cover-lightning vertical hierarchical coupling feature set, and the cloud cover-lightning temporal rhythm coupling feature set, combined with the pre-established lightning-related cloud type labels, spatial and temporal features are extracted, and a spatiotemporal coupling feature set of cloud cover dynamic evolution and lightning trend is generated through weighted fusion.

[0016] By performing cluster analysis on the spatiotemporal coupling feature set, the dynamic evolution pattern of lightning trends is obtained; based on the dynamic evolution pattern, the time series analysis method is used to predict the lightning trends, and the final lightning trend prediction results are obtained.

[0017] In the above technical solution, the spatial distribution and time series of cloud cover are first obtained, and then the first cloud cover texture feature vector is extracted to generate a cloud cover hierarchical feature matrix. This process can fully explore the key features of cloud cover at different spatial levels and time dimensions. The first cloud cover texture feature vector is dimensionality reduced to obtain the second cloud cover texture feature vector, and combined with the cloud cover hierarchical feature matrix, a regionalized cloud cover distribution feature set is generated through clustering. Dimensionality reduction can not only reduce data dimensions and computational complexity, but also remove redundant information, retain core features, and achieve regionalized summarization of cloud cover distribution. Key meteorological element data such as humidity, temperature, and wind field are extracted to generate a lightning triggering factor dataset. These meteorological elements are closely related to the formation of lightning, and their changes play a role in inducing lightning occurrence. The regionalized cloud cover distribution feature set is analyzed using time series analysis and optical flow methods to form a dynamic cloud cover feature set. Time series analysis can capture the trend of cloud cover changes over time, while optical flow can reflect the movement state and evolution direction of clouds. Combining the two can comprehensively characterize the dynamic features of cloud cover. Based on the dynamic feature set of cloud cover and the cloud cover stratification feature matrix, combined with pre-established lightning-related cloud type labels, a lightning intensity level classification feature set is generated. By associating cloud type labels, cloud cover features are directly linked to lightning intensity levels, enabling the classification and prediction of lightning intensity. The coupling weight coefficient between cloud cover vertical stratification and lightning intensity grading is calculated, generating a cloud cover-lightning vertical stratification coupling feature set. This process fully considers the interaction between the vertical distribution of clouds and lightning intensity, providing a basis for more accurate lightning prediction. Combining lightning monitoring data, a time-series analysis method is used to generate a lightning occurrence probability distribution and a lightning spatial location coordinate sequence, forming a lightning probability distribution feature set. Analysis of historical lightning monitoring data allows the determination of the spatial and temporal probability distribution patterns of lightning occurrence. Based on the lightning probability distribution feature set and the cloud cover periodic feature vector, a time-series analysis method is used to generate a lightning occurrence probability time series and a lightning time window range series, forming a lightning time window prediction feature set. The periodic variations in cloud cover are often correlated with the temporal patterns of lightning activity. By analyzing the periodic feature vector of cloud cover, the time window in which lightning may occur can be predicted. The rhythmic matching degree between the periodic fluctuations of cloud cover and the lightning time window is calculated, generating a cloud cover-lightning time rhythm coupling feature set. This further reveals the intrinsic rhythmic relationship between cloud cover changes and the temporal distribution of lightning, providing more comprehensive temporal information for lightning trend prediction. Based on the lightning time window prediction feature set and the lightning triggering factor dataset, time series analysis is used to generate lightning movement path sequences and lightning duration sequences, forming a lightning trend prediction feature set. This enables the prediction of the spatial movement path and duration of lightning.This study comprehensively analyzes the lightning trend prediction feature set, the cloud cover-lightning vertical hierarchical coupling feature set, and the cloud cover-lightning temporal rhythm coupling feature set. Combined with lightning-related cloud type tags, spatial and temporal features are extracted, and a weighted fusion is used to generate a spatiotemporal coupling feature set of cloud cover dynamic evolution and lightning trend. This process organically combines the spatiotemporal characteristics of lightning to form a comprehensive feature set that fully reflects the spatiotemporal evolution law of lightning trends. Cluster analysis of the spatiotemporal coupling feature set yields dynamic evolution patterns of lightning trends. Different clustering results represent different lightning trend evolution paths, and these patterns reflect the changing characteristics of lightning activity under different spatiotemporal conditions. Based on the dynamic evolution patterns, time series analysis is used to predict lightning trends, obtaining the final lightning trend prediction results and providing a scientific basis for lightning protection and disaster reduction decisions.

[0018] This lightning nowcasting method, through multi-dimensional feature extraction, analysis, and fusion, fully utilizes multi-source data such as cloud cover and meteorological elements to construct a relatively complete lightning prediction system. This method can provide meteorological departments with more accurate and timely lightning nowcasting forecasts, helping them to take appropriate lightning protection measures in advance and reduce the losses to people's lives and property caused by lightning disasters. It has broad application prospects and significant practical value.

[0019] In some embodiments, humidity, temperature, and wind field data are extracted to generate a lightning triggering factor dataset. Time series analysis and optical flow methods are used to analyze the regionalized cloud cover distribution characteristic set, forming a dynamic cloud cover characteristic set, including:

[0020] Humidity, temperature and wind field data were obtained from multi-source meteorological data, and a lightning triggering factor dataset was generated through data cleaning and standardization.

[0021] Long Short-Term Memory (LSTM) networks were used to analyze the time series data in the regionalized cloud cover distribution feature set, extracting the temporal variation features of cloud cover distribution to obtain the cloud cover time series feature set.

[0022] By analyzing the cloud cover time series feature set using optical flow, the periodic changes and movement speed of cloud cover distribution are calculated, generating a periodic feature vector of cloud cover and a sequence of cloud cover movement speed.

[0023] Based on the periodic feature vector of cloud cover and the sequence of cloud cover movement speed, and by integrating humidity, temperature and wind field data, a set of dynamic evolution features of cloud cover is constructed.

[0024] In the aforementioned technical solutions, accurately capturing lightning triggering conditions and dynamic changes in cloud cover is crucial for improving forecast accuracy in the field of lightning proximity trend forecasting. The steps described above focus on a method that generates a dynamic cloud cover feature set by combining time-series analysis and optical flow methods, starting from humidity, temperature, and wind field data.

[0025] Humidity, temperature, and wind field data were acquired from a wide range of meteorological sources, including ground meteorological station observations, satellite remote sensing, and radar detection. The raw data underwent data cleaning to remove errors, outliers, and missing values. Standardization was then applied to transform data of different dimensions and ranges to a uniform scale, generating a lightning triggering factor dataset. A Long Short-Term Memory (LSTM) network was used to analyze the time series data in the regionalized cloud cover distribution feature set. LSTM is a special type of recurrent neural network capable of effectively handling long-term dependencies in time series data. Through its structure of memory units, input gates, forget gates, and output gates, it learns the cloud cover distribution time series, extracting the temporal variation characteristics of cloud cover distribution to obtain a cloud cover time series feature set, capturing the evolution of cloud cover at different time scales. Optical flow was used to analyze the cloud cover time series feature set. Optical flow calculates the periodic changes and movement speed of cloud cover distribution based on the motion information of pixels in an image sequence. First, the cloud cover time series feature set was treated as an image sequence, with the cloud cover distribution characteristics at each time step represented as an image. Then, the motion vectors of pixels between adjacent images are estimated using an optical flow algorithm to obtain the direction and speed of cloud cover movement. Based on this, the periodic changes in cloud cover distribution are further analyzed, generating periodic feature vectors and cloud cover movement speed sequences, providing crucial information for understanding the dynamic evolution of cloud cover. Based on the periodic feature vectors and cloud cover movement speed sequences, humidity, temperature, and wind field data are integrated to construct a set of dynamic evolution features for cloud cover. Humidity, temperature, and wind field, as important meteorological elements, are closely related to the formation, development, and movement of cloud cover. Combining these data with the periodicity and movement speed characteristics of cloud cover can comprehensively reflect the dynamic evolution process of cloud cover under different meteorological conditions, forming a comprehensive feature set that provides rich information for subsequent lightning trend forecasting.

[0026] This method constructs a lightning triggering factor dataset, extracts dynamic cloud cover features using LSTM and optical flow methods, and integrates multi-source meteorological data to build a set of dynamic cloud cover evolution features, achieving accurate capture of lightning triggering conditions and dynamic changes in cloud cover. This process provides crucial data support and feature descriptions for nowcasting lightning trends, helping to improve the accuracy and timeliness of lightning prediction, and has significant practical implications for early warning and prevention of lightning disasters.

[0027] In some embodiments, a lightning intensity level classification feature set is generated based on a dynamic cloud cover feature set and a hierarchical cloud cover feature matrix, combined with pre-established lightning-related cloud type labels, including:

[0028] Obtain the dynamic feature set and hierarchical feature matrix of cloud amount, and obtain the standardized feature dataset through data preprocessing; if there are missing values ​​in the standardized feature dataset, the mean imputation method is used to obtain the complete feature dataset.

[0029] Based on the complete feature dataset and pre-established lightning-related cloud type labels, a classification model is trained using the random forest algorithm to obtain a lightning intensity level classifier. The lightning intensity level classifier is then used to predict the complete feature dataset to generate an initial lightning intensity level classification result. If the confidence level of the initial lightning intensity level classification result is lower than the threshold, the classification result is adjusted using an ensemble evaluation method to obtain an optimized lightning intensity level classification result.

[0030] Based on the optimized lightning intensity level classification results, high-confidence classification features are extracted to generate a lightning intensity level classification feature set. By combining the lightning intensity level classification feature set with the spatiotemporal distribution of cloud cover dynamic features and hierarchical features, a spatiotemporal feature mapping of lightning intensity level is generated.

[0031] In the above technical solution, a dynamic feature set and a hierarchical feature matrix of cloud cover are obtained. These data contain rich information about cloud cover in time, space, and different levels. Data preprocessing is performed on the raw data, including noise removal and normalization, to eliminate differences in units and inconsistent data distribution, resulting in a standardized feature dataset. During data preprocessing, if missing values ​​are found, mean imputation is used to fill them in, estimating the missing values ​​using the mean of the existing data to ensure data integrity, thus obtaining a complete feature dataset. Based on the complete feature dataset and pre-established lightning-related cloud type labels, a classification model is trained using the random forest algorithm. Random forest is an ensemble learning algorithm that improves the accuracy and stability of classification by constructing multiple decision trees and combining their results. After sufficient training, a lightning intensity level classifier is obtained. This classifier is used to predict the complete feature dataset, generating initial lightning intensity level classification results. The initial results include predictions of the lightning intensity level for each sample; however, due to potential uncertainties in the model, the confidence level of some predictions is relatively low. To address the issue of initial lightning intensity classification results with confidence levels below a set threshold, an ensemble evaluation method is introduced to adjust the classification results. This method comprehensively considers the opinions of multiple models or evaluation indicators, optimizing the classification results through weighted voting and other methods, ultimately yielding an optimized lightning intensity classification result. The optimized classification result exhibits higher accuracy and reliability. Based on the optimized classification result, high-confidence classification features are extracted. These features represent the essence of cloud cover dynamics and stratification features closely related to lightning intensity levels, thus generating a lightning intensity level classification feature set. This feature set is then combined with the spatiotemporal distribution of cloud cover dynamics and stratification features to generate a spatiotemporal feature mapping for lightning intensity levels. This process links the classification results to the temporal and spatial distribution characteristics of cloud cover, visually demonstrating the distribution of different lightning intensity levels at different times and spatial locations.

[0032] This method achieves an effective transformation from dynamic and hierarchical cloud cover features to a set of classification features for lightning intensity levels through steps such as data preprocessing, classification model training, classification result optimization, and spatiotemporal feature mapping generation. By leveraging the random forest algorithm and ensemble evaluation method, the accuracy and reliability of the classification results are improved, while the generated spatiotemporal feature mapping provides an intuitive basis for lightning trend prediction.

[0033] In some embodiments, based on the lightning intensity level classification feature set and the cloud cover stratification feature matrix, the coupling weight coefficient between cloud cover vertical stratification and lightning intensity classification is calculated to generate a cloud cover-lightning vertical stratification coupling feature set, including:

[0034] From the lightning intensity level classification feature set and cloud cover stratification feature matrix, the data matrix of lightning intensity level and cloud cover stratification features is obtained. The main feature components are extracted by principal component analysis to obtain the feature dimensionality reduction set.

[0035] Based on the feature dimensionality reduction set, calculate the Pearson correlation coefficient between lightning intensity level and cloud cover stratification features to determine the correlation analysis results; if the absolute value of the Pearson correlation coefficient in the correlation analysis results is greater than the preset range, then mark the corresponding lightning intensity level and cloud cover stratification feature pair as a highly correlated feature pair to obtain a set of highly correlated feature pairs.

[0036] For the set of highly correlated feature pairs, a weighted linear combination method is used to calculate the coupling weight coefficients between lightning intensity level and cloud cover stratification features, thus obtaining a set of coupling weight coefficients. From the set of coupling weight coefficients, feature pairs with weight coefficients greater than a preset threshold are obtained to generate a set of cloud cover-lightning vertical stratification coupling features.

[0037] Based on the cloud cover-lightning vertical layered coupling feature set, the K-means clustering algorithm is used to classify the feature set to obtain the vertical layered structure of cloud cover-lightning coupling features. From the vertical layered structure, the correspondence between cloud cover distribution patterns and lightning activity intensity is extracted to generate the final cloud cover-lightning vertical layered coupling feature set.

[0038] In the above technical solution, a data matrix is ​​obtained from the lightning intensity level classification feature set and the cloud cover stratification feature matrix. These data represent the characteristics of lightning intensity and cloud cover at different vertical strata. Principal component analysis (PCA) is used to reduce the dimensionality of the data, extracting the main feature components to obtain a feature dimensionality reduction set. PCA can effectively remove redundant information in the data, retain the core features of the data, and reduce the data dimensionality, thereby improving the efficiency of subsequent calculations. Based on the feature dimensionality reduction set, the Pearson correlation coefficient between lightning intensity level and cloud cover stratification features is calculated to quantify the linear correlation between the two. A preset range is set. When the absolute value of the Pearson correlation coefficient is greater than this range, the corresponding lightning intensity level and cloud cover stratification feature are considered to have a strong correlation, and these are marked as highly correlated feature pairs, forming a set of highly correlated feature pairs. This step selects feature combinations with potentially important value for lightning intensity prediction. For the set of highly correlated feature pairs, a weighted linear combination method is used to calculate the coupling weight coefficient between lightning intensity level and cloud cover stratification features. The coupling weight coefficients reflect the relative influence of different cloud cover stratification features on lightning intensity, resulting in a set of coupling weight coefficients. Feature pairs with weight coefficients greater than a preset threshold are selected from this set; these feature pairs play a dominant role in the cloud cover-lightning vertical stratification coupling, thus generating a cloud cover-lightning vertical stratification coupling feature set, achieving focus on key features. The K-means clustering algorithm is used to classify the cloud cover-lightning vertical stratification coupling feature set, obtaining the vertical stratification structure of the cloud cover-lightning coupling features. K-means clustering can group similar feature sets into the same category, revealing the vertical distribution pattern of cloud cover-lightning coupling features. The correspondence between cloud cover distribution patterns and lightning activity intensity is extracted from the vertical stratification structure, ultimately generating an optimized cloud cover-lightning vertical stratification coupling feature set, providing a clearer and more regular feature representation for lightning trend prediction.

[0039] This method achieves effective coupling between cloud cover vertical stratification and lightning intensity classification through steps including data acquisition and feature extraction, correlation analysis and screening of highly correlated feature pairs, calculation of coupling weight coefficients and generation of coupled feature sets, and cluster analysis and extraction of vertical hierarchical structure. The use of principal component analysis and Pearson correlation coefficients ensures the scientific rigor of feature extraction and screening, while weighted linear combination and K-means clustering effectively reveal the vertical hierarchical relationship between cloud cover and lightning intensity. The generated cloud cover-lightning vertical hierarchical coupled feature set provides high-quality feature data for lightning nowcasting, contributing to improved precision and accuracy of lightning prediction.

[0040] In some embodiments, based on the lightning intensity level classification feature set, cloud cover dynamic feature set, and cloud cover-lightning vertical hierarchical coupling feature set, combined with lightning monitoring data, a time series analysis method is used to generate a lightning occurrence probability distribution and a lightning spatial location coordinate sequence, forming a lightning probability distribution feature set, including:

[0041] We obtained the lightning intensity level classification feature set, cloud cover dynamic feature set, and cloud cover-lightning vertical hierarchical coupling feature set. We extracted the raw spatiotemporal data from the lightning monitoring data and used data preprocessing methods to obtain a standardized feature dataset.

[0042] By using a standardized feature dataset and a long short-term memory network, a lightning occurrence probability prediction model is trained to obtain the lightning occurrence probability distribution.

[0043] Based on the probability distribution of lightning occurrence and combined with spatial location coordinates, a convolutional neural network is used to extract the spatial distribution features of lightning and obtain the sequence of lightning spatial location coordinates.

[0044] If the probability distribution of lightning occurrence exceeds the preset range, then through spatiotemporal sequence analysis, the spatial location coordinate sequence of lightning is fused to generate a set of lightning probability distribution features.

[0045] In the above technical solutions, the accurate generation of lightning occurrence probability distribution and spatial positioning coordinate sequence in the lightning nowcasting system is of vital importance for early warning of lightning activity and reducing the risk of lightning disasters.

[0046] This study collects a set of features for classifying lightning intensity levels, a set of dynamic cloud cover features, and a set of vertically coupled cloud cover-lightning features. Simultaneously, raw spatiotemporal data is extracted from lightning monitoring data. These raw data undergo preprocessing operations such as data cleaning and normalization to eliminate noise and dimensional differences, resulting in a standardized feature dataset. Using this standardized feature dataset as input, a Long Short-Term Memory (LSTM) network is used to train a lightning occurrence probability prediction model. LSTM effectively captures long-term dependencies in time-series data. Through its memory units and gating mechanism, it learns the temporal evolution of lightning-related features, thereby predicting the lightning occurrence probability distribution. This probability distribution reflects the likelihood of lightning occurrence at different time points and spatial locations. The lightning occurrence probability distribution is combined with spatial location coordinates, and a Convolutional Neural Network (CNN) is used to extract the spatial distribution features of lightning, obtaining a sequence of lightning spatial location coordinates. CNN excels at processing grid-like data, such as images, and can automatically learn the local and global features of lightning spatial distribution, thus achieving accurate localization and sequential representation of lightning activity spatial locations. If the lightning probability distribution exceeds a preset range, it indicates that lightning activity is relatively active and the probability of occurrence is high. In this case, spatiotemporal series analysis is used to integrate the spatial location coordinate sequence of lightning events to generate a lightning probability distribution feature set. Spatiotemporal series analysis comprehensively considers the temporal and spatial trends of lightning probability, incorporating the spatial location coordinate sequence of lightning events into the probability distribution. This ensures that the final lightning probability distribution feature set not only contains lightning occurrence probability information but also reflects the spatial evolution of lightning activity, providing a more comprehensive and detailed feature description for lightning trend prediction.

[0047] This method achieves an effective transformation from a multi-source feature set to a lightning probability distribution feature set through steps such as data acquisition and preprocessing, lightning occurrence probability prediction model training, lightning spatial distribution feature extraction, and generation of a lightning probability distribution feature set. The application of LSTM and CNN fully leverages their respective advantages in time series analysis and spatial feature extraction, ensuring the accurate generation of the lightning occurrence probability distribution and spatial location coordinate sequences. The generated lightning probability distribution feature set provides high-quality, information-rich feature data for lightning nowcasting, contributing to improved precision and accuracy in lightning prediction.

[0048] In some embodiments, based on the lightning probability distribution feature set and the cloud cover periodic feature vector, a time series analysis method is used to generate a lightning occurrence probability time series and a lightning time window range series, forming a lightning time window prediction feature set, including:

[0049] Input data is obtained from the lightning probability distribution feature set and cloud cover periodic feature vector, and standardized feature data input is obtained through data preprocessing;

[0050] A long short-term memory network is used to train standardized feature data input to generate a time series of lightning occurrence probabilities;

[0051] Based on the lightning occurrence probability time series, and by judging through preset thresholds, the time points with high probability of lightning occurrence are determined, and the lightning time window range sequence is obtained;

[0052] For the lightning time window range sequence, cluster analysis is used to extract the periodic features of the time window, and a set of periodic features of the time window is obtained;

[0053] Feature vectors are obtained from the periodic feature set of the time window, and then a second training is performed through a long short-term memory network to generate a lightning time window prediction feature set.

[0054] If the matching degree between the lightning time window prediction feature set and the lightning occurrence probability time series is lower than the preset range, the parameters of the long short-term memory network are adjusted through iterative optimization to obtain an optimized prediction feature set.

[0055] Based on the optimized prediction feature set, the time series is smoothed using the sliding window method to obtain the final lightning time window prediction feature set.

[0056] In the above technical solution, input data is obtained from the lightning probability distribution feature set and the cloud cover periodic feature vector. These data integrate key information on lightning occurrence probability and cloud cover periodic changes. The raw data undergoes preprocessing, including noise removal and normalization, to eliminate differences in data units and noise interference, resulting in standardized feature data input to ensure the accuracy of subsequent analysis. A Long Short-Term Memory (LSTM) network is used to train the standardized feature data input. LSTM can effectively capture long-term dependencies in time series data. Through its unique memory units and gating mechanism, it learns the temporal evolution of lightning probability and cloud cover periodic features, thereby generating a lightning occurrence probability time series. This time series details the trend of lightning occurrence probability over time. Based on the lightning occurrence probability time series, a preset threshold is used to determine high-probability lightning occurrence time points. That is, when the lightning occurrence probability exceeds the set threshold, it is considered that lightning may occur at that time point, thus obtaining a lightning time window range sequence. This sequence initially delineates the time interval where lightning may occur, providing a foundation for subsequent analysis. For the lightning time window range sequence, cluster analysis is used to extract the periodic features of the time windows, resulting in a set of periodic features. Cluster analysis can group similar time window ranges into one class, thus revealing the periodic distribution pattern of lightning time windows over time and further deepening the understanding of lightning activity timing patterns. Feature vectors are obtained from the set of periodic features of the time windows, and a Long Short-Term Memory (LSTM) network is used for secondary training to generate a set of predicted features for lightning time windows. The secondary training aims to further explore the deep correlation between the periodic features of the time windows and the lightning occurrence time, improving the accuracy and relevance of the predicted feature set. If the matching degree between the predicted feature set of lightning time windows and the lightning occurrence probability time series is lower than a preset range, the parameters of the LSM network are adjusted through iterative optimization. The iterative optimization process continuously fine-tunes the model parameters to improve the consistency between the predicted feature set and the actual lightning probability time series, thereby obtaining an optimized predicted feature set. Based on the optimized predicted feature set, a sliding window method is used to smooth the time series. The sliding window method can effectively reduce short-term fluctuations and noise interference in time series, making the prediction results smoother and more stable, and finally obtaining a high-quality lightning time window prediction feature set, providing a reliable basis for lightning time prediction.

[0057] This method achieves an effective transformation from a lightning probability distribution feature set and cloud cover periodic feature vector to a lightning time window prediction feature set through steps such as data acquisition and preprocessing, generation of lightning occurrence probability time series, determination of lightning time window range sequence, extraction of periodic features of the time window, generation of lightning time window prediction feature set, optimization of prediction feature set, and time series smoothing. The application of LSTM and cluster analysis fully leverages their respective advantages in time series analysis and periodic feature extraction, ensuring the accuracy and reliability of lightning time window prediction. The generated lightning time window prediction feature set provides high-quality, information-rich feature data for lightning nowcasting, contributing to improved precision and accuracy of lightning prediction.

[0058] In some embodiments, based on the periodic feature vector of cloud cover and the lightning time window range sequence, the rhythm matching degree between the periodic fluctuation of cloud cover and the lightning time window is calculated to generate a set of cloud cover-lightning time rhythm coupling features, including:

[0059] The periodic feature vector of cloud cover and the time window range sequence of lightning are obtained. Data preprocessing methods are used to normalize and denoise the data to obtain the cloud cover feature sequence and lightning time sequence.

[0060] By using signal processing techniques, periodic fluctuation features are extracted from cloud cover feature sequences to generate a cloud cover periodic feature set. Cross-correlation analysis is then used to calculate the temporal correlation between the cloud cover periodic feature set and the lightning time series, resulting in a rhythm matching degree sequence.

[0061] If a value in the rhythm matching degree sequence exceeds the preset range, the corresponding cloud cover periodic feature is paired with the lightning time window to generate a preliminary coupling feature set. For the preliminary coupling feature set, the principal component analysis method is used to extract the main feature components to obtain a simplified cloud cover-lightning time rhythm coupling feature set.

[0062] By analyzing time series data, periodic patterns are detected in a simplified set of coupled features, generating the final set of cloud cover-lightning time rhythm coupled features.

[0063] In the above technical solution, periodic feature vectors of cloud cover and lightning time window range sequences are obtained. These data respectively contain the periodic variation pattern of cloud cover over time and the time interval in which lightning may occur. Data preprocessing methods are used on the raw data, including normalization to scale the data to a uniform scale, and denoising algorithms to remove noise interference, resulting in smooth and dimensionlessly consistent cloud cover feature sequences and lightning time sequences, providing a high-quality data foundation for subsequent analysis. Periodic fluctuation features are extracted from the cloud cover feature sequences using signal processing techniques. Methods such as Fast Fourier Transform are used to identify the main periodic components of cloud cover changes, generating a periodic feature set of cloud cover. Then, cross-correlation analysis is used to calculate the temporal correlation between the periodic feature set of cloud cover and the lightning time sequence. Cross-correlation analysis can quantify the delayed correlation between the two time series, thus obtaining a rhythm matching degree sequence, reflecting the degree of correlation between cloud cover periodic fluctuations and lightning time windows at different time lags. If any value in the rhythm matching sequence exceeds a preset range, it is considered that there is a significant correlation between the corresponding cloud cover periodicity feature and the lightning time window. These features are then paired to generate a preliminary coupled feature set. Principal component analysis (PCA) is used to extract the main feature components from this preliminary coupled feature set. PCA effectively reduces data dimensionality, removes redundant information, and retains the core features in the coupled feature set, resulting in a simplified cloud cover-lightning time rhythm coupled feature set, improving the data's compactness and representativeness. Time series analysis is then used to detect periodic patterns in the simplified coupled feature set. Autocorrelation analysis and spectral analysis are employed to identify stable periodic patterns in the coupled feature set, further optimizing and refining the feature set to generate the final cloud cover-lightning time rhythm coupled feature set. This final feature set not only retains the key correlation features between cloud cover periodic fluctuations and the lightning time window but also reflects its periodic variation pattern, providing a more accurate and efficient feature representation for lightning trend prediction.

[0064] This method, through steps including data preprocessing, feature extraction and correlation analysis, preliminary coupled feature set generation and simplification, and final coupled feature set generation, achieves the calculation of the rhythm matching degree between cloud cover periodic fluctuations and lightning time windows, and the generation of coupled feature sets. The application of signal processing techniques and cross-correlation analysis ensures the scientific rigor of feature extraction and correlation analysis, while principal component analysis and time series analysis effectively improve the quality and representativeness of the feature sets. The generated cloud cover-lightning time rhythm coupled feature set provides high-quality, information-rich feature data for lightning nowcasting, contributing to improved precision and accuracy in lightning prediction.

[0065] In some embodiments, based on the lightning time window prediction feature set and the lightning triggering factor dataset, a time series analysis method is used to generate a lightning movement path sequence and a lightning duration sequence, forming a lightning trend prediction feature set, including:

[0066] Obtain the lightning time window prediction feature set and lightning triggering factor dataset, and perform standardization processing through the data preprocessing module to obtain the standardized feature set and standardized triggering factor dataset;

[0067] A long short-term memory network is used to perform time series analysis on a standardized feature set to generate lightning movement path sequences, thus obtaining path sequence data.

[0068] Based on a standardized triggering factor dataset, a long short-term memory network is used to predict the duration of lightning, generating a lightning duration sequence and obtaining duration sequence data.

[0069] By fusion of features, path sequence data and duration sequence data are merged to construct a feature set for lightning trend prediction.

[0070] In the above technical solution, a lightning time window prediction feature set and a lightning triggering factor dataset are obtained. These data contain the temporal patterns of lightning occurrence and meteorological conditions that trigger lightning, respectively. The raw data is standardized through a data preprocessing module, including outlier removal and normalization, to eliminate differences in data dimensions and noise interference, resulting in a standardized feature set and a standardized triggering factor dataset. A Long Short-Term Memory (LSTM) network is used to perform time series analysis on the standardized feature set. LSTM can effectively capture long-term dependencies in time series data. Through its memory units and gating mechanism, it learns the spatiotemporal information in the lightning time window prediction feature set, thereby generating a lightning movement path sequence, resulting in path sequence data. This path sequence details the spatial trajectory of lightning activity, providing a basis for predicting the direction of lightning propagation. Based on the standardized triggering factor dataset, the LSTM network is again used for duration prediction. LSTM analyzes the time series of meteorological elements in the triggering factor dataset, mining the intrinsic correlation between factors such as humidity, temperature, and wind field and lightning duration, thereby generating a lightning duration sequence, resulting in duration sequence data. This sequence reflects the temporal continuity of lightning activity, helping to assess the duration and intensity variations of lightning events. Feature fusion technology is used to merge path sequence data and duration sequence data. The feature fusion process comprehensively considers two key aspects: lightning movement path and duration, organically integrating the data from both to construct a feature set that comprehensively reflects the trend of lightning activity—the lightning trend prediction feature set. This set not only includes spatial movement information of lightning but also incorporates its temporal persistence, providing a more comprehensive and accurate feature description for lightning trend prediction.

[0071] This method, through steps such as data acquisition and preprocessing, lightning movement path sequence generation, lightning duration sequence generation, and construction of a lightning trend prediction feature set, achieves an effective transformation from a lightning time window prediction feature set and a lightning triggering factor dataset to a lightning trend prediction feature set. The two applications of the Long Short-Term Memory (LSTM) network fully leverage its advantages in time series analysis, accurately capturing the characteristics of lightning movement paths and durations, respectively. The final generated lightning trend prediction feature set provides high-quality, information-rich feature data for near-term lightning trend forecasting, contributing to improved precision and accuracy in lightning prediction.

[0072] In some embodiments, based on the lightning trend prediction feature set, the cloud cover-lightning vertical hierarchical coupling feature set, and the cloud cover-lightning temporal rhythm coupling feature set, combined with pre-established lightning-related cloud type labels, spatial and temporal features are extracted, and a spatiotemporal coupling feature set of cloud cover dynamic evolution and lightning trend is generated through weighted fusion, including:

[0073] The lightning trend prediction feature set, the cloud cover-lightning vertical hierarchical coupling feature set, and the cloud cover-lightning temporal rhythm coupling feature set are obtained. The feature sets are normalized using a standardization processing method to obtain a normalized feature set.

[0074] Based on the normalized feature set, a convolutional neural network is used to process the lightning occurrence probability and cloud coverage of the vertical cloud cover data in the lightning trend prediction feature set to extract spatial distribution features and obtain a spatial feature set.

[0075] Based on the normalized feature set, a long short-term memory network is used to process the time rhythm coupled data, extract the time dynamic features, and obtain the time feature set;

[0076] A weighted fusion method is used to integrate the spatial feature set and the temporal feature set to obtain a fused feature set;

[0077] Based on the fusion feature set and combined with the pre-established lightning-related cloud type labels, a fully connected neural network is used for classification to obtain a spatiotemporal coupled feature set of cloud cover dynamic evolution and lightning trend.

[0078] In the above technical solution, a feature set for lightning trend prediction, a feature set for cloud cover-lightning vertical layer coupling, and a feature set for cloud cover-lightning temporal rhythm coupling are obtained. These feature sets reflect the patterns of lightning activity and the correlation between cloud cover and lightning from different perspectives. A standardization method is used to normalize the feature sets, eliminating dimensional differences and inconsistent data distributions between different features, resulting in a normalized feature set. Based on the normalized feature set, a convolutional neural network (CNN) is used to process the cloud cover vertical layer data. CNN excels at processing grid-like data and can automatically learn the spatial distribution characteristics of cloud cover in vertical layers. Through operations such as convolutional layers and pooling layers, local and global spatial distribution patterns of cloud cover in different vertical layers are extracted, resulting in a spatial feature set, providing key information for the spatial localization of lightning activity and the spatial patterns of cloud cover evolution. Also based on the normalized feature set, a long short-term memory network (LSTM) is used to process the temporal rhythm coupling data. LSTM effectively captures long-term dependencies in time-series data. Through its memory units and gating mechanisms, it learns the coupled features of lightning time rhythms, extracting the dynamic changes in lightning activity over time to obtain a temporal feature set. This set reflects the evolutionary trends of lightning occurrence probability and duration over time, providing support for predicting the temporal patterns of lightning activity. If the spatial and temporal feature sets have the same dimensionality, a weighted fusion method is used to integrate them. Weighted fusion assigns different weights to spatial and temporal features based on their importance in lightning prediction, linearly combining them into a fused feature set. If the dimensions are inconsistent, the feature set with the lower dimension is first linearly interpolated to match the other feature set before weighted fusion to obtain the fused feature set. The fused feature set retains both the spatial distribution information of cloud cover and lightning activity and incorporates temporal evolution patterns, forming a comprehensive feature representation. Based on the fused feature set and pre-established lightning-related cloud type labels, a fully connected neural network is used for classification. Fully connected neural networks can deeply learn and abstract the fused features. Through the transformation of multiple layers of neurons, they ultimately output a spatiotemporally coupled feature set of cloud cover dynamics and lightning trends. This set comprehensively reflects the complex correlation between cloud cover in vertical stratification and temporal rhythms and lightning activity, providing high-quality and information-rich feature data for lightning trend prediction.

[0079] This method achieves effective fusion of lightning trend prediction features, cloud cover-lightning vertical hierarchical coupling features, and cloud cover-lightning temporal rhythm coupling features through steps such as data acquisition and standardization, spatial feature extraction, temporal feature extraction, feature fusion, and generation of spatiotemporally coupled feature sets. The combination of convolutional neural networks and long short-term memory networks fully leverages their respective advantages in spatial feature extraction and time series analysis, ensuring the high-quality generation of spatiotemporally coupled feature sets. The generated spatiotemporally coupled feature sets provide more accurate and comprehensive feature data for lightning nowcasting, contributing to improved precision and accuracy in lightning prediction.

[0080] According to another aspect of the present invention, a lightning now-near trend forecasting system is provided, the system comprising:

[0081] The first feature set module is used to obtain the spatial distribution and time series of cloud cover, extract the first cloud cover texture feature vector, and generate a cloud cover hierarchical feature matrix; reduce the dimensionality of the first cloud cover texture feature vector to generate the second cloud cover texture feature vector, and combine it with the cloud cover hierarchical feature matrix to generate a regional cloud cover distribution feature set through clustering.

[0082] The second feature set module is used to extract humidity, temperature, and wind field data to generate a lightning triggering factor dataset. It uses time series analysis and optical flow methods to analyze the regional cloud cover distribution feature set and form a dynamic cloud cover feature set.

[0083] The third feature set module is used to generate a lightning intensity level classification feature set based on the cloud cover dynamic feature set and cloud cover hierarchical feature matrix, combined with pre-established lightning-related cloud type labels.

[0084] The fourth feature set module is used to calculate the coupling weight coefficient between cloud cover vertical stratification and lightning intensity classification based on the lightning intensity level classification feature set and cloud cover stratification feature matrix, and generate the cloud cover-lightning vertical stratification coupling feature set.

[0085] The fifth feature set module is used to classify feature sets based on lightning intensity level, cloud cover dynamic feature set, and cloud cover-lightning vertical layer coupling feature set. Combined with lightning monitoring data, it uses time series analysis to generate lightning occurrence probability distribution and lightning spatial location coordinate sequence, forming lightning probability distribution feature set.

[0086] The sixth feature set module is used to generate a lightning occurrence probability time series and a lightning time window range series based on the lightning probability distribution feature set and the cloud cover periodic feature vector, using time series analysis to form a lightning time window prediction feature set.

[0087] The seventh feature set module is used to calculate the rhythm matching degree between cloud cover periodic fluctuations and lightning time windows based on the cloud cover periodic feature vector and the lightning time window range sequence, and generate a cloud cover-lightning time rhythm coupling feature set.

[0088] The eighth feature set module is used to generate lightning movement path sequences and lightning duration sequences based on the lightning time window prediction feature set and lightning triggering factor dataset, and to form a lightning trend prediction feature set.

[0089] The ninth feature set module is used to extract spatial and temporal features based on the lightning trend prediction feature set, the cloud cover-lightning vertical hierarchical coupling feature set, and the cloud cover-lightning temporal rhythm coupling feature set, combined with pre-established lightning-related cloud type labels, and generate a spatiotemporal coupling feature set of cloud cover dynamic evolution and lightning trend through weighted fusion.

[0090] The prediction module is used to obtain the dynamic evolution pattern of lightning trends by performing cluster analysis on the spatiotemporal coupling feature set; based on the dynamic evolution pattern, the time series analysis method is used to predict the lightning trend and obtain the final lightning trend prediction result.

[0091] In order to better utilize the above methods, this application proposes a lightning proximity trend forecasting system. Each module corresponds to a step of the above methods, and its specific principles have been described above and will not be repeated here. Attached Figure Description

[0092] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0093] Figure 1 This is a flowchart illustrating an embodiment of a lightning proximity trend forecasting method according to the present invention;

[0094] Figure 2 This is a schematic diagram of an embodiment of a lightning proximity trend forecasting system according to the present invention. Detailed Implementation

[0095] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be particularly noted that the following embodiments are for illustrative purposes only and do not limit the scope of the invention. Similarly, the following embodiments are only some, not all, embodiments of the present invention, and all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0096] Example 1

[0097] Please see Figure 1 A method for predicting the imminent trend of lightning, the method comprising:

[0098] S1. Obtain the spatial distribution and time series of cloud cover, extract the first cloud cover texture feature vector, and generate a cloud cover hierarchical feature matrix; reduce the dimensionality of the first cloud cover texture feature vector to generate the second cloud cover texture feature vector, and combine it with the cloud cover hierarchical feature matrix to generate a regionalized cloud cover distribution feature set through clustering.

[0099] In this embodiment, S1, acquire the spatial distribution and time series of cloud cover, extract the first cloud cover texture feature vector, and generate a cloud cover hierarchical feature matrix; reduce the dimensionality of the first cloud cover texture feature vector to generate a second cloud cover texture feature vector, and combine it with the cloud cover hierarchical feature matrix to generate a regionalized cloud cover distribution feature set through clustering, including:

[0100] S11. Obtain the spatial distribution and time series of cloud cover from multi-source meteorological data, use a convolutional neural network to extract the first cloud cover texture feature vector, combine the gradient operator to generate a cloud cover hierarchical feature matrix, and generate a set of cloud cover spatiotemporal features by feature splicing.

[0101] S12. Perform principal component analysis on the first cloud texture feature vector in the cloud spatiotemporal feature set, retain the features whose principal component dimension is lower than the preset threshold, generate the second cloud texture feature vector, and combine the cloud layer feature matrix layered by cloud height to generate a regional cloud distribution feature set through K-means clustering.

[0102] For example, cloud cover spatial distribution and time series are obtained from multi-source meteorological data. First, satellite remote sensing data (such as MODIS) and ground station observation data (such as the ERA5 reanalysis dataset) are fused to obtain cloud cover grid data with a resolution of 0.25°×0.25°, spanning from January 1 to December 31, 2023, with data collected every 3 hours, for a total of 2920 time periods. Data preprocessing employs a weighted average algorithm, with satellite data weighted at 0.7 and ground station data weighted at 0.3, ensuring a combination of high resolution and ground-based measurement accuracy, generating a cloud cover spatial distribution matrix (720×1440 dimensions) and a time series vector (2920 dimensions). Next, a convolutional neural network (CNN) is used to extract the first cloud cover texture feature vector. The network structure includes three convolutional layers (3×3 kernels, 1 stride, 1 padding), with each layer using the ReLU activation function, outputting a 32-dimensional feature vector. The specific process is as follows: The cloud cover spatial distribution matrix is ​​input into a CNN, and convolutional operations are used to extract local texture features (such as cloud boundary roughness). Dimensionality reduction is achieved through max pooling (2×2), resulting in a 32-dimensional texture feature vector per time interval, reflecting the spatial heterogeneity of cloud cover. Then, the Sobel gradient operator is used to calculate the cloud cover hierarchical feature matrix. Horizontal and vertical convolutional kernels ([-1,0,1;-2,0,2;-1,0,1] and their transposes) are applied to the cloud cover grid data to generate a gradient magnitude matrix (720×1440), representing the intensity of cloud cover edges. The maximum gradient value is normalized to 1.0, reflecting the drastic changes in cloud cover boundaries. Finally, a spatiotemporal feature set of cloud cover is generated through feature concatenation. The 32-dimensional texture feature vector is concatenated with the 16-dimensional principal components (extracted via PCA, with a variance contribution rate of 95%) after dimensionality reduction of the gradient matrix, resulting in a 48-dimensional spatiotemporal feature vector per time interval. After feature concatenation, cloud cover spatiotemporal patterns can be divided using K-means clustering (K=5). For example, high cloud cover areas are concentrated in the Intertropical Convergence Zone, while low cloud cover areas appear in the subtropical high region. The time series shows that the fluctuation range of cloud cover in summer (standard deviation 0.15) is greater than that in winter (0.08).

[0103] S2. Extract humidity, temperature, and wind field data to generate a lightning triggering factor dataset. Use time series analysis and optical flow methods to analyze the regional cloud cover distribution feature set and form a dynamic cloud cover feature set.

[0104] In this embodiment, S2, humidity, temperature, and wind field data are extracted to generate a lightning triggering factor dataset. Time series analysis and optical flow methods are used to analyze the regionalized cloud cover distribution characteristic set, forming a dynamic cloud cover characteristic set, including:

[0105] S21. Obtain humidity, temperature and wind field data from multi-source meteorological data, and generate a lightning triggering factor dataset through data cleaning and standardization.

[0106] S22. Long Short-Term Memory Network is used to analyze the time series in the regional cloud cover distribution feature set, extract the time change features of cloud cover distribution, and obtain the cloud cover time series feature set.

[0107] S23. Analyze the cloud cover time series feature set using optical flow method, calculate the periodic changes and movement speed of cloud cover distribution, and generate cloud cover periodic feature vectors and cloud cover movement speed sequences.

[0108] S24. Based on the periodic feature vector of cloud cover and the sequence of cloud cover movement speed, integrate humidity, temperature and wind field data to construct a set of dynamic evolution features of cloud cover.

[0109] For example, a lightning triggering factor dataset is generated by extracting humidity, temperature, and wind field data from multi-source meteorological data. First, data preprocessing is performed to obtain relative humidity (%), 2-meter temperature (°C), and 10-meter wind field (u and v components, m / s) at a resolution of 0.25°×0.25° from the ECMWF ERA5 dataset. The temporal resolution is 1 hour, covering the area from 30°N to 40°N and 110°E to 120°E. Assuming data from 00:00 to 23:00 on July 1, 2025, relative humidity (range 60%-90%), temperature (20-35°C), and wind speed (0-15 m / s) are extracted. Principal component analysis (PCA) is used for dimensionality reduction, retaining 90% of the variance, to generate the lightning triggering factor dataset, which includes a dominant humidity factor (weight 0.6), a temperature factor (weight 0.3), and a wind field factor (weight 0.1). A Long Short-Term Memory (LSTM) network was used to analyze the time series of regional cloud cover distribution characteristics. The input cloud cover data was MODIS cloud coverage (0-100%), with a time step of 24 hours. A 3-layer LSTM model (128, 64, and 32 hidden units) was constructed, trained for 100 epochs using the Adam optimizer (learning rate 0.001) and the mean squared error (MSE) loss function. The output was a predicted cloud cover sequence (error <5%). A periodic feature vector and movement velocity sequence of cloud cover were generated using optical flow. Based on the Farneback optical flow algorithm, the cloud cover movement vector was calculated from GOES-16 satellite cloud imagery (2km resolution). Assuming the cloud cluster velocity range was 0-10 m / s, the periodic features were extracted using Fast Fourier Transform (FFT), with dominant periods of 12 hours and 24 hours. This generated a feature vector (dimension 10, including period intensities of 0.2-0.8) and a velocity sequence (mean 5 m / s, standard deviation 1.2 m / s). The final result is a set of dynamic cloud cover evolution features, which integrates triggering factors, periodic features, and velocity sequences. Cloud cover evolution patterns are classified using K-means clustering (K=3), resulting in three feature sets: high dynamic (velocity >7 m / s), medium dynamic (3-7 m / s), and low dynamic (<3 m / s), with a classification accuracy of 85%. Triggering factors provide the environmental basis for lightning occurrence, LSTM captures the temporal patterns of cloud cover, optical flow quantifies motion characteristics, and clustering integrates dynamic features to provide data support for lightning prediction.

[0110] S3. Based on the dynamic feature set of cloud cover and the hierarchical feature matrix of cloud cover, combined with the pre-established lightning-related cloud type labels, generate a lightning intensity level classification feature set.

[0111] In this embodiment, S3, based on the cloud cover dynamic feature set and cloud cover hierarchical feature matrix, and combined with pre-established lightning-related cloud type labels, a lightning intensity level classification feature set is generated, including:

[0112] S31. Obtain the dynamic feature set and hierarchical feature matrix of cloud cover. Through data preprocessing, obtain a standardized feature dataset. If there are missing values ​​in the standardized feature dataset, use the mean imputation method to obtain a complete feature dataset.

[0113] S32. Based on the complete feature dataset and the pre-established lightning-related cloud type labels, a classification model is trained using the random forest algorithm to obtain a lightning intensity level classifier; the lightning intensity level classifier is used to predict the complete feature dataset to generate an initial lightning intensity level classification result; if the confidence of the initial lightning intensity level classification result is lower than the threshold, the classification result is adjusted using an ensemble evaluation method to obtain an optimized lightning intensity level classification result.

[0114] S33. Based on the optimized lightning intensity level classification results, extract high-confidence classification features and generate a lightning intensity level classification feature set; through the lightning intensity level classification feature set, combine the spatiotemporal distribution of cloud cover dynamic features and hierarchical features to generate a spatiotemporal feature mapping of lightning intensity levels.

[0115] For example, the implementation method for generating a lightning intensity level classification feature set using the random forest algorithm, based on a dynamic cloud cover feature set and a hierarchical cloud cover feature matrix, combined with pre-established lightning-related cloud type labels, is as follows. Assume the dynamic cloud cover feature set includes cloud cover change rate and cloud movement speed, with 1000 sample points. Each point includes the cloud cover change rate (0.1 to 0.5, unit: % / minute) and cloud movement speed (5 to 20, unit: m / s). The first hierarchical cloud cover feature matrix includes the cloud cover percentages of low-level clouds (0-2km), mid-level clouds (2-6km), and high-level clouds (6-12km), with a matrix dimension of 1000×3 and values ​​ranging from 0 to 1.

[0116] For example, a sample point with data of [0.3, 0.4, 0.2] represents the proportion of low, medium, and high-altitude cloud cover. Lightning-related cloud type labels are based on historical data and are categorized into cumulonimbus (Cb), stratocumulus (Sc), etc., with a label set of 1000, labeled as 0 (no lightning), 1 (weak lightning), and 2 (strong lightning). First, dynamic and hierarchical features are fused to construct a feature vector with a dimension of 1000×5 (2 dynamic features + 3 hierarchical features). The feature vector is then normalized using the Min-Max method, mapping values ​​to [0,1]. For example, the cloud cover change rate of 0.3 is normalized to (0.3-0.1) / (0.5-0.1)=0.5. A random forest algorithm is used, with 100 decision trees and a maximum depth of 10. The Gini coefficient is used for feature selection. After splitting the training set (80%) and the test set (20%), the model is trained. The model outputs the probability of lightning intensity level for each sample. For example, if a sample's prediction result is [0.1, 0.6, 0.3], the highest probability corresponds to level 1 (weak lightning). The model performance is evaluated by calculating the test set accuracy (e.g., 85%) and analyzing feature importance. It is found that the cloud cover change rate and the proportion of high-level clouds contribute the most to the classification, with importance scores of 0.4 and 0.35, respectively. Finally, a lightning intensity level classification feature set is generated, containing the predicted levels and probability distributions for 1000 samples.

[0117] S4. Based on the lightning intensity level classification feature set and the cloud cover stratification feature matrix, calculate the coupling weight coefficient between cloud cover vertical stratification and lightning intensity classification, and generate the cloud cover-lightning vertical stratification coupling feature set.

[0118] In this embodiment, S4, based on the lightning intensity level classification feature set and the cloud cover stratification feature matrix, calculate the coupling weight coefficient between cloud cover vertical stratification and lightning intensity classification, and generate a cloud cover-lightning vertical stratification coupling feature set, including:

[0119] S41. Obtain the data matrix of lightning intensity level and cloud cover stratification features from the lightning intensity level classification feature set and cloud cover stratification feature matrix, and use principal component analysis to extract the main feature components to obtain the feature dimensionality reduction set.

[0120] S42. Based on the feature dimensionality reduction set, calculate the Pearson correlation coefficient between lightning intensity level and cloud cover stratification feature, and determine the correlation analysis results; if the absolute value of the Pearson correlation coefficient in the correlation analysis results is greater than the preset range, then mark the corresponding lightning intensity level and cloud cover stratification feature pair as a highly correlated feature pair, and obtain the set of highly correlated feature pairs.

[0121] S43. For the set of highly correlated feature pairs, a weighted linear combination method is used to calculate the coupling weight coefficients between lightning intensity level and cloud cover stratification features, and a set of coupling weight coefficients is obtained. From the set of coupling weight coefficients, feature pairs with weight coefficients greater than a preset threshold are obtained to generate a set of cloud cover-lightning vertical stratification coupling features.

[0122] S44. Based on the cloud cover-lightning vertical layered coupling feature set, the K-means clustering algorithm is used to classify the feature set to obtain the vertical layered structure of cloud cover-lightning coupling features; from the vertical layered structure, the correspondence between cloud cover distribution patterns and lightning activity intensity is extracted to generate the final cloud cover-lightning vertical layered coupling feature set.

[0123] For example, suppose the first lightning intensity level classification feature set contains lightning intensity levels (divided into levels 1 to 5, corresponding to current intensities of 10kA, 20kA, 30kA, 40kA, and 50kA respectively), and the first cloud cover stratification feature matrix contains three layers of cloud cover data (low-level clouds 0-2km, mid-level clouds 2-5km, and high-level clouds 5-10km), with each layer's cloud cover percentage represented as 0-100%. The data sample consists of 100 meteorological observation points, each containing the lightning intensity level and the corresponding three-layer cloud cover value. For example, sample 1: lightning intensity level 3 (30kA), low-level clouds 80%, mid-level clouds 60%, and high-level clouds 20%. First, construct the feature matrix, converting the lightning intensity level and the three-layer cloud cover values ​​into matrix X, where each row of X represents a sample, and the columns are [lightning intensity, low-level cloud cover, mid-level cloud cover, high-level cloud cover], for example, [30, 80, 60, 20]. To calculate the coupling weighting coefficient between cloud cover vertical stratification and lightning intensity classification, Pearson correlation coefficient analysis was used, with the formula r = cov(Xi, Y) / [std(Xi)·std(Y)], where Xi is the cloud cover at a certain layer and Y is the lightning intensity. For the calculation of low-level cloud cover and lightning intensity, assuming the sample means are μ_X1=70 (mean low-level cloud cover) and μ_Y=25 (mean lightning intensity), the covariance cov(X1, Y)=150, and the standard deviations std(X1)=15 and std(Y)=10, then r1=150 / (15·10)=1, indicating a high positive correlation between low-level cloud cover and lightning intensity. Similarly, the mid-level cloud cover was calculated as r2=0.8, and the high-level cloud cover as r3=0.4. The weighting coefficients are obtained by normalizing the correlation coefficients, using the formula wi=ri / Σri, where Σri=1+0.8+0.4=2.2. Therefore, w1=1 / 2.2≈0.455, w2=0.8 / 2.2≈0.364, and w3=0.4 / 2.2≈0.182. A vertically layered coupled feature set of cloud cover and lightning is generated. Cloud cover data is fused using a weighted fusion formula F=Σ(wi·Xi). For sample 1, F=0.455·80+0.364·60+0.182·20=36.4+21.84+3.64=61.88, generating a new feature vector [30, 61.88]. This calculation is repeated for 100 samples to form the coupled feature set.

[0124] S5. Based on the lightning intensity level classification feature set, cloud cover dynamic feature set, and cloud cover-lightning vertical layer coupling feature set, combined with lightning monitoring data, the lightning occurrence probability distribution and lightning spatial location coordinate sequence are generated by time series analysis, forming a lightning probability distribution feature set.

[0125] In this embodiment, S5, based on the lightning intensity level classification feature set, cloud cover dynamic feature set, and cloud cover-lightning vertical layer coupling feature set, combined with lightning monitoring data, a time series analysis method is used to generate a lightning occurrence probability distribution and a lightning spatial location coordinate sequence, forming a lightning probability distribution feature set, including:

[0126] S51. Obtain the lightning intensity level classification feature set, cloud cover dynamic feature set, and cloud cover-lightning vertical hierarchical coupling feature set. Extract the original spatiotemporal data from the lightning monitoring data and use data preprocessing methods to obtain a standardized feature dataset.

[0127] S52. Using a standardized feature dataset and a long short-term memory network, train a lightning occurrence probability prediction model to obtain the lightning occurrence probability distribution.

[0128] S53. Based on the probability distribution of lightning occurrence and combined with spatial location coordinates, a convolutional neural network is used to extract the spatial distribution features of lightning to obtain the sequence of lightning spatial location coordinates.

[0129] S54. If the probability distribution of lightning occurrence exceeds the preset range, then through spatiotemporal sequence analysis, the spatial location coordinate sequence of lightning is fused to generate a set of lightning probability distribution features.

[0130] For example, based on the lightning intensity level classification feature set, cloud cover dynamic feature set, and cloud cover-lightning vertical hierarchical coupling feature set, combined with lightning monitoring data, a Long Short-Term Memory (LSTM) network is used to generate the first lightning occurrence probability distribution and the first lightning spatial location coordinate sequence, forming the first lightning probability distribution feature set. The implementation method is as follows: Assuming the lightning intensity level classification feature set includes lightning intensity levels (divided into levels 1 to 5, corresponding to current intensities of 10kA, 20kA, 30kA, 40kA, and 50kA respectively), classification is performed using radar reflectivity data (in dBZ, ranging from 20-60). A Support Vector Machine (SVM) algorithm is used, with the radial basis function (RBF) kernel function selected, the penalty parameter C set to 1.0, and the kernel parameter γ set to 0.01. Training is performed on 1000 sets of sample data (each set including radar reflectivity, wind speed, and humidity), resulting in a classification accuracy of 85%. The cloud cover dynamic feature set extracts cloud cover coverage (0-100%) and cloud top height (2-15km) from meteorological satellite data. It uses a convolutional neural network (CNN) to extract dynamic features with a kernel size of 3×3, a stride of 1, and max pooling. The output is a 64-dimensional feature vector. The cloud cover change rate is calculated by analyzing 10 consecutive frames of cloud cover images (with a time interval of 5 minutes).

[0131] For example, cloud cover increased from 70% to 90%, a change rate of 4% / min. The cloud cover-lightning vertical hierarchical coupling feature set was obtained by analyzing the coupling relationship between cloud height (low layer 0-3km, mid layer 3-8km, high layer 8-15km) and lightning occurrence location. Principal component analysis (PCA) was used for dimensionality reduction, retaining 95% of the variance, resulting in a 5-dimensional coupling feature vector. Lightning monitoring data included the spatial coordinates (latitude, longitude, and altitude) and timestamps of 1000 lightning events, with a time resolution of 1 second and a spatial resolution of 0.1°. The above feature set (64-dimensional intensity features, 64-dimensional cloud cover dynamic features, and 5-dimensional coupling features) was concatenated into a 133-dimensional input vector, which was then input into an LSTM network. The network structure contained two layers of LSTM units (128 neurons per layer), with a time step of 10, an activation function of tanh, an optimizer of Adam, a learning rate of 0.001, and training for 100 epochs. The loss function was mean squared error (MSE), and the final MSE converged to 0.015. The LSTM outputs a lightning occurrence probability distribution (range 0-1, e.g., 0.75 for a certain area) and a spatial location coordinate sequence (e.g., latitude 39.9°, longitude 116.3°, altitude 5km). The probability is normalized using the softmax function to generate a lightning probability distribution feature set containing probability values ​​and corresponding coordinates.

[0132] S6. Based on the lightning probability distribution feature set and the cloud cover periodic feature vector, the time series analysis method is used to generate the lightning occurrence probability time series and the lightning time window range series, forming a lightning time window prediction feature set;

[0133] In this embodiment, S6, based on the lightning probability distribution feature set and the cloud cover periodic feature vector, a time series analysis method is used to generate a lightning occurrence probability time series and a lightning time window range series, forming a lightning time window prediction feature set, including:

[0134] S61. Obtain input data from the lightning probability distribution feature set and cloud cover periodic feature vector, and obtain standardized feature data input through data preprocessing;

[0135] S62. A long short-term memory network is used to train the standardized feature data input to generate a time series of lightning occurrence probability.

[0136] S63. Based on the lightning occurrence probability time series, determine the high-probability lightning occurrence time points by judging through preset thresholds, and obtain the lightning time window range sequence.

[0137] S64. For the lightning time window range sequence, cluster analysis is used to extract the periodic features of the time window and obtain the set of periodic features of the time window;

[0138] S65. Obtain feature vectors from the periodic feature set of the time window, and perform secondary training through a long short-term memory network to generate a lightning time window prediction feature set.

[0139] S66. If the matching degree between the lightning time window prediction feature set and the lightning occurrence probability time series is lower than the preset range, the parameters of the long short-term memory network are adjusted through iterative optimization to obtain an optimized prediction feature set.

[0140] S67. Based on the optimized prediction feature set, the time series is smoothed using the sliding window method to obtain the final lightning time window prediction feature set.

[0141] For example, based on the lightning probability distribution feature set and the cloud cover periodicity feature vector, a lightning occurrence probability time series and a lightning time window range series are first generated to form a lightning time window prediction feature set. Assume the lightning probability distribution feature set contains the hourly lightning occurrence probability within 24 hours, for example, [0.1, 0.15, 0.2, …, 0.05], a total of 24 values, representing the lightning probability for each hour of the day in a certain area. The cloud cover periodicity feature vector is the cloud cover coverage rate over a 12-hour period, for example, [0.3, 0.5, 0.7, ..., 0.2], reflecting the periodic trend of cloud cover changes. First, these two feature sets are normalized using the min-max normalization method, mapping the lightning probability and cloud cover value to the [0, 1] interval. For example, the lightning probability of 0.2 is normalized to (0.2-0.05) / (0.3-0.05)=0.6. Next, a Long Short-Term Memory (LSTM) network was constructed. The network structure consists of two LSTM layers, each with 64 neurons. The input dimension is 2 (probability of lightning and cloud cover), and the time step is 12 hours. The training data consists of normalized features from the past 30 days, with a batch size of 32. The Adam optimizer is used with a learning rate of 0.001, and the loss function is mean squared error. After training, the LSTM predicts the time series of lightning probability for the next 24 hours, for example, [0.12, 0.18, 0.25, ..., 0.08]. Subsequently, based on a probability threshold of 0.15, time points with a probability greater than 0.15 are extracted to form a time window range sequence, for example, [14:00-16:00, 20:00-22:00]. To ensure forecast accuracy, wind speed data within the time window is analyzed in conjunction with meteorological operational logic (assuming an average of 5 m / s). If the wind speed is greater than 4 m / s, the probability of lightning is confirmed to be increased. The weight of the time window is adjusted to generate the final forecast feature set, such as {[14:00-16:00, 0.8], [20:00-22:00, 0.9]}.

[0142] S7. Based on the periodic feature vector of cloud cover and the lightning time window range sequence, calculate the rhythm matching degree between the periodic fluctuation of cloud cover and the lightning time window, and generate a set of cloud cover-lightning time rhythm coupling features.

[0143] In this embodiment, S7, based on the periodic feature vector of cloud cover and the lightning time window range sequence, the rhythm matching degree between the periodic fluctuation of cloud cover and the lightning time window is calculated, and a set of cloud cover-lightning time rhythm coupling features is generated, including:

[0144] S71. Obtain the periodic feature vector of cloud cover and the time window range sequence of lightning. Use data preprocessing methods to normalize and denoise the data to obtain the cloud cover feature sequence and lightning time sequence.

[0145] S72. Using signal processing techniques, periodic fluctuation features are extracted from the cloud cover feature sequence to generate a cloud cover periodic feature set; using cross-correlation analysis, the temporal correlation between the cloud cover periodic feature set and the lightning time series is calculated to obtain the rhythm matching degree sequence.

[0146] S73. If there are values ​​in the rhythm matching degree sequence that exceed the preset range, the corresponding cloud cover periodic features are paired with the lightning time window to generate a preliminary coupling feature set. For the preliminary coupling feature set, the principal component analysis method is used to extract the main feature components to obtain a simplified cloud cover-lightning time rhythm coupling feature set.

[0147] S74. Through time series analysis, periodic patterns are detected in the simplified set of coupling features to generate the final set of cloud cover-lightning time rhythm coupling features.

[0148] For example, suppose the input data is a cloud cover periodicity feature vector, representing the periodic changes in cloud cover over 24 hours. The data is an hourly percentage of cloud cover, such as: [10, 15, 20, 30, 40, 50, 60, 70, 65, 55, 45, 35, 30, 25, 20, 15, 10, 10, 15, 20, 25, 30, 25, 20], with the unit being percentage, reflecting the fluctuation of cloud cover within a day. The lightning time window range sequence is the time period during which lightning occurs, such as: [3, 4, 5, 15, 16, 17], indicating that lightning occurs in the 3rd, 4th, 5th, 15th, 16th, and 17th hours, with a time resolution of 1 hour. First, a periodic feature vector of cloud cover is generated. The periodic features are extracted using Fast Fourier Transform (FFT), and the spectrum is calculated using Python's numpy.fft module, yielding a main period of 12 hours and an amplitude of 30. The cloud cover data is then standardized, with a mean of 35 and a variance of 400, resulting in a range of [-1.25, 1.75]. The lightning time window sequence is converted into a binary sequence with the same length as the cloud cover vector, setting the value to 1 for the hour of lightning occurrence and 0 for the rest, resulting in the sequence: [0, 0, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0]. Subsequently, cross-correlation analysis was performed using the `numpy.correlate` function in "full" mode to calculate the cross-correlation between the standardized cloud cover vector and the binary lightning sequence. The maximum cross-correlation coefficient was found to be 0.65, with a lag time of 2 hours, indicating that the peak cloud cover lags the lightning event by 2 hours. The rhythm matching degree was defined as the maximum cross-correlation coefficient, i.e., 0.65. A cloud cover-lightning time rhythm coupling feature set was generated, including a cross-correlation coefficient of 0.65, a lag time of 2 hours, a main period of 12 hours, a cloud cover amplitude of 30, and a lightning time window proportion of 6 / 24 = 0.25. The periodic fluctuations of cloud cover and the lightning time window were established with a time-series correlation through cross-correlation analysis. The coupling feature set can be used for subsequent meteorological prediction models to optimize the accuracy of lightning warnings.

[0149] S8. Based on the lightning time window prediction feature set and the lightning triggering factor dataset, use time series analysis to generate lightning movement path sequence and lightning duration sequence, forming a lightning trend prediction feature set.

[0150] In this embodiment, S8, based on the lightning time window prediction feature set and the lightning triggering factor dataset, a time series analysis method is used to generate a lightning movement path sequence and a lightning duration sequence, forming a lightning trend prediction feature set, including:

[0151] S81. Obtain the lightning time window prediction feature set and lightning triggering factor dataset, and perform standardization processing through the data preprocessing module to obtain the standardized feature set and standardized triggering factor dataset.

[0152] S82. Use a long short-term memory network to perform time series analysis on the standardized feature set to generate lightning movement path sequences and obtain path sequence data;

[0153] S83. Based on the standardized triggering factor dataset, a long short-term memory network is used to predict the duration of lightning, generating a lightning duration sequence and obtaining duration sequence data.

[0154] S84. By fusion of features, the path sequence data and duration sequence data are merged to construct a feature set for lightning trend prediction.

[0155] For example, the implementation method for generating lightning movement path sequences and lightning duration sequences using a Long Short-Term Memory (LSTM) network, based on a lightning time window prediction feature set and a lightning triggering factor dataset, to form a lightning trend prediction feature set is as follows. Assume the lightning time window prediction feature set contains meteorological data from 18:00 to 20:00 on July 22, 2025, including temperature 30.5°C, humidity 75%, wind speed 5.2 m / s, and air pressure 1002 hPa, with a spatial resolution of 1 km × 1 km grid. The lightning triggering factor dataset includes historical lightning occurrence time, location (latitude and longitude, e.g., 120.5°E, 30.2°N), intensity (current 50 kA), and cloud thickness of 500 m. First, the predicted feature set is preprocessed using Z-score standardization, transforming features such as temperature and humidity into a distribution with a mean of 0 and a standard deviation of 1. The calculation formula is z=(x-μ) / σ, where μ is the mean and σ is the standard deviation. For example, after standardization, temperature becomes (30.5-28) / 2.5=1.0. Next, a Long Short-Term Memory (LSTM) network model is constructed, consisting of 3 layers of LSTM with 64 units per layer. The input dimension is 4 (temperature, humidity, wind speed, and air pressure), and the time step is 12 (one data point every 10 minutes, covering 2 hours). The Adam optimizer is used with a learning rate of 0.001 and the loss function is mean squared error. During training, the standardized feature set and the triggering factor dataset are input. The model learns the relationship between the location and duration of lightning strikes and outputs a sequence of movement paths (e.g., from 120.5°E, 30.2°N to 120.6°E, 30.3°N, with a step size of 0.1°) and a duration sequence. For example, lightning lasting 30 minutes or 45 minutes. Path prediction uses a two-dimensional Gaussian distribution to estimate the movement probability, with the probability density function f(x,y)=1 / (2πσ). 2 )exp(-((x-μx) 2 +(y-μy) 2) / (2σ 2 The duration prediction is calculated using LSTM regression values, combined with the historical average lightning duration of 40 minutes, adjusting the predicted value range to 20-60 minutes. Finally, the path sequence and duration sequence are merged to generate a trend prediction feature set, such as {(120.6°E, 30.3°N, 30min), (120.7°E, 30.4°N, 45min)}.

[0156] S9. Based on the lightning trend prediction feature set, the cloud cover-lightning vertical hierarchical coupling feature set, and the cloud cover-lightning temporal rhythm coupling feature set, combined with the pre-established lightning-related cloud type labels, spatial and temporal features are extracted, and a spatiotemporal coupling feature set of cloud cover dynamic evolution and lightning trend is generated through weighted fusion.

[0157] In this embodiment, S9, based on the lightning trend prediction feature set, the cloud cover-lightning vertical hierarchical coupling feature set, and the cloud cover-lightning temporal rhythm coupling feature set, combined with pre-established lightning-related cloud type labels, spatial and temporal features are extracted, and a spatiotemporal coupling feature set of cloud cover dynamic evolution and lightning trend is generated through weighted fusion, including:

[0158] S91. Obtain the lightning trend prediction feature set, the cloud cover-lightning vertical hierarchical coupling feature set, and the cloud cover-lightning time rhythm coupling feature set. Normalize the feature set using a standardization processing method to obtain the normalized feature set.

[0159] S92. Based on the normalized feature set, a convolutional neural network is used to process the lightning occurrence probability and cloud coverage rate of the vertical cloud cover data in the lightning trend prediction feature set, extract spatial distribution features, and obtain a spatial feature set.

[0160] S93. Based on the normalized feature set, a long short-term memory network is used to process the time rhythm coupled data, extract the time dynamic features, and obtain the time feature set.

[0161] S94. The spatial feature set and the temporal feature set are integrated using a weighted fusion method to obtain a fused feature set;

[0162] S95. Based on the fusion feature set and combined with the pre-established lightning-related cloud type labels, a fully connected neural network is used for classification processing to obtain a spatiotemporal coupling feature set of cloud cover dynamic evolution and lightning trend.

[0163] For example, based on the lightning trend prediction feature set, the cloud cover-lightning vertical hierarchical coupling feature set, and the cloud cover-lightning temporal rhythm coupling feature set, combined with pre-established lightning-related cloud type labels, a convolutional neural network (CNN) is used to extract spatial features, and a long short-term memory network (LSTM) is used to extract temporal features. The method for generating a spatiotemporal coupling feature set of cloud cover dynamic evolution and lightning trend through weighted fusion is as follows: Assume the input lightning trend prediction feature set includes the probability of lightning occurrence (0.7), lightning frequency (3 times per minute), and location coordinates (latitude and longitude grid 100×100). The cloud cover-lightning vertical hierarchical coupling feature set includes cloud height (5000 meters), cloud cover coverage (80%), and vertical charge distribution (positive charge concentrated at the cloud top). The cloud cover-lightning temporal rhythm coupling feature set includes the time series cloud cover change rate (10% per hour) and lightning activity cycle (peak every 2 hours). The pre-established lightning-related cloud type labels are based on meteorological radar data and are divided into cumulonimbus clouds (80%), stratocumulus clouds (15%), and other cloud types (5%). First, spatial features are processed using a CNN, taking the lightning probability and cloud cover rate of a 100×100 grid as input. The convolution kernel size is set to 3×3, the stride is 1, and the pooling layer uses 2×2 max pooling to extract spatial correlation features, outputting a 64-dimensional feature map. The activation function is ReLU, and the calculation formula is f(x)=max(0,x). Next, the time series data is processed using an LSTM, taking the cloud cover change rate and lightning activity cycle as input. The number of hidden layer units is set to 128, the time step is 6 (representing 6 hours), the forget gate weight is initialized to 0.5, and the update formula is ft=σ(Wf·[ht-1,xt]+bf), generating a 32-dimensional time feature vector. To fuse the features, a weighted fusion algorithm is adopted, with a spatial feature weight of 0.6 and a temporal feature weight of 0.4. The fusion formula is F = 0.6·Fs + 0.4·Ft, where Fs is the CNN output and Ft is the LSTM output. The fused feature dimension is 96, which is input to a fully connected layer and outputs a spatiotemporally coupled feature set of cloud cover dynamic evolution and lightning trend (10 dimensions), corresponding to 10 lightning intensity levels.

[0164] S10. By performing cluster analysis on the spatiotemporal coupling feature set, the dynamic evolution pattern of lightning trend is obtained; based on the dynamic evolution pattern, the time series analysis method is used to predict the lightning trend, and the final lightning trend prediction result is obtained.

[0165] This embodiment discloses a lightning prediction and analysis method based on multi-source meteorological data fusion. Addressing operational challenges related to lightning occurrence probability, spatial location, time windows, and trend prediction under complex meteorological conditions, it achieves high-precision lightning prediction through multi-level feature extraction and coupling analysis. First, this invention extracts the spatiotemporal distribution features of cloud cover from multi-source meteorological data. It then uses convolutional neural networks and gradient operators to generate texture and edge features, and optimizes the cloud cover feature set through principal component analysis and K-means clustering. Combining humidity, temperature, and wind field data, it extracts dynamic evolution features of cloud cover using long short-term memory networks and optical flow methods. Finally, it constructs cloud cover-lightning coupling features using random forest algorithms and correlation analysis. Finally, it generates lightning probability distribution, time windows, and movement path predictions through weighted fusion of convolutional neural networks and long short-term memory networks. This invention significantly improves the accuracy and timeliness of lightning prediction through deep fusion and dynamic optimization of spatiotemporal features, providing efficient technical support for meteorological disaster early warning.

[0166] Example 2

[0167] Please see Figure 2 A lightning nowcasting system, based on the method described in one embodiment, the system comprising:

[0168] The first feature set module is used to obtain the spatial distribution and time series of cloud cover, extract the first cloud cover texture feature vector, and generate a cloud cover hierarchical feature matrix; reduce the dimensionality of the first cloud cover texture feature vector to generate the second cloud cover texture feature vector, and combine it with the cloud cover hierarchical feature matrix to generate a regional cloud cover distribution feature set through clustering.

[0169] The second feature set module is used to extract humidity, temperature, and wind field data to generate a lightning triggering factor dataset. It uses time series analysis and optical flow methods to analyze the regional cloud cover distribution feature set and form a dynamic cloud cover feature set.

[0170] The third feature set module is used to generate a lightning intensity level classification feature set based on the cloud cover dynamic feature set and cloud cover hierarchical feature matrix, combined with pre-established lightning-related cloud type labels.

[0171] The fourth feature set module is used to calculate the coupling weight coefficient between cloud cover vertical stratification and lightning intensity classification based on the lightning intensity level classification feature set and cloud cover stratification feature matrix, and generate the cloud cover-lightning vertical stratification coupling feature set.

[0172] The fifth feature set module is used to classify feature sets based on lightning intensity level, cloud cover dynamic feature set, and cloud cover-lightning vertical layer coupling feature set. Combined with lightning monitoring data, it uses time series analysis to generate lightning occurrence probability distribution and lightning spatial location coordinate sequence, forming lightning probability distribution feature set.

[0173] The sixth feature set module is used to generate a lightning occurrence probability time series and a lightning time window range series based on the lightning probability distribution feature set and the cloud cover periodic feature vector, using time series analysis to form a lightning time window prediction feature set.

[0174] The seventh feature set module is used to calculate the rhythm matching degree between cloud cover periodic fluctuations and lightning time windows based on the cloud cover periodic feature vector and the lightning time window range sequence, and generate a cloud cover-lightning time rhythm coupling feature set.

[0175] The eighth feature set module is used to generate lightning movement path sequences and lightning duration sequences based on the lightning time window prediction feature set and lightning triggering factor dataset, and to form a lightning trend prediction feature set.

[0176] The ninth feature set module is used to extract spatial and temporal features based on the lightning trend prediction feature set, the cloud cover-lightning vertical hierarchical coupling feature set, and the cloud cover-lightning temporal rhythm coupling feature set, combined with pre-established lightning-related cloud type labels, and generate a spatiotemporal coupling feature set of cloud cover dynamic evolution and lightning trend through weighted fusion.

[0177] The prediction module is used to obtain the dynamic evolution pattern of lightning trends by performing cluster analysis on the spatiotemporal coupling feature set; based on the dynamic evolution pattern, the time series analysis method is used to predict the lightning trend and obtain the final lightning trend prediction result.

[0178] In this embodiment, in order to better utilize the method described in one of the embodiments, this application proposes a lightning proximity trend forecasting system. Each module corresponds to each step of the above method, and its specific principle has been described above and will not be repeated here.

[0179] The above description is only a part of the embodiments of the present invention and does not limit the scope of protection of the present invention. Any equivalent device or equivalent process transformation made based on the content of the present invention specification and drawings, or direct or indirect application in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for predicting the imminent trend of lightning, characterized in that, The method includes: Obtain the spatial distribution and time series of cloud cover, extract the first cloud cover texture feature vector, and generate a cloud cover hierarchical feature matrix; reduce the dimensionality of the first cloud cover texture feature vector to generate the second cloud cover texture feature vector, and combine the cloud cover hierarchical feature matrix to generate a regionalized cloud cover distribution feature set through clustering. Humidity, temperature, and wind field data are extracted to generate a lightning triggering factor dataset. Time series analysis and optical flow methods are used to analyze the regional cloud cover distribution characteristics set and form a dynamic cloud cover characteristic set. Based on the dynamic cloud cover feature set and the cloud cover hierarchical feature matrix, combined with the pre-established lightning-related cloud type labels, a lightning intensity level classification feature set is generated. Based on the lightning intensity level classification feature set and the cloud cover stratification feature matrix, calculate the coupling weight coefficient between cloud cover vertical stratification and lightning intensity classification, and generate the cloud cover-lightning vertical stratification coupling feature set. Based on the lightning intensity level classification feature set, cloud cover dynamic feature set, and cloud cover-lightning vertical layer coupling feature set, combined with lightning monitoring data, the lightning occurrence probability distribution and lightning spatial location coordinate sequence are generated by time series analysis, forming a lightning probability distribution feature set. Based on the lightning probability distribution feature set and the cloud cover periodic feature vector, the time series analysis method is used to generate the lightning occurrence probability time series and the lightning time window range series, forming a lightning time window prediction feature set; Based on the periodic feature vector of cloud cover and the time window range sequence of lightning, the rhythm matching degree between the periodic fluctuation of cloud cover and the time window of lightning is calculated, and a set of cloud cover-lightning time rhythm coupling features is generated. Based on the lightning time window prediction feature set and the lightning triggering factor dataset, the lightning movement path sequence and lightning duration sequence are generated by time series analysis, forming a lightning trend prediction feature set. Based on the lightning trend prediction feature set, the cloud cover-lightning vertical hierarchical coupling feature set, and the cloud cover-lightning temporal rhythm coupling feature set, combined with the pre-established lightning-related cloud type labels, spatial and temporal features are extracted, and a spatiotemporal coupling feature set of cloud cover dynamic evolution and lightning trend is generated through weighted fusion. By performing cluster analysis on the spatiotemporal coupling feature set, the dynamic evolution pattern of lightning trends is obtained; based on the dynamic evolution pattern, the time series analysis method is used to predict the lightning trends, and the final lightning trend prediction results are obtained. Humidity, temperature, and wind field data were extracted to generate a lightning triggering factor dataset. Time-series analysis and optical flow methods were used to analyze the regionalized cloud cover distribution characteristics, forming a dynamic cloud cover feature set, including: Humidity, temperature and wind field data were obtained from multi-source meteorological data, and a lightning triggering factor dataset was generated through data cleaning and standardization. Long Short-Term Memory (LSTM) networks were used to analyze the time series data in the regionalized cloud cover distribution feature set, extracting the temporal variation features of cloud cover distribution to obtain the cloud cover time series feature set. By analyzing the cloud cover time series feature set using optical flow, the periodic changes and movement speed of cloud cover distribution are calculated, generating a periodic feature vector of cloud cover and a sequence of cloud cover movement speed. Based on the periodic feature vector of cloud cover and the sequence of cloud cover movement speed, and by integrating humidity, temperature and wind field data, a set of dynamic evolution features of cloud cover is constructed.

2. The lightning proximity trend forecasting method as described in claim 1, characterized in that, Based on the dynamic cloud cover feature set and the cloud cover hierarchical feature matrix, combined with pre-established lightning-related cloud type labels, a lightning intensity level classification feature set is generated, including: Obtain the dynamic feature set and hierarchical feature matrix of cloud amount, and obtain the standardized feature dataset through data preprocessing; if there are missing values ​​in the standardized feature dataset, the mean imputation method is used to obtain the complete feature dataset. Based on the complete feature dataset and pre-established lightning-related cloud type labels, a classification model is trained using the random forest algorithm to obtain a lightning intensity level classifier. The lightning intensity level classifier is then used to predict the complete feature dataset to generate an initial lightning intensity level classification result. If the confidence level of the initial lightning intensity level classification result is lower than the threshold, the classification result is adjusted using an ensemble evaluation method to obtain an optimized lightning intensity level classification result. Based on the optimized lightning intensity level classification results, high-confidence classification features are extracted to generate a lightning intensity level classification feature set. By combining the lightning intensity level classification feature set with the spatiotemporal distribution of cloud cover dynamic features and hierarchical features, a spatiotemporal feature mapping of lightning intensity level is generated.

3. The lightning proximity trend forecasting method as described in claim 1, characterized in that, Based on the lightning intensity level classification feature set and the cloud cover stratification feature matrix, the coupling weight coefficient between cloud cover vertical stratification and lightning intensity classification is calculated, generating a cloud cover-lightning vertical stratification coupling feature set, including: From the lightning intensity level classification feature set and cloud cover stratification feature matrix, the data matrix of lightning intensity level and cloud cover stratification features is obtained. The main feature components are extracted by principal component analysis to obtain the feature dimensionality reduction set. Based on the feature dimensionality reduction set, calculate the Pearson correlation coefficient between lightning intensity level and cloud cover stratification features to determine the correlation analysis results; if the absolute value of the Pearson correlation coefficient in the correlation analysis results is greater than the preset range, then mark the corresponding lightning intensity level and cloud cover stratification feature pair as a highly correlated feature pair to obtain a set of highly correlated feature pairs. For the set of highly correlated feature pairs, a weighted linear combination method is used to calculate the coupling weight coefficients between lightning intensity level and cloud cover stratification features, thus obtaining a set of coupling weight coefficients. From the set of coupling weight coefficients, feature pairs with weight coefficients greater than a preset threshold are obtained to generate a set of cloud cover-lightning vertical stratification coupling features. Based on the cloud cover-lightning vertical layered coupling feature set, the K-means clustering algorithm is used to classify the feature set to obtain the vertical layered structure of cloud cover-lightning coupling features. From the vertical layered structure, the correspondence between cloud cover distribution patterns and lightning activity intensity is extracted to generate the final cloud cover-lightning vertical layered coupling feature set.

4. The lightning proximity trend forecasting method as described in claim 1, characterized in that, Based on the lightning intensity level classification feature set, cloud cover dynamic feature set, and cloud cover-lightning vertical stratification coupling feature set, combined with lightning monitoring data, a time series analysis method is used to generate a lightning occurrence probability distribution and a lightning spatial location coordinate sequence, forming a lightning probability distribution feature set, including: We obtained the lightning intensity level classification feature set, cloud cover dynamic feature set, and cloud cover-lightning vertical hierarchical coupling feature set. We extracted the raw spatiotemporal data from the lightning monitoring data and used data preprocessing methods to obtain a standardized feature dataset. By using a standardized feature dataset and a long short-term memory network, a lightning occurrence probability prediction model is trained to obtain the lightning occurrence probability distribution. Based on the probability distribution of lightning occurrence and combined with spatial location coordinates, a convolutional neural network is used to extract the spatial distribution features of lightning and obtain the sequence of lightning spatial location coordinates. If the probability distribution of lightning occurrence exceeds the preset range, then through spatiotemporal sequence analysis, the spatial location coordinate sequence of lightning is fused to generate a set of lightning probability distribution features.

5. The lightning proximity trend forecasting method as described in claim 1, characterized in that, Based on the lightning probability distribution feature set and the cloud cover periodic feature vector, time series analysis is used to generate a lightning occurrence probability time series and a lightning time window range series, forming a lightning time window prediction feature set, including: Input data is obtained from the lightning probability distribution feature set and cloud cover periodic feature vector, and standardized feature data input is obtained through data preprocessing; A long short-term memory network is used to train standardized feature data input to generate a time series of lightning occurrence probabilities; Based on the lightning occurrence probability time series, and by judging through preset thresholds, the time points with high probability of lightning occurrence are determined, and the lightning time window range sequence is obtained; For the lightning time window range sequence, cluster analysis is used to extract the periodic features of the time window, and a set of periodic features of the time window is obtained; Feature vectors are obtained from the periodic feature set of the time window, and then a second training is performed through a long short-term memory network to generate a lightning time window prediction feature set. If the matching degree between the lightning time window prediction feature set and the lightning occurrence probability time series is lower than the preset range, the parameters of the long short-term memory network are adjusted through iterative optimization to obtain an optimized prediction feature set. Based on the optimized prediction feature set, the time series is smoothed using the sliding window method to obtain the final lightning time window prediction feature set.

6. The lightning proximity trend forecasting method as described in claim 1, characterized in that, Based on the periodic feature vector of cloud cover and the sequence of lightning time windows, the rhythm matching degree between the periodic fluctuations of cloud cover and the lightning time windows is calculated, generating a set of cloud cover-lightning time rhythm coupling features, including: The periodic feature vector of cloud cover and the time window range sequence of lightning are obtained. Data preprocessing methods are used to normalize and denoise the data to obtain the cloud cover feature sequence and lightning time sequence. By using signal processing techniques, periodic fluctuation features are extracted from cloud cover feature sequences to generate a cloud cover periodic feature set. Cross-correlation analysis is then used to calculate the temporal correlation between the cloud cover periodic feature set and the lightning time series, resulting in a rhythm matching degree sequence. If a value in the rhythm matching degree sequence exceeds the preset range, the corresponding cloud cover periodic feature is paired with the lightning time window to generate a preliminary coupling feature set. For the preliminary coupling feature set, the principal component analysis method is used to extract the main feature components to obtain a simplified cloud cover-lightning time rhythm coupling feature set. By analyzing time series data, periodic patterns are detected in a simplified set of coupled features, generating the final set of cloud cover-lightning time rhythm coupled features.

7. The lightning proximity trend forecasting method as described in claim 1, characterized in that, Based on the lightning time window prediction feature set and the lightning triggering factor dataset, time series analysis is used to generate lightning movement path sequences and lightning duration sequences, forming a lightning trend prediction feature set, including: Obtain the lightning time window prediction feature set and lightning triggering factor dataset, and perform standardization processing through the data preprocessing module to obtain the standardized feature set and standardized triggering factor dataset; A long short-term memory network is used to perform time series analysis on a standardized feature set to generate lightning movement path sequences, thus obtaining path sequence data. Based on a standardized triggering factor dataset, a long short-term memory network is used to predict the duration of lightning, generating a lightning duration sequence and obtaining duration sequence data. By fusion of features, path sequence data and duration sequence data are merged to construct a feature set for lightning trend prediction.

8. The lightning proximity trend forecasting method as described in claim 1, characterized in that, Based on the lightning trend prediction feature set, the cloud cover-lightning vertical hierarchical coupling feature set, and the cloud cover-lightning temporal rhythm coupling feature set, combined with pre-established lightning-related cloud type labels, spatial and temporal features are extracted. Through weighted fusion, a spatiotemporal coupling feature set of cloud cover dynamic evolution and lightning trend is generated, including: The lightning trend prediction feature set, the cloud cover-lightning vertical hierarchical coupling feature set, and the cloud cover-lightning temporal rhythm coupling feature set are obtained. The feature sets are normalized using a standardization processing method to obtain a normalized feature set. Based on the normalized feature set, a convolutional neural network is used to process the lightning occurrence probability and cloud coverage of the vertical cloud cover data in the lightning trend prediction feature set to extract spatial distribution features and obtain a spatial feature set. Based on the normalized feature set, a long short-term memory network is used to process the time rhythm coupled data, extract the time dynamic features, and obtain the time feature set; A weighted fusion method is used to integrate the spatial feature set and the temporal feature set to obtain a fused feature set; Based on the fusion feature set and combined with the pre-established lightning-related cloud type labels, a fully connected neural network is used for classification to obtain a spatiotemporal coupled feature set of cloud cover dynamic evolution and lightning trend.

9. A lightning imminent trend forecasting system, characterized in that, Based on the method of any one of claims 1-8, the system comprises: The first feature set module is used to obtain the spatial distribution and time series of cloud cover, extract the first cloud cover texture feature vector, and generate a cloud cover hierarchical feature matrix; reduce the dimensionality of the first cloud cover texture feature vector to generate the second cloud cover texture feature vector, and combine it with the cloud cover hierarchical feature matrix to generate a regional cloud cover distribution feature set through clustering. The second feature set module is used to extract humidity, temperature, and wind field data to generate a lightning triggering factor dataset. It uses time series analysis and optical flow methods to analyze the regional cloud cover distribution feature set and form a dynamic cloud cover feature set. The third feature set module is used to generate a lightning intensity level classification feature set based on the cloud cover dynamic feature set and cloud cover hierarchical feature matrix, combined with pre-established lightning-related cloud type labels. The fourth feature set module is used to calculate the coupling weight coefficient between cloud cover vertical stratification and lightning intensity classification based on the lightning intensity level classification feature set and cloud cover stratification feature matrix, and generate the cloud cover-lightning vertical stratification coupling feature set. The fifth feature set module is used to classify feature sets based on lightning intensity level, cloud cover dynamic feature set, and cloud cover-lightning vertical layer coupling feature set. Combined with lightning monitoring data, it uses time series analysis to generate lightning occurrence probability distribution and lightning spatial location coordinate sequence, forming lightning probability distribution feature set. The sixth feature set module is used to generate a lightning occurrence probability time series and a lightning time window range series based on the lightning probability distribution feature set and the cloud cover periodic feature vector, using time series analysis to form a lightning time window prediction feature set. The seventh feature set module is used to calculate the rhythm matching degree between cloud cover periodic fluctuations and lightning time windows based on the cloud cover periodic feature vector and the lightning time window range sequence, and generate a cloud cover-lightning time rhythm coupling feature set. The eighth feature set module is used to generate lightning movement path sequences and lightning duration sequences based on the lightning time window prediction feature set and lightning triggering factor dataset, and to form a lightning trend prediction feature set. The ninth feature set module is used to extract spatial and temporal features based on the lightning trend prediction feature set, the cloud cover-lightning vertical hierarchical coupling feature set, and the cloud cover-lightning temporal rhythm coupling feature set, combined with pre-established lightning-related cloud type labels, and generate a spatiotemporal coupling feature set of cloud cover dynamic evolution and lightning trend through weighted fusion. The prediction module is used to obtain the dynamic evolution pattern of lightning trends by performing cluster analysis on the spatiotemporal coupling feature set; based on the dynamic evolution pattern, the time series analysis method is used to predict the lightning trend and obtain the final lightning trend prediction result.

Images