Method and system for improving thunder and lightning early warning accuracy and medium

Abstract

The invention discloses a method and system for improving thunder and lightning early warning accuracy and a medium, and relates to the technical field of meteorological monitoring and early warning, and the method comprises the steps: connecting a three-dimensional monitoring network, and obtaining multi-source monitoring data; fusing multi-source monitoring data in real time to generate a three-dimensional thunder and lightning risk thermodynamic diagram; adopting a space-time convolutional neural network embedded with a physical constraint mechanism to predict lightning occurrence probability and motion trail in a future time range based on the thermodynamic diagram; and according to the prediction confidence coefficient, carrying out early warning grading on the lightning occurrence probability and the motion trail, and updating an early warning area geofence based on the real-time lightning positioning data. The technical problems that the predicted lightning occurrence probability and motion trail are inaccurate and the requirements of accurate early warning and effective prevention and control cannot be met due to the fact that complex lightning activity information is difficult to comprehensively capture in a traditional lightning early warning technology are solved, and the purposes of improving the lightning occurrence probability and motion trail prediction accuracy and improving the lightning early warning efficiency are achieved. The technical effects of accurate early warning and effective prevention and control of lightning disasters are achieved.

Description

Technical Field

[0001] This invention relates to the field of meteorological monitoring and early warning technology, and in particular to a method, system and medium for improving the accuracy of lightning early warning. Background Technology

[0002] Lightning disasters pose a serious threat to people's lives, production activities, and equipment operation. Improving the accuracy of lightning warnings is crucial for disaster prevention and mitigation. Current lightning warning technologies largely rely on data from single or limited monitoring devices, such as traditional radar and limited electric field monitoring equipment. While these methods have proven effective in localized, small-scale warnings, their limitations have become apparent as the demands for timeliness and accuracy increase, particularly when dealing with large-scale lightning activity under complex weather conditions. Due to the complexity of lightning formation and movement, traditional warning methods struggle to comprehensively capture lightning-related information, leading to inaccurate predictions of lightning occurrence probabilities and trajectories, and unreasonable warning grading, ultimately failing to meet the needs for accurate early warning and effective prevention of lightning disasters. Summary of the Invention

[0003] This application provides a method, system, and medium for improving the accuracy of lightning warnings, which addresses the technical problem that traditional lightning warning technologies are unable to fully capture complex lightning activity information, resulting in inaccurate predictions of lightning occurrence probability and trajectory, and thus failing to meet the needs of accurate warnings and effective prevention and control.

[0004] The first aspect of this application provides a method for improving the accuracy of lightning warnings. The method includes: connecting a three-dimensional monitoring network to acquire multi-source monitoring data, including radar data, atmospheric electric field data, and lightning electromagnetic data; fusing the multi-source monitoring data in real time to generate a three-dimensional lightning risk heat map; using a spatiotemporal convolutional neural network with embedded physical constraints to predict the probability of lightning occurrence and its trajectory within a future time range based on the three-dimensional lightning risk heat map; classifying the lightning occurrence probability and trajectory into warning levels according to the prediction confidence level; and updating the geofence of the warning area based on real-time lightning location data.

[0005] A second aspect of this application provides a system for improving the accuracy of lightning warnings. The system includes: a multi-source monitoring data acquisition module for connecting to a three-dimensional monitoring network and acquiring multi-source monitoring data, including radar data, atmospheric electric field data, and lightning electromagnetic data; a three-dimensional lightning risk heat map acquisition module for real-time fusion of the multi-source monitoring data to generate a three-dimensional lightning risk heat map; a lightning occurrence probability prediction module for using a spatiotemporal convolutional neural network with embedded physical constraints to predict the probability of lightning occurrence and its trajectory within a future time range based on the three-dimensional lightning risk heat map; and a warning area update module for classifying the lightning occurrence probability and trajectory according to the prediction confidence level, and updating the geofence of the warning area based on real-time lightning location data.

[0006] A third aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements a method for improving the accuracy of lightning warnings provided in this application.

[0007] One or more technical solutions provided in this application have at least the following technical effects or advantages: This application utilizes a three-dimensional monitoring network consisting of a millimeter-wave phased array radar array, a distributed atmospheric electric field meter sensor network, and a wide-area lightning location device to acquire multi-source monitoring data and fuse them in real time to generate a three-dimensional lightning risk heat map. It employs a spatiotemporal convolutional neural network with embedded physical constraints to predict the probability of future lightning occurrences and their trajectories. Furthermore, it classifies warnings based on prediction confidence levels and updates the geofence of the warning area based on real-time lightning location data. This improves the accuracy of lightning warnings, making the warning results more reliable and achieving the technical goal of enhancing the accuracy of lightning occurrence probability and trajectory prediction, thus meeting the requirements for precise early warning and effective prevention and control of lightning disasters. Attached Figure Description

[0008] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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.

[0009] Figure 1 This is a flowchart illustrating a method for improving the accuracy of lightning warnings provided in an embodiment of this application.

[0010] Figure 2 This is a schematic diagram of the structure of a system for improving the accuracy of lightning warnings provided in an embodiment of this application.

[0011] Figure labeling: Module 1 for multi-source monitoring data acquisition, Module 2 for 3D lightning risk heat map acquisition, Module 3 for lightning occurrence probability prediction, and Module 4 for early warning area update. Detailed Implementation

[0012] This application provides a method, system, and medium for improving the accuracy of lightning warnings, which addresses the technical problem that traditional lightning warning technologies are unable to fully capture complex lightning activity information, resulting in inaccurate predictions of lightning occurrence probability and trajectory, and thus failing to meet the needs of accurate warnings and effective prevention and control.

[0013] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0014] It should be noted that the terms "first," "second," etc., in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or modules not explicitly listed or inherent to such processes, methods, products, or devices.

[0015] Example 1, as Figure 1 As shown, a method for improving the accuracy of lightning warnings includes: Step A100: Connect to the three-dimensional monitoring network and acquire multi-source monitoring data, including radar data, atmospheric electric field data, and lightning electromagnetic data.

[0016] Specifically, existing lightning warning technologies largely rely on single-type monitoring equipment, such as traditional weather radar or a few electric field monitors, resulting in limited data dimensions and an inability to comprehensively reflect the complex process of lightning formation and development. The first step is to construct and connect a three-dimensional monitoring network covering the target area. This network consists of millimeter-wave phased array radar arrays deployed in the target area, a distributed atmospheric electric field sensor network, and wide-area lightning location devices working together. It establishes a connection with the processing center through a unified data transmission protocol, ensuring that monitoring data from all types of equipment can be transmitted to the processing system in real time and stably.

[0017] After connecting to the three-dimensional monitoring network, the system begins to synchronously acquire multi-source monitoring data. The millimeter-wave phased array radar captures radar data by transmitting and receiving millimeter-wave signals, including radar reflectivity factor reflecting the density of cloud clusters, vertical liquid water content indicating the total amount of liquid water in the cloud, and velocity divergence reflecting differences in airflow motion. The distributed atmospheric electric field sensor network continuously records atmospheric electric field data through sensors distributed in different locations, from which the rate of change of electric field intensity reflecting the rate of charge accumulation and the change of electric field polarity indicating charge polarity conversion can be extracted. The wide-area lightning location device acquires lightning electromagnetic data by monitoring the electromagnetic signals generated by lightning, including the frequency of lightning per unit time, the discharge intensity of a single lightning strike, and the lightning type that distinguishes between cloud-to-ground lightning.

[0018] These multi-source monitoring data characterize lightning activity from different perspectives: radar data reflects the macroscopic physical state of clouds, atmospheric electric field data reflects the charge conditions for lightning formation, and lightning electromagnetic data records the discharge activities that have occurred. The three complement each other to form a comprehensive monitoring of lightning activity.

[0019] By connecting to a three-dimensional monitoring network, radar data including radar reflectivity factor, vertical liquid water content, and velocity divergence, atmospheric electric field data including electric field intensity change rate and electric field polarity change, and lightning electromagnetic data including lightning frequency, lightning intensity, and lightning type are acquired simultaneously. This achieves the effect of comprehensively capturing multi-dimensional information on lightning activity and compensating for the one-sidedness of traditional single-device monitoring data.

[0020] Step A200: In real time, the multi-source monitoring data is fused to generate a three-dimensional lightning risk heat map.

[0021] Optionally, a method for generating a three-dimensional lightning risk heat map by real-time fusion of multi-source monitoring data includes extracting specific features from radar, atmospheric electric field, and lightning electromagnetic data, aligning and fusing them according to spatiotemporal relationships to obtain fused features, and then using them to predict risk values ​​to generate a heat map. Specific steps are detailed in A210-A250.

[0022] Step A300: Using a spatiotemporal convolutional neural network with embedded physical constraints, the probability of lightning occurrence and its trajectory within a future time range are predicted based on the three-dimensional lightning risk heat map.

[0023] In one embodiment of this application, a spatiotemporal convolutional neural network with an embedded physical constraint mechanism and a spatiotemporal convolutional layer is used to predict the probability of future lightning occurrence (between 0 and 1) and the trajectory (including speed and direction) of the lightning based on a three-dimensional lightning risk heat map. The specific steps are described in detail in A310-A350.

[0024] Step A400: Based on the prediction confidence level, classify the probability of lightning occurrence and the trajectory of lightning movement into early warning levels, and update the geofence of the early warning area based on real-time lightning location data.

[0025] Specifically, firstly, the prediction confidence is calculated based on the output of the spatiotemporal convolutional neural network, and the probability of lightning occurrence and its trajectory are classified into early warning levels. The method of dividing different early warning level areas and dynamically adjusting the boundaries is explained in detail in A410-A430.

[0026] Next, establish a real-time data connection with the wide-area lightning location device to continuously receive the latest lightning location data, including the precise latitude and longitude of the lightning occurrence, altitude, occurrence time, lightning intensity (e.g., a single lightning strike with an intensity of 30kA or 60kA), lightning type (e.g., cloud lightning or ground lightning), and lightning frequency per unit time.

[0027] Upon receiving real-time lightning location data, it is spatially compared with the current warning area geofence. For example, if the current warning fence defines the area as 116.2°–116.5°E and 39.9°–40.1°N, and three newly detected lightning strikes occurred at 116.55°E and 40.05°N, outside the fence, and these lightning strikes all exceeded 50kA in intensity and occurred frequently, it indicates that lightning activity has exceeded the original warning range. Simultaneously, the lightning's trajectory characteristics are analyzed, and its direction and speed are extracted from the real-time data to determine the areas it may affect in the future.

[0028] Finally, based on the comparison results and trajectory analysis, the geofence of the warning area is dynamically updated. If the lightning moves northwest at a stable speed, the northwest boundary of the fence needs to be expanded accordingly, for example, from 116.2°E to 116.1°E, to cover areas that may be affected in the future. If no new lightning activity is detected within the original fence for an extended period, and the frequency of lightning in the surrounding area decreases significantly, the fence area should be appropriately reduced to minimize unnecessary warning coverage. The updated fence must accurately match the real-time spatial distribution and movement trends of lightning to ensure that the boundary of each area reflects the actual impact range of current lightning activity.

[0029] By receiving and analyzing lightning location data in real time, including location, intensity, frequency, and type, and comparing it with the current warning fence, and dynamically adjusting the fence boundary in conjunction with the lightning movement trajectory, the effect of synchronizing the geographic fence of the warning area with the actual lightning activity in real time is achieved, thereby improving the accuracy of the warning area.

[0030] Furthermore, step A100 in the method provided in this application embodiment includes: A110: The three-dimensional monitoring network includes: a millimeter-wave phased array radar array deployed in the target area, a distributed atmospheric electric field sensor network, and a wide-area lightning location device.

[0031] Specifically, a specialized three-dimensional monitoring network is deployed in the target area. This network consists of a millimeter-wave phased array radar array, a distributed atmospheric electric field sensor network, and a wide-area lightning location device. The millimeter-wave phased array radar array captures radar data, including radar reflectivity factor, vertical liquid water content, and velocity divergence, which reflect the physical state and motion characteristics of clouds. The distributed atmospheric electric field sensor network focuses on collecting atmospheric electric field data, from which information such as the rate of change of electric field intensity and the change of electric field polarity can be extracted, directly correlated with the charge accumulation during lightning formation. The wide-area lightning location device primarily acquires lightning electromagnetic data, extracting key features such as lightning frequency, lightning intensity, and lightning type, and recording the discharge activity that has occurred.

[0032] These devices work collaboratively within the target area to form a multi-dimensional, all-round monitoring coverage. The millimeter-wave phased array radar array, through its array layout, monitors clouds at different altitudes and ranges, ensuring the spatial continuity of radar data; the distributed atmospheric electric field meter sensor network is distributed at a certain density, enabling atmospheric electric field data to reflect differences in electric field changes in different local areas; and the wide-area lightning location device covers a wider geographical area, ensuring that lightning electromagnetic data can capture all lightning activity within the region. All three operate simultaneously, continuously collecting their respective monitoring data, providing comprehensive and real-time raw material for subsequent multi-source data fusion.

[0033] By deploying a three-dimensional monitoring network consisting of a millimeter-wave phased array radar array, a distributed atmospheric electric field sensor network, and a wide-area lightning location device, comprehensive information from multiple sources, including radar data, atmospheric electric field data, and lightning electromagnetic data, is acquired. This lays a comprehensive and accurate data source foundation for improving the accuracy of lightning warnings, achieving the effect of breaking through the limitations of traditional single-device monitoring and realizing multi-dimensional capture of lightning activity.

[0034] Furthermore, step A200 in the method provided in this application embodiment includes: A210: Based on the radar data, obtain the radar reflectivity factor, vertical liquid water content, and velocity divergence.

[0035] A220: Extract the rate of change of electric field intensity and the change of electric field polarity from atmospheric electric field data.

[0036] A230: Extract lightning frequency, lightning intensity, and lightning type from lightning electromagnetic data.

[0037] A240: Align and merge different data sources according to their temporal and spatial relationships to obtain fused features.

[0038] A250: Utilizes fusion features to predict lightning risk, obtains lightning risk values ​​for various temporal and spatial relationships, and generates a three-dimensional lightning risk heat map.

[0039] Optionally, key features are extracted from various monitoring data. For radar data, the focus is on acquiring radar reflectivity factor, vertical liquid water content, and velocity divergence. Radar reflectivity factor reflects the concentration and size of particles in the cloud, vertical liquid water content reflects the total amount of liquid water in the cloud, and velocity divergence indicates the convergence and divergence state of airflow. These features collectively characterize the macroscopic physical properties of the cloud. From atmospheric electric field data, the rate of change of electric field intensity and the change of electric field polarity are extracted. The rate of change of electric field intensity is directly related to the speed of charge accumulation, while the change of electric field polarity may indicate the adjustment of charge structure before lightning discharge. From lightning electromagnetic data, lightning frequency, lightning intensity, and lightning type are extracted. Lightning frequency reflects the activity level of discharge per unit time, lightning intensity reflects the energy of a single discharge, and lightning type (such as cloud-to-ground lightning) can distinguish the spatial location of the discharge.

[0040] After feature extraction, different data sources need to be aligned and fused according to their temporal and spatial relationships. Temporally, all data is calibrated using a unified timestamp; for example, the radar data acquisition time, the atmospheric electric field meter recording time, and the lightning location device monitoring time are synchronized to the same time coordinate system to ensure the correspondence of different data in the temporal dimension. Spatially, a unified spatial reference system is established based on the geographic coordinates and altitude information of the target area, so that the area covered by radar data, the deployment location of the atmospheric electric field meter, and the location of lightning occurrence can all be mapped to the same spatial grid. Through this spatiotemporal alignment, the originally scattered radar features, atmospheric electric field features, and lightning electromagnetic features are integrated into a set of interconnected fused features, preserving the unique information of each data source while forming a multi-dimensional, multi-scale characterization of lightning activity.

[0041] Finally, a method for generating a three-dimensional lightning risk heat map by using fusion features to predict lightning risk includes aligning multi-source monitoring data according to the collection timestamp, projecting it onto a unified three-dimensional spatial grid, establishing a model to calculate the risk value of each grid point and generating a three-dimensional lightning risk heat map. The specific steps are explained in detail in A251-A254.

[0042] By extracting specific features from radar data, atmospheric electric field data, and lightning electromagnetic data, and aligning and fusing them according to temporal and spatial relationships to obtain fused features, the system achieves the effect of integrating multi-source information, eliminating the impact of data fragmentation, and providing comprehensive and coherent input for subsequent lightning risk assessment.

[0043] Furthermore, step A250 in the method provided in this application embodiment includes: A251: Align multi-source monitoring data according to the collection timestamp.

[0044] A252: Project all monitoring data into a unified three-dimensional spatial grid. Radar data is filled into the grid using interpolation methods, while atmospheric electric field data and lightning electromagnetic data are directly mapped to the grid based on location information.

[0045] A253: Establish a lightning risk assessment model, input the fused feature data into the model, and calculate the lightning risk value for each grid point.

[0046] A254: Generate the three-dimensional lightning risk heat map according to the lightning risk value of each grid point.

[0047] Specifically, the first step is to unify the time dimension of the multi-source monitoring data. Radar data, atmospheric electric field data, and lightning electromagnetic data are synchronized to a unified time coordinate system according to their respective acquisition timestamps. For example, the acquisition time of all data is calibrated in milliseconds to ensure that different data sources correspond one-to-one in the time dimension, laying a foundation for time consistency for subsequent spatial fusion.

[0048] After time alignment, all monitoring data are projected onto a unified three-dimensional spatial grid. This grid can be divided according to the latitude, longitude, and altitude of the target area; for example, each grid cell consists of 0.01 degrees of longitude, 0.01 degrees of latitude, and 100 meters of altitude. For radar data, since sampling points may have spatial intervals, interpolation methods, such as linear interpolation, are used to fill the data into each grid, ensuring continuous coverage of radar data in three-dimensional space. Atmospheric electric field data comes from distributed sensors, each corresponding to specific latitude, longitude, and altitude, and can be directly mapped to the corresponding grid based on its location information. Lightning electromagnetic data records the precise location of lightning strikes and is similarly directly matched to the corresponding grid based on latitude, longitude, and altitude information, achieving unified integration of all data in the spatial dimension.

[0049] After achieving spatiotemporal unification, a lightning risk assessment model needs to be established. This process requires combining multi-source fusion characteristics with historical lightning data and proceeding in stages: Step a: Collect and organize historical training data. The input data consists of features derived from the fusion of historical multi-source monitoring data, including radar reflectivity factor, vertical liquid water content, velocity divergence, electric field intensity change rate, electric field polarity change, lightning frequency, lightning intensity, and lightning type in each three-dimensional spatial grid. The label data consists of historical records of actual lightning activity at corresponding grid points within the same time range, such as whether lightning occurred, the probability range of lightning occurrence, and the intensity level. These are used to construct "feature-label" training sample pairs.

[0050] Step b: Preprocess the input features. Standardize the fused feature data to unify the dimensions of different features, such as mapping radar reflectivity factor and electric field intensity change rate to the 0-1 interval, eliminating the impact of magnitude differences on model training; for grid features with missing values, use interpolation based on neighboring grid features to fill in the missing values, ensuring the continuity of the 3D spatial grid data.

[0051] Step c: Design the model structure. Based on the nonlinear correlation between lightning risk and multi-source features, select a model architecture suitable for handling high-dimensional spatiotemporal features, such as gradient boosting trees or deep learning neural networks. The model should be able to capture the correlation between spatial grids (such as the conduction of electric field changes between adjacent grids) and time series features (such as the trend of lightning frequency changes over time) in order to accurately map the relationship between features and risk values.

[0052] Step d: Model training and optimization. The preprocessed training samples are divided into training and validation sets. The training set data drives model learning by minimizing the error (e.g., mean squared error) between the predicted risk value and the actual lightning activity label. Model parameters (e.g., the depth of the tree model, the weights of the neural network) are iteratively adjusted. The validation set is used to evaluate model performance. If overfitting exists (e.g., low error on the training set but high error on the validation set), optimization is performed through regularization, reducing model complexity, etc., until the model's prediction accuracy on the validation set stabilizes.

[0053] Step e: Determine the model output format. The trained model must be able to receive fused feature data from a real-time 3D spatial grid, calculate the lightning risk value for each grid point using a built-in algorithm, and output a value from 0 to 100. The higher the value, the higher the risk, providing a quantitative basis for the subsequent generation of a 3D lightning risk heat map.

[0054] Finally, a three-dimensional lightning risk heat map is generated based on the lightning risk value of each grid point. Different colors are used to indicate the level of risk, such as red for high risk, yellow for medium risk, and blue for low risk, to visually present the distribution of lightning risk in the target area at different times and locations.

[0055] By aligning multi-source monitoring data according to timestamps and projecting it onto a unified three-dimensional spatial grid, inputting it into a lightning risk assessment model to calculate the risk value of grid points and generate a three-dimensional lightning risk heat map, the effect of eliminating data spatiotemporal misalignment and achieving accurate spatial characterization of lightning risk is achieved.

[0056] Furthermore, step A240 in the method provided in this application embodiment includes: A241: Based on the lightning occurrence cycle nodes, the fusion weight of each data source is dynamically adjusted, and feature fusion is performed based on the fusion weight. Specifically, when it is in the initial development stage of a thunderstorm, the weight of atmospheric electric field data is ≥0.7; when it is in the mature stage of a thunderstorm, the weight of radar echo intensity data is ≥0.6; and when it is the first lightning occurrence, the weight of lightning location data is automatically increased to 0.8.

[0057] Specifically, in existing lightning warning technologies, multi-source data fusion often employs fixed weight allocation, failing to consider the differences in core driving factors at different stages of lightning development. This results in the fused features failing to accurately reflect the essence of lightning activity at each stage. For example, the key influencing factors differ between the initial and mature stages of a thunderstorm; fixed weights weaken data features that dominate the current stage, affecting the fusion effect.

[0058] First, the periodic nodes of lightning occurrence are identified, and the current stage is determined by real-time monitoring data: when the atmospheric electric field begins to change continuously and the radar echo is weak and scattered, it is determined to be the initial development stage of thunderstorm; when the radar echo intensity increases significantly, the vertical liquid water content rises sharply, and the cloud structure tends to stabilize, it is determined to be the mature stage of thunderstorm; when the first lightning electromagnetic signal is detected, it is marked as the first lightning occurrence.

[0059] After identifying the periodic nodes, the fusion weights of each data source are dynamically adjusted. In the initial development stage of a thunderstorm, charge accumulation is the core process of lightning formation. At this time, atmospheric electric field data, such as the rate of change of electric field intensity and polarity change, can better reflect early characteristics. Therefore, its weight is adjusted to ≥0.7, while the weights of radar data and lightning electromagnetic data are reduced accordingly. In the mature stage of a thunderstorm, the physical state of the cloud, such as radar reflectivity factor and vertical liquid water content, has a more significant impact on lightning occurrence. Therefore, the weight of radar echo intensity data is adjusted to ≥0.6 to highlight its role in fusion. After the first lightning strike, the existing discharge activity provides a direct reference for subsequent lightning development. The predictive value of lightning location data (such as lightning frequency, intensity, and type) is significantly improved. Therefore, its weight is automatically increased to 0.8, dominating the composition of fusion features.

[0060] When performing feature fusion based on adjusted weights, features from each data source participate in the calculation according to their corresponding weights. For example, in the initial stage, features such as the rate of change of electric field intensity and polarity change of the atmospheric electric field account for more than 70% of the fusion, while radar and lightning data features account for less than 30%; in the mature stage, radar echo intensity-related features account for more than 60%; after the first lightning strike, the features from lightning location data account for 80%, ensuring that the fused features can accurately match the lightning activity patterns of the current stage.

[0061] By identifying the periodic nodes of lightning occurrence, dynamically adjusting the fusion weights of atmospheric electric field data, radar echo intensity data, and lightning location data according to stages, and performing feature fusion, the fused features are made to conform to the core laws of lightning activity at each stage, thereby improving the accuracy of the features.

[0062] Furthermore, step A300 in the method provided in this application embodiment includes: A310: The design includes a network architecture with multiple spatiotemporal convolutional layers, pooling layers, and fully connected layers. The spatiotemporal convolutional layers use three-dimensional convolution operations, with the convolutional kernels sliding in the three dimensions of time, height, and width to extract local spatiotemporal features. The pooling layers are used to reduce the dimensionality of the feature maps.

[0063] A320: The connection layer maps the extracted spatiotemporal features to the predicted output of lightning occurrence probability and motion trajectory.

[0064] A330: Embed physical constraint mechanisms in the network, including the physical laws of lightning propagation and atmospheric dynamics.

[0065] A340: Using a converged training model, a real-time generated 3D lightning risk heat map is used as input to perform lightning occurrence probability analysis and prediction, and output the probability of lightning occurrence within a future time range. The probability value is between 0 and 1. The higher the probability value, the greater the possibility of lightning occurrence.

[0066] A350: Outputs the trajectory of the lightning cloud, including its speed and direction.

[0067] Specifically, a targeted network architecture is designed first, which includes multiple spatiotemporal convolutional layers, pooling layers, and fully connected layers. The spatiotemporal convolutional layers employ three-dimensional convolution operations, with the convolutional kernels sliding across three dimensions: time, height, and width. For example, a 3×3×3 convolutional kernel can simultaneously cover grid data from three time nodes, three height levels, and three horizontal width units, thereby extracting local spatiotemporal features from the three-dimensional lightning risk heatmap, such as the trend of risk value changes and spatial distribution correlation at different times and heights.

[0068] The pooling layer follows the spatiotemporal convolutional layer. It reduces the dimensionality of the feature map by downsampling (such as max pooling or average pooling), for example, compressing a 64×64×10 feature map to 32×32×5. This reduces computational cost while preserving key features, thus improving model efficiency. The fully connected layer then maps the high-dimensional spatiotemporal features processed by convolution and pooling to specific prediction outputs, establishing a nonlinear relationship between features and the probability of lightning occurrence and the trajectory of the lightning strike.

[0069] Next, physical constraint mechanisms are embedded in the network to make the prediction results conform to actual meteorological laws, including the physical laws of lightning propagation, such as lightning usually developing along the direction of the maximum electric field intensity gradient, and atmospheric dynamic relationships, such as the influence of wind speed and air pressure on cloud movement. By adding constraint terms to the loss function, it is ensured that the model output will not produce results that violate these laws, such as avoiding the prediction of lightning clouds moving at high speed against the wind.

[0070] After the network design and training are completed and convergence is achieved, the real-time generated 3D lightning risk heat map is input into the model. The model calculates the probability of lightning occurrence for each grid in the future time range by analyzing the risk value and its spatiotemporal changes in each grid in the heat map. The probability value is between 0 and 1. For example, if a grid in a certain area outputs 0.85, it means that there is a high probability of lightning occurring in that area in the future. At the same time, the model outputs the trajectory parameters of the lightning cloud, including the movement speed and direction, such as 5 m / s and 30 degrees east of northeast.

[0071] By designing a network architecture that includes spatiotemporal convolutional layers, pooling layers, and fully connected layers with 3D convolutional operations, and embedding a constraint mechanism that incorporates the physical laws of lightning propagation and the relationship with atmospheric dynamics, the model is trained to converge and analyze the 3D lightning risk heat map. The output is the probability of lightning occurrence between 0 and 1 and the motion trajectory including velocity and direction, which improves the consistency between lightning prediction and actual activity and increases the accuracy of prediction.

[0072] Furthermore, step A400 in the method provided in this application embodiment includes: A410: Based on the predicted probability of lightning occurrence and the predicted trajectory output by the spatiotemporal convolutional neural network, calculate the prediction confidence. The confidence is measured by the standard deviation of the model's output probability.

[0073] A420: Set a confidence threshold to divide the prediction results into three levels: high confidence, medium confidence, and low confidence, and classify the warning level according to the probability of lightning occurrence and the confidence level.

[0074] A430: Divides the target area into different warning level zones, and the boundary of each zone is dynamically adjusted according to the movement trajectory of the lightning cloud.

[0075] In one embodiment, firstly, the prediction confidence is calculated based on the lightning occurrence probability and trajectory output by the spatiotemporal convolutional neural network. The confidence is measured by the standard deviation of the model's output probability. For example, if the model predicts the lightning occurrence probability of a certain area for the next hour multiple times with results of 0.82, 0.83, and 0.81 respectively, the smaller standard deviation indicates stable prediction and high confidence; if the prediction results fluctuate between 0.5 and 0.9, the larger standard deviation indicates lower confidence.

[0076] After calculating the confidence level, confidence thresholds are set to classify levels: a confidence level above 0.9 is considered high confidence, between 0.7 and 0.9 is medium confidence, and below 0.7 is low confidence. Simultaneously, warning levels are classified based on the probability of lightning occurrence: Level 1 (high risk) areas require a lightning occurrence probability greater than 0.8 and a confidence level greater than 0.9; Level 2 (medium risk) areas require a lightning occurrence probability between 0.5 and 0.8 and a confidence level between 0.7 and 0.9; and Level 3 (low risk) areas require a lightning occurrence probability less than 0.5 and a confidence level less than 0.7.

[0077] Subsequently, the target area was divided into different warning level zones according to the aforementioned grading standards. Since the movement trajectory (including speed and direction) of lightning clouds continuously changes, the boundaries of each zone need to be dynamically adjusted accordingly. For example, if a lightning cloud is detected moving southeast at a speed of 10 meters per second, the boundary of the high-risk zone will expand southeastward accordingly, and the ranges of the medium- and low-risk zones will also be adjusted synchronously to ensure that the warning area is consistent with the actual lightning development trend.

[0078] By calculating the prediction confidence level based on the standard deviation of the model output probability, setting thresholds to classify high, medium, and low confidence levels, and combining the probability of lightning occurrence to classify first-level, second-level, and third-level warning areas, and dynamically adjusting the area boundaries according to the movement trajectory of lightning clouds, the accuracy and timeliness of the warning classification and the targeted nature of lightning disaster prevention and control have been improved.

[0079] In summary, the method for improving the accuracy of lightning warnings provided in this application has the following technical effects: This application acquires radar data, atmospheric electric field data, and lightning electromagnetic data by connecting to a three-dimensional monitoring network. These data are then fused in real time to generate a three-dimensional lightning risk heat map. A spatiotemporal convolutional neural network with embedded physical constraints is used to predict the probability of lightning occurrence and its trajectory within a future timeframe. The prediction confidence level is calculated, and early warning levels are assigned. Real-time lightning location data is combined to update the geofence of the warning area, thereby improving the accuracy of lightning early warnings and making the warning results more precise and reliable. This achieves the technical goal of improving the accuracy of lightning occurrence probability and trajectory prediction, and fulfilling the requirements for precise early warning and effective prevention and control of lightning disasters.

[0080] Example 2, as Figure 2 As shown, based on the same inventive concept as in Embodiment 1 above, this application provides a system for improving the accuracy of lightning warnings, the system comprising: Multi-source monitoring data acquisition module 1 is used to connect to a three-dimensional monitoring network and acquire multi-source monitoring data, including radar data, atmospheric electric field data, and lightning electromagnetic data.

[0081] The three-dimensional lightning risk heat map acquisition module 2 is used to fuse the multi-source monitoring data in real time to generate a three-dimensional lightning risk heat map.

[0082] The lightning occurrence probability prediction module 3 is used to predict the probability of lightning occurrence and its trajectory in the future time range based on the three-dimensional lightning risk heat map by employing a spatiotemporal convolutional neural network with embedded physical constraint mechanism.

[0083] The warning area update module 4 is used to classify the probability of lightning occurrence and the movement trajectory according to the prediction confidence level, and update the geofence of the warning area based on real-time lightning location data.

[0084] Furthermore, the multi-source monitoring data acquisition module 1 is used to perform the following steps: The three-dimensional monitoring network includes: a millimeter-wave phased array radar array deployed in the target area, a distributed atmospheric electric field sensor network, and a wide-area lightning location device.

[0085] Furthermore, the three-dimensional lightning risk heat map acquisition module 2 is used to perform the following steps: Based on the radar data, radar reflectivity factor, vertical liquid water content, and velocity divergence are obtained; electric field intensity change rate and electric field polarity change are extracted from atmospheric electric field data; lightning frequency, lightning intensity, and lightning type are extracted from lightning electromagnetic data; different data sources are aligned and fused according to temporal and spatial relationships to obtain fusion features; lightning risk prediction is performed using fusion features to obtain lightning risk values ​​for each temporal and spatial relationship, and a three-dimensional lightning risk heat map is generated.

[0086] Furthermore, the three-dimensional lightning risk heat map acquisition module 2 is used to perform the following steps: Multi-source monitoring data are aligned according to the collection timestamp; all monitoring data are projected into a unified three-dimensional spatial grid, where radar data is filled into the grid using interpolation, and atmospheric electric field data and lightning electromagnetic data are directly mapped to the grid based on location information; a lightning risk assessment model is established, and the fused feature data is input into the model to calculate the lightning risk value for each grid point; the three-dimensional lightning risk heat map is generated according to the lightning risk value for each grid point.

[0087] Furthermore, the three-dimensional lightning risk heat map acquisition module 2 is used to perform the following steps: Based on the lightning occurrence cycle, the fusion weights of each data source are dynamically adjusted, and feature fusion is performed based on the fusion weights. Specifically, when the lightning is in the initial development stage, the weight of atmospheric electric field data is ≥0.7; when the lightning is in the mature stage, the weight of radar echo intensity data is ≥0.6; and when the lightning is the first lightning strike, the weight of lightning location data is automatically increased to 0.8.

[0088] Furthermore, the lightning occurrence probability prediction module 3 is used to perform the following steps: The design incorporates a network architecture with multiple spatiotemporal convolutional layers, pooling layers, and fully connected layers. The spatiotemporal convolutional layers utilize 3D convolutional operations, with the convolutional kernels sliding along the time, height, and width dimensions to extract local spatiotemporal features. Pooling layers reduce the dimensionality of the feature maps. Connecting layers map the extracted spatiotemporal features to predicted outputs of lightning occurrence probability and trajectory. Physical constraints are embedded in the network, including the physical laws of lightning propagation and atmospheric dynamics. Using a converged training model, a real-time generated 3D lightning risk heatmap is used as input to perform lightning occurrence probability analysis and prediction, outputting the probability of lightning occurrence within a future time range. The probability value ranges from 0 to 1, with higher probability values ​​indicating a greater likelihood of lightning occurrence. The design also outputs the trajectory of the lightning cloud, including its speed and direction.

[0089] Furthermore, the early warning area update module 4 is used to perform the following steps: Based on the predicted probability and trajectory of lightning output from the spatiotemporal convolutional neural network, the prediction confidence is calculated, and the confidence is measured by the standard deviation of the model's output probability. A confidence threshold is set, and the prediction results are divided into three levels: high confidence, medium confidence, and low confidence. Warning levels are also classified according to the probability and confidence of lightning occurrence. The target area is divided into different warning level areas, and the boundary of each area is dynamically adjusted according to the trajectory of the lightning cloud.

[0090] In embodiment three, this application also provides a computer-readable storage medium that can be used to store software programs, computer-executable programs, and modules, such as the program instructions / modules corresponding to a method for improving the accuracy of lightning warnings in this application embodiment, thereby realizing the aforementioned method for improving the accuracy of lightning warnings.

[0091] It should be understood that the embodiments disclosed in this application and the above description enable those skilled in the art to implement this application. However, this application is not limited to the embodiments mentioned above. It should be understood that those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this invention, and should all be included within the protection scope of this application.

Claims

1. A method for improving the accuracy of lightning warnings, characterized in that, include: Connect to a three-dimensional monitoring network to acquire multi-source monitoring data, including radar data, atmospheric electric field data, and lightning electromagnetic data; The multi-source monitoring data are fused in real time to generate a three-dimensional lightning risk heat map; A spatiotemporal convolutional neural network with embedded physical constraints is used to predict the probability of lightning occurrence and its trajectory within a future time range based on the three-dimensional lightning risk heat map. Based on the prediction confidence level, the probability of lightning occurrence and the trajectory of lightning are classified into early warning levels, and the geofence of the early warning area is updated based on real-time lightning location data.

2. The method for improving the accuracy of lightning warnings according to claim 1, characterized in that, The three-dimensional monitoring network includes: a millimeter-wave phased array radar array deployed in the target area, a distributed atmospheric electric field sensor network, and a wide-area lightning location device.

3. The method for improving the accuracy of lightning warnings according to claim 1, characterized in that, The multi-source monitoring data is fused in real time to generate a three-dimensional lightning risk heat map, including: Based on the radar data, obtain the radar reflectivity factor, vertical liquid water content, and velocity divergence. Extract the rate of change of electric field intensity and the change of electric field polarity from atmospheric electric field data; Extract lightning frequency, lightning intensity, and lightning type from lightning electromagnetic data; Different data sources are aligned and fused according to their temporal and spatial relationships to obtain fused features; Lightning risk prediction is performed using fusion features to obtain lightning risk values ​​for various temporal and spatial relationships, and a three-dimensional lightning risk heat map is generated.

4. The method for improving the accuracy of lightning warning according to claim 3, characterized in that, Lightning risk prediction is performed using fusion features to obtain lightning risk values ​​for various temporal and spatial relationships, generating a three-dimensional lightning risk heatmap, including: The multi-source monitoring data are aligned according to the collection timestamp; All monitoring data are projected onto a unified three-dimensional spatial grid. Radar data is filled into the grid using interpolation methods, while atmospheric electric field data and lightning electromagnetic data are directly mapped to the grid based on location information. Establish a lightning risk assessment model, input the fused feature data into the model, and calculate the lightning risk value for each grid point; The three-dimensional lightning risk heat map is generated based on the lightning risk value of each grid point.

5. The method for improving the accuracy of lightning warning according to claim 3, characterized in that, Obtain fusion features, including: Based on the lightning occurrence cycle, the fusion weights of each data source are dynamically adjusted, and feature fusion is performed based on the fusion weights. Specifically, when the lightning is in the initial development stage, the weight of atmospheric electric field data is ≥0.7; when the lightning is in the mature stage, the weight of radar echo intensity data is ≥0.6; and when the lightning is the first lightning strike, the weight of lightning location data is automatically increased to 0.

8.

6. The method for improving the accuracy of lightning warning according to claim 1, characterized in that, A spatiotemporal convolutional neural network with embedded physical constraints is used to predict the probability of lightning occurrence and its trajectory within a future time range based on the aforementioned three-dimensional lightning risk heat map, including: The design incorporates a network architecture with multiple spatiotemporal convolutional layers, pooling layers, and fully connected layers. The spatiotemporal convolutional layers use three-dimensional convolution operations, with the convolutional kernels sliding along the time, height, and width dimensions to extract local spatiotemporal features. The pooling layers are used to reduce the dimensionality of the feature maps. The fully connected layer maps the extracted spatiotemporal features to the predicted output of lightning occurrence probability and motion trajectory; Embed physical constraint mechanisms in the network, including the physical laws of lightning propagation and atmospheric dynamics; Using a converged training model, a real-time generated 3D lightning risk heat map is used as input to perform lightning occurrence probability analysis and prediction, and output the probability of lightning occurrence in the future time range. The probability value is between 0 and 1. The higher the probability value, the greater the possibility of lightning occurrence. It also outputs the trajectory of the lightning cloud, including its speed and direction.

7. The method for improving the accuracy of lightning warning according to claim 6, characterized in that, Based on the prediction confidence level, the probability of lightning occurrence and its trajectory are classified into early warning levels, including: Based on the predicted probability of lightning occurrence and the predicted trajectory output by the spatiotemporal convolutional neural network, the prediction confidence is calculated. The confidence is measured by the standard deviation of the model's output probability. Set a confidence threshold and divide the prediction results into three levels: high confidence, medium confidence, and low confidence. Then, classify the warning level according to the probability of lightning occurrence and the confidence level. The target area is divided into different warning level zones, and the boundary of each zone is dynamically adjusted according to the movement trajectory of the lightning cloud.

8. A system for improving the accuracy of lightning warnings, characterized in that, A system for implementing a method for improving the accuracy of lightning warnings according to any one of claims 1-7, the system comprising: The multi-source monitoring data acquisition module is used to connect to the three-dimensional monitoring network and acquire multi-source monitoring data, including radar data, atmospheric electric field data, and lightning electromagnetic data. A three-dimensional lightning risk heat map acquisition module is used to fuse the multi-source monitoring data in real time to generate a three-dimensional lightning risk heat map; The lightning occurrence probability prediction module is used to predict the probability of lightning occurrence and its trajectory in the future time range based on the three-dimensional lightning risk heat map using a spatiotemporal convolutional neural network with embedded physical constraint mechanism. The warning area update module is used to classify the probability of lightning occurrence and the movement trajectory according to the prediction confidence level, and update the geofence of the warning area based on real-time lightning location data.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements a method for improving the accuracy of lightning warnings as described in any one of claims 1 to 7.

Images