A lightning intelligent targeting early warning method and system for local key targets

Abstract

The application relates to the technical field of lightning mode recognition, and discloses a lightning intelligent target early warning method and system for local key targets, which comprises the following steps: collecting target feature description, estimating storm movement information based on lightning positioning data and radar echo images at continuous time points, generating a risk corridor in the upstream direction of each target area, and forming a target area around the target area in combination with the risk corridor. Unified space-time data is obtained, space-time feature extraction is carried out to obtain storm features, cross-modal correlation selection and weighting are carried out on the storm features driven by the target feature description, and target features are constructed. An early warning result is generated according to the target features within a predetermined time window. When the lightning risk probability reaches the early warning threshold and the spatial risk distribution intersects with the buffer range of the target area, target early warning information is generated. The application realizes lightning target early warning, improves the hit rate and reduces false alarms.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of lightning pattern recognition, in particular to a lightning intelligent targeting early warning method and system for local key targets. BACKGROUND

[0002] Lightning has characteristics such as high voltage, large current and strong electromagnetic radiation, which poses a threat to power communication, major facilities and personnel safety. There is a clear engineering need to improve the accuracy and timeliness of lightning early warning. Existing research has developed from single threshold method to feature mining and identification modeling based on observed signals, such as decomposition, feature extraction and model training based on atmospheric electric field signals to reduce false alarms and improve hit rate.

[0003] The disclosed scheme is mostly from the perspective of global or single source for early warning, lacking target domain recognition and upstream corridor modeling for specific receptors: for example, CN112818912A-lightning early warning method based on integrated empirical mode decomposition and extreme gradient boosting, only based on atmospheric electric field signal decomposition, fusion and classification to output early warning signal, without determining and distinguishing the risk corridor of a single target area and the target area. For example, CN113378915A-determination method of lightning aggregation center, the lightning aggregation center is determined by rasterizing and clustering the lightning location data, which focuses on the overall distribution and aggregation area positioning, and does not construct a dynamic corridor and target evaluation of arrival time in the upstream direction for a target area. CN112396116A-a lightning detection method, device, computer equipment and readable medium, focusing on image detection or electromagnetic waveform recognition detection / prediction method, respectively focusing on the probability output of lightning target detection in image frame or low-frequency electromagnetic wave recognition, without target feature description driven cross-modal correlation selection and weighting, and spatial risk superposition and trigger logic in target buffer range. Therefore, although the existing technology contains various signal processing and identification ideas, the spatiotemporal recognition link for local key targets in pattern recognition is still insufficient, lacking technical means and trigger mechanism for upstream risk corridor, target area, target-driven correlation selection, arrival time and spatial risk distribution.

[0004] Therefore, it is necessary to design a lightning intelligent targeting early warning method and system for local key targets to solve the problems existing in the prior art. SUMMARY

[0005] In view of this, the present application provides a lightning intelligent targeting early warning method and system for local key targets, aiming to solve the problem that the prior art is mostly at the regional or single source level, and it is difficult to align multi-source spatiotemporal data with storm movement at the scale of local key targets and output accurate early warning of target lightning risk, arrival time and spatial impact range.

[0006] In one aspect, this invention proposes a smart targeted early warning method for lightning strikes targeting key local areas, comprising:

[0007] The geographical location, structural parameters, grounding parameters and surrounding terrain parameters of each target area are collected to obtain a target feature description, which is used to characterize lightning vulnerability.

[0008] Storm motion information is estimated based on lightning location data and radar echo images at continuous time intervals, and a risk corridor is generated in the upstream direction of each target area according to the storm motion information. The risk corridor is combined to form a target area around the target area, and the shape and scale of the target area are determined according to the storm motion information.

[0009] Multi-source observation data within the target area are preprocessed to obtain unified spatiotemporal data; spatiotemporal features are extracted from the unified spatiotemporal data to obtain storm features; the target feature description is used as a driver to perform cross-modal correlation selection and weighting on the storm features to construct targeted features for the target area.

[0010] Based on the target features, an early warning result is generated within a predetermined time window. The early warning result includes the probability of lightning strike risk, the arrival time and spatial risk distribution.

[0011] When the probability of lightning strike reaches the warning threshold and the spatial risk distribution intersects with the buffer zone of the target area, targeted warning information is generated. The targeted warning information includes the identifier of the target area, the risk level, and the arrival time. The buffer zone is a preset distance band outside the target area.

[0012] Furthermore, when obtaining the target feature description, it includes:

[0013] Determine the boundaries, geographical location and altitude of each target area, collect structural parameters, grounding parameters and surrounding terrain parameters, perform time consistency verification, clean up duplicate records and remove outliers, and mark missing data for data that cannot be verified.

[0014] Based on digital elevation data and the height of nearby facilities, the 360° range is divided into at least four equal-angle sectors. The relative rise and shading angle of each sector are calculated to form a circumferential exposure distribution.

[0015] Based on the multi-year wind direction statistics and historical lightning records of the target area, the dominant sector is identified and marked as a sensitive sector to form a windward exposure description.

[0016] The grounding parameters are corrected according to the season and soil moisture; the structural parameters are classified according to height, degree of conductive exposure and down conductor type to obtain structural classification; the grounding parameters are classified according to preset intervals to obtain grounding classification; the historical lightning records are counted in layers according to the number of occurrences and distance to obtain historical record count; and the boundary, geographical location and altitude, structural classification, grounding classification, windward exposure description and historical record count of the target area are combined in sequence to form the target feature description.

[0017] Furthermore, when estimating storm motion information based on lightning location data and radar echo images at continuous time intervals, the following steps are included:

[0018] Align lightning location data at continuous intervals with radar echo images, the alignment including uniform grid and uniform time step;

[0019] The radar observation results are determined based on the correlation matching of radar echoes at adjacent time points. The radar observation results include the radar observation direction of motion and the radar observation speed.

[0020] The displacement direction and displacement velocity of the centroid of lightning activity density at adjacent moments are calculated to obtain the positioning observation results, which include the positioning observation motion direction and positioning observation motion velocity.

[0021] The radar observation results are compared with the location observation results. When the radar observation results and the location observation results are consistent, the storm movement direction and storm movement speed are obtained by merging them; otherwise, the radar observation results are corrected according to the location observation results, and abnormal segments are marked.

[0022] The spatial expansion trend is determined by the changes in radar echo profile area and lightning activity range, and the abnormal segments are removed to obtain the storm movement information.

[0023] Furthermore, based on the storm movement information, a risk corridor is generated upstream of each target area, and when the risk corridor forms a target area around the target area, it includes:

[0024] A risk corridor is established upstream of each target area along the storm's movement direction. Its length is determined by a predetermined time window and the storm's movement speed, while its width is determined by the spatial expansion trend and historical prediction errors. Uncertainty buffer zones are set on both sides. When the risk corridor intersects with the buffer zone of the target area, a target area is generated around the target area. The target area is an elliptical region with its major axis along the storm's movement direction, where the major axis increases with the storm's movement speed, and the minor axis increases with the spatial expansion trend and the increase of the uncertainty buffer zone. When the corridor does not intersect, the monitoring status is recorded, and the target area is not generated.

[0025] Furthermore, when preprocessing the multi-source observation data within the target area to obtain unified spatiotemporal data, the process includes:

[0026] The multi-source observation data includes radar echo images, lightning location data, camera images, ground electric field observation data, and low-frequency electromagnetic observation data.

[0027] The radar echo images are subjected to ground object occlusion correction and ground clutter suppression; the lightning location data are subjected to event deduplication and time window aggregation; the camera images are subjected to geometric correction and brightness normalization; and the ground electric field observation data and low-frequency electromagnetic observation data are subjected to abnormal pulse removal and baseline correction. A quality score and missing measurement marker are generated for each time step, and the unified spatiotemporal data are constructed according to a fixed channel order and a fixed time order. When any data source is missing, the missing measurement marker is used instead of interpolation.

[0028] Furthermore, when extracting storm features from the unified spatiotemporal data, the process includes:

[0029] Based on the unified spatiotemporal data, rapid change features are extracted within a first time window, and continuous change features are extracted within a second time window, where the first time window is shorter than the second time window. Energy distribution within the target area and the risk corridor is converged at a first spatial scale and a second spatial scale, respectively, where the first spatial scale is shorter than the second spatial scale. The propagation coherence and boundary steepness along the risk corridor direction are enhanced. Furthermore, the influence of low-quality and missing information is suppressed based on the quality score and missing data markers to obtain the storm features, which include local intensity, boundary steepness, propagation coherence, and upstream persistence.

[0030] Furthermore, when constructing targeted features for the target region by driving cross-modal correlation selection and weighting on the storm features using the target feature description, the process includes:

[0031] Sensitive directions and sectors are determined based on the structural classification, grounding classification, and windward exposure description in the target feature description. Storm features located in sensitive sectors and varying along the risk corridor are given increased weight, while storm feature data corresponding to locations and time periods marked in historical false alarm and missed alarm records are given decreased weight. Low-quality and missing information is suppressed based on the quality score and missing data markers. Correlation scores between each location and time period are calculated, correlation thresholds are set and ranked, and locations and time periods corresponding to storm feature data exceeding the correlation threshold are retained. The storm feature data for these locations and time periods are then aggregated to form targeted features for the target area.

[0032] Furthermore, when generating an early warning result within a predetermined time window based on the target features, the process includes:

[0033] The target features are statistically analyzed and sorted. A preliminary assessment of the lightning risk probability is obtained based on local intensity, boundary steepness, propagation coherence, and upstream persistence. A preliminary assessment of the arrival time is determined along the risk corridor direction. A preliminary assessment of the spatial risk distribution is generated within the target area. The fusion weight is determined based on the confidence and quality scores of each data point in the multi-source observation data. Evidence fusion is performed according to the consistency correction rule to obtain the fused lightning risk probability, fused arrival time, and fused spatial risk distribution. When determining the warning threshold, the warning threshold is corrected based on the historical false alarm records and historical missed alarm records of the target area.

[0034] Furthermore, when generating targeted early warning information, the following are included:

[0035] The risk level is positively correlated with the lightning risk probability, arrival time, and spatial risk distribution.

[0036] Lightning location data is acquired within a predetermined time window corresponding to the arrival time, and the correlation between the actual lightning location hit result and the target area is determined based on a preset distance range and a preset time deviation. When there is a correlated actual lightning location hit result, the warning is marked as a hit and is not counted as a false alarm in the historical false alarm record. When there is no correlated actual lightning location hit result, the warning is counted in the historical false alarm record and the current quality score and data missing mark are retained. When an actual lightning location hit result correlated with the target area occurs in the same time window without meeting the triggering condition, it is counted in the historical missed alarm record. When multiple targeted warnings cover the same target area and the time windows overlap, the earliest triggered targeted warning information is correlated with the actual lightning location hit result once, and subsequent overlapping warnings are not correlated repeatedly.

[0037] Compared with existing technologies, the beneficial effects of this invention are as follows: It compresses wide-area thunderstorm activity into the spatiotemporal window most relevant to the target by using upstream risk corridors and target areas, weakening background interference unrelated to the target and reducing false alarms; within the target area, it unifies radar, lightning location, and other accessible observations into unified spatiotemporal data, maintaining data accuracy even under conditions of sensor gaps, noise, and scene switching; it synchronously characterizes sudden enhancement and persistent organizational structure through multi-time window and multi-scale spatiotemporal feature extraction, providing more accurate priors for arrival time and impact range; it uses target feature descriptions as queries to perform cross-modal correlation selection and weighting of storm features, aligning storm evidence with the target's structure, grounding, and exposure attributes one by one, outputting the probability of lightning exposure and spatial risk distribution, achieving targeted discrimination of the same storm, different targets, and different risks; and it provides manageable early warning, balancing accuracy and interpretability, improving hit rate and reducing false alarm rate.

[0038] On the other hand, this application also provides a lightning intelligent targeted early warning system for key local targets, used to apply the above-mentioned lightning intelligent targeted early warning method for key local targets, including:

[0039] The acquisition unit is configured to acquire the geographical location, structural parameters, grounding parameters and surrounding terrain parameters of each target area to obtain a target feature description, which is used to characterize lightning vulnerability.

[0040] The first processing unit is configured to estimate storm motion information based on lightning location data and radar echo images at continuous time intervals, and generate a risk corridor in the upstream direction of each target area according to the storm motion information, and form a target area around the target area by combining the risk corridor, wherein the shape and scale of the target area are determined according to the storm motion information.

[0041] The second processing unit is configured to preprocess the multi-source observation data within the target area to obtain unified spatiotemporal data; extract spatiotemporal features from the unified spatiotemporal data to obtain storm features; and use the target feature description to drive cross-modal correlation selection and weighting on the storm features to construct targeted features for the target area.

[0042] The early warning unit is configured to generate an early warning result within a predetermined time window based on the target characteristics. The early warning result includes the probability of lightning strike risk, the time and spatial risk distribution of arrival.

[0043] The early warning unit is further configured to generate targeted early warning information when the probability of lightning strike reaches an early warning threshold and the spatial risk distribution intersects with the buffer range of the target area. The targeted early warning information includes the identifier of the target area, the risk level, and the arrival time.

[0044] It is understandable that the aforementioned intelligent targeted early warning methods and systems for lightning targeting key local areas have the same beneficial effects, and will not be elaborated upon here. Attached Figure Description

[0045] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0046] Figure 1 A flowchart of a lightning intelligent targeted early warning method for key local targets provided in an embodiment of the present invention;

[0047] Figure 2 This is a functional block diagram of a lightning intelligent targeted early warning system for key local targets provided in an embodiment of the present invention. Detailed Implementation

[0048] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0049] Traditional lightning warning methods, based on global monitoring data or single signal sources, struggle to achieve precise protection for key local targets. Existing technologies rely on static thresholds or single-modal feature analysis, failing to dynamically identify upstream risk corridors in the target area, resulting in insufficient spatial resolution for warnings. The heterogeneity of multi-source observation data limits spatiotemporal alignment accuracy, and the lack of consideration for the dynamic correlation between target features and storm propagation paths during cross-modal data fusion leads to a simultaneous increase in false alarm and missed alarm rates. The absence of a spatial coupling mechanism between the target area's buffer zone and the risk corridor results in spatiotemporal discrepancies between the warning triggering logic and the actual lightning trajectory.

[0050] For example, in power facility protection scenarios, traditional methods use fixed-radius warning area delineation without dynamically adjusting the target area based on storm movement speed and spatial expansion trends. Differences in time steps between radar echo images and lightning location data lead to motion vector estimation errors exceeding three meters per second. The lack of an uncertainty buffer in the risk corridor generation process causes accumulated prediction bias in the upstream direction. The absence of a quality scoring mechanism in the multi-source data preprocessing stage allows pulse interference in ground electric field observation data to directly affect the accuracy of spatiotemporal feature extraction, resulting in misjudgments of storm boundary steepness. Furthermore, the circumferential exposure distribution of the target area is not weighted and matched with the risk corridor direction; when the sensitive sector identification deviation exceeds fifteen degrees, the warning information misaligns with the actual lightning vulnerability of the facilities.

[0051] If the above problems are not addressed, the early warning system will be unable to accurately capture the consistent characteristics of lightning propagation upstream of key local targets, leading to delayed or false triggering of protective devices. Static early warning area models struggle to adapt to the nonlinear expansion during storm movement, resulting in excessive deployment of protective resources in non-critical areas. Defects in the quality of multi-source data directly impact the reliability of feature fusion; the continuous accumulation of historical false alarms and missed alarms will reduce system credibility. The lack of a spatial coupling mechanism between the target area buffer zone and the risk corridor may create protection blind spots, increasing the risk of structural damage and equipment failure caused by lightning strikes.

[0052] For this, please refer to Figure 1 As shown, this application proposes a smart targeted lightning early warning method for key local targets, including:

[0053] S100: Collect the geographical location, structural parameters, grounding parameters and surrounding terrain parameters of each target area to obtain a target feature description, which is used to characterize lightning vulnerability.

[0054] S200: Based on lightning location data and radar echo images from continuous time intervals, storm motion information is estimated, and a risk corridor is generated upstream of each target area according to the storm motion information. The risk corridor is combined to form a target area around the target area, and the shape and scale of the target area are determined according to the storm motion information.

[0055] S300: Preprocess multi-source observation data within the target area to obtain unified spatiotemporal data. Extract storm features from the unified spatiotemporal data. Using the target feature description as a driving force, perform cross-modal correlation selection and weighting on the storm features to construct targeted features for the target area.

[0056] S400: Generates early warning results within a predetermined time window based on target characteristics. The early warning results include the probability of lightning strike risk, arrival time, and spatial risk distribution.

[0057] S500: When the probability of lightning strike reaches the warning threshold and the spatial risk distribution intersects with the buffer zone of the target area, a targeted warning message is generated. The targeted warning message includes the identifier of the target area, the risk level and the arrival time. The buffer zone is a preset distance band outside the target area.

[0058] Specifically, target feature description refers to a dataset that comprehensively characterizes the lightning vulnerability of a target area by collecting its geographical location, structural parameters, grounding parameters, and surrounding terrain parameters. This can be achieved through methods such as using a geographic information system to collect coordinates and altitude, sensors to obtain grounding resistance values, 3D modeling to extract structural parameters, and topographic map analysis to determine surrounding obstruction or uplift angles. This description quantifies the target area's exposure and vulnerability to lightning. The risk corridor refers to a dynamic path region generated upstream of the target area based on storm movement information. This can be achieved by using radar echo image matching and lightning location data displacement analysis to determine the storm's movement direction and speed, and combining historical prediction errors to set the length and width range. The risk corridor is used to dynamically track the potential threat path of thunderstorms to the target area. The target area refers to the risk coverage area formed around the target area. This can be achieved using an elliptical model combined with adjusting the major axis according to the risk corridor direction, adjusting the minor axis according to the spatial expansion trend, and setting buffer zone expansion boundaries. This area is used to focus on the direct impact range of thunderstorms on the target area. Multi-source observation data preprocessing refers to the standardization of radar, lightning location, camera, electric field, and electromagnetic data. This can be achieved using ground feature correction algorithms, event deduplication rules, geometric correction models, pulse rejection thresholds, and baseline calibration parameters. Preprocessing is used to eliminate data noise and unify spatiotemporal references. Cross-modal correlation selection and weighting involves filtering and enhancing the parts of storm features associated with target vulnerability based on target characteristics. This can be achieved using sensitive sector direction matching, historical false alarm record filtering, and quality scoring weighting mechanisms. This step increases the weight of relevant signals in the target area's features. Spatial risk distribution and buffer zone intersection determination involves spatially overlaying the predicted risk area with the target's preset distance zone. This can be achieved using geospatial overlay algorithms to calculate the intersection area ratio and set an intersection threshold. This determination is used to confirm whether the thunderstorm threat has entered the target's defense range.

[0059] This application achieves precise early warning of lightning threats to key local targets by dynamically constructing risk corridors and target areas, combined with multi-source data-driven cross-modal feature fusion and spatial overlay triggering mechanisms. This method breaks through the traditional global or single-source early warning model by dynamically linking thunderstorm movement paths, target vulnerability characteristics, and spatial risk distribution, thereby improving the targeting and timeliness of the early warning.

[0060] The working process and principle of this application are as follows: First, the geographical location, structural parameters, grounding parameters, and surrounding terrain parameters of each target area are collected to obtain a target feature description, which is used to characterize lightning vulnerability. Then, based on continuous lightning location data and radar echo images, storm movement information is estimated. According to the storm movement information, a risk corridor is generated in the upstream direction of each target area. The risk corridor is combined to form a target area around the target area. The shape and scale of the target area are determined according to the storm movement information.

[0061] Next, the multi-source observation data within the target area are preprocessed to obtain unified spatiotemporal data. Spatiotemporal features are extracted from the unified spatiotemporal data to obtain storm features. The target feature description is used as the driving force to perform cross-modal correlation selection and weighting on the storm features, thereby constructing targeted features for the target area.

[0062] Based on the target characteristics, early warning results are generated within a predetermined time window, including the probability of lightning strike risk, arrival time, and spatial risk distribution. When the probability of lightning strike risk reaches the early warning threshold and the spatial risk distribution intersects with the buffer zone of the target area, targeted early warning information is generated, including the target area identifier, risk level, and arrival time.

[0063] This scheme achieves precise early warning for key local targets by dynamically generating risk corridors and target areas. Spatiotemporal alignment of multi-source data and cross-modal feature fusion improve the accuracy of early warnings. Target feature-driven relevance selection ensures that early warnings match target vulnerability. A cross-judgment mechanism between spatial risk distribution and target buffer range further enhances the accuracy of early warnings.

[0064] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0065] First, the geographical location, structural parameters, grounding parameters, and surrounding terrain parameters of the target area are collected. For example, for a substation, its latitude and longitude coordinates, altitude, building area, number of main transformers, lightning rod height, grounding grid resistance value, and surrounding building heights are recorded to form a target feature description.

[0066] Then, weather radar and lightning location systems were used to acquire radar echo images and lightning location data at continuous intervals. Image processing algorithms were used to calculate the displacement and deformation of radar echoes at adjacent time points to obtain the storm's direction and velocity. Simultaneously, the movement trajectory of the center of gravity of lightning activity density was analyzed to cross-validate the accuracy of the storm's movement information.

[0067] Based on the storm's movement direction, a fan-shaped risk corridor is generated upstream of the substation. The corridor length is determined by the warning time window and the storm speed, while the width takes into account the storm's expansion trend. This risk corridor, combined with the storm's movement direction, forms an elliptical target area around the substation.

[0068] Spatiotemporal alignment and preprocessing were performed on multi-source data, including radar echoes, lightning location data, and ground electric fields, within the target area. Spatiotemporal features such as echo intensity, lightning frequency, and electric field change rate were extracted to construct a storm feature vector. Based on the substation's structural characteristics and historical lightning strike records, the storm features were correlated and weighted to obtain targeted features for the substation.

[0069] The probability of lightning strike, estimated arrival time, and spatial risk distribution are calculated based on targeted features. When the probability of lightning strike exceeds the warning threshold and the spatial risk distribution intersects with the 500-meter buffer zone around the substation, targeted warning information containing the substation number, risk level, and estimated arrival time is generated.

[0070] Through the above-described scheme, this application achieves precise lightning warnings for key local targets. Dynamically generated risk corridors and target areas improve the spatial resolution of the warnings and reduce false alarms. Spatiotemporal alignment of multi-source data and cross-modal feature fusion enhance the reliability of the warnings. Target feature-driven correlation selection ensures that the warnings match the target's vulnerability, improving the warnings' targeting effectiveness. The cross-judgment mechanism between spatial risk distribution and target buffer range further enhances the spatiotemporal accuracy of the warnings. The targeted warning method can provide more timely and accurate lightning protection guidance for key facilities, reducing the risk of damage caused by lightning strikes.

[0071] In some of the schemes described above in this application, target feature description is used to characterize lightning vulnerability. However, in practical applications, the geographical location, structural parameters, grounding parameters and surrounding terrain parameters of the target area have problems such as data acquisition errors, time inconsistencies and dynamic environmental changes, which lead to insufficient accuracy of feature description and affect the accuracy of subsequent risk corridor and target area generation.

[0072] This application further proposes a process including determining the boundaries, geographical location, and altitude of each target area; collecting structural parameters, grounding parameters, and surrounding terrain parameters; performing time consistency verification, duplicate record cleanup, and outlier removal; and marking unverifiable data as missing data. Based on digital elevation data and the height of nearby facilities, the 360° range is divided into at least four equal-angle sectors. The relative uplift and shading angle of each sector are statistically analyzed to form a circumferential exposure distribution. Based on multi-year wind direction statistics and the distribution of historical lightning strikes in the target area, dominant sectors are identified and marked as sensitive sectors to form a windward exposure description. Grounding parameters are corrected according to season and soil moisture. Structural parameters are classified according to height, degree of conductive exposure, and down conductor type to obtain a structural classification. Grounding parameters are classified according to preset intervals to obtain a grounding classification. Historical lightning records are counted layer by layer according to the number of occurrences and distance to obtain a historical record count. Finally, the boundaries, geographical location and altitude, structural classification, grounding classification, windward exposure description, and historical record count of the target area are combined sequentially to form a target feature description.

[0073] The system includes several key features: Time consistency verification identifies data exceeding a preset fluctuation range by comparing the collected values ​​of the same parameter at different time points; Duplicate record cleanup uses a hash algorithm to detect redundant entries; Outlier removal identifies data deviating from the mean by more than three standard deviations based on a statistical distribution model; Equal-angle sector division uses a polar coordinate system with the center of the target area as the origin, with each sector's angle not exceeding 90°; Relative rise angle is calculated using a digital elevation model to determine the elevation difference between the target area and adjacent facilities; and Obstruction angle is determined by line-of-sight analysis to assess the degree of obstruction of the surrounding terrain to the lightning current path; Windward exposure description uses the joint probability distribution of wind direction data and lightning event direction over ten years to filter sectors with a frequency exceeding a set threshold; Grounding parameter correction uses an empirical formula for soil resistivity variation with the seasons, dynamically adjusting the correction coefficient based on real-time humidity sensor data; Structural grading divides height into five-meter intervals; Conductivity exposure is categorized into three levels based on the proportion of metal surface area; and Downlead type is categorized into surface-mounted, concealed, and mixed installation based on the installation method; and Historical record counting is divided into five-hundred-meter intervals based on the distance between the lightning occurrence location and the target area boundary.

[0074] Specifically, the boundary and geographical location of the target area were acquired through a geographic information system (GIS), and altitude data was obtained through lidar measurements. Structural parameters included building height, roof metal component area, and down conductor installation method; grounding parameters included grounding resistance and soil type. Surrounding terrain parameters were obtained by generating a 3D model using UAV aerial surveying, extracting slope and obstacle height. During time consistency verification, if the same structural parameter fluctuated by more than 20% in three consecutive acquisitions, a manual review process was triggered. Duplicate record cleanup used the MD5 hash algorithm to generate data fingerprints and deleted entries corresponding to duplicate fingerprints. After outlier removal, missing data was marked as missing to prevent erroneous interpolation from affecting subsequent analysis. In the generation of the circumferential exposure distribution, the relative uplift angle of each sector was calculated by comparing the altitude difference between the center point of the target area and the highest point within a 500-meter radius; the shading angle was calculated using a line-of-sight shading model to determine the maximum obstacle elevation angle in each direction. Windward exposure description was achieved by fusing ten years of meteorological data and lightning location data using kernel density estimation to identify dominant sectors with a lightning event probability exceeding 60%. The seasonal correction coefficient for grounding resistance is obtained from tables based on monthly average precipitation and soil moisture content. A frozen soil resistivity correction model is used in winter, while a saturated soil correction model is used in summer. In structural grading, height grading intervals are set as 0-10 meters, 10-15 meters, and 15-20 meters, etc. Conductivity exposure is divided into three levels based on the metal surface area ratio: less than 30%, 30%-60%, and greater than 60%. Historical record counts are divided into levels based on the distance between lightning events and the target area boundary: 0-500 meters, 500-1000 meters, and 1000-2000 meters. The number of events in each level is counted using a sliding time window. The final combined target feature description is arranged in a fixed field order, including boundary coordinates, altitude, structural grading code, grounding grading code, sensitive sector azimuth range, and historical event count matrix, forming a standardized data structure for downstream modules to use.

[0075] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0076] Determine the boundaries, geographical location, and altitude of each target area. Collect structural parameters, grounding parameters, and surrounding terrain parameters. Perform time consistency checks, duplicate record cleanup, and outlier removal. Mark any data that cannot be verified as missing.

[0077] Based on digital elevation data and the height of nearby facilities, the 360° area was divided into eight equal-angle sectors. The relative uplift and shading angles of each sector were calculated to form a circumferential exposure distribution.

[0078] Based on the distribution of wind direction statistics and historical lightning records over many years in the target area, the dominant sector is identified and marked as a sensitive sector to form a windward exposure description.

[0079] Grounding parameters are corrected according to season and soil moisture. Structural parameters are classified according to height, degree of conductor exposure, and down conductor type to obtain structural classification. Grounding parameters are classified according to preset intervals to obtain grounding classification. Historical lightning records are stratified and counted according to the number of occurrences and distance to obtain historical record counts.

[0080] The target area's boundaries, geographical location and altitude, structural classification, grounding classification, windward exposure description, and historical record count are combined in sequence to form a target feature description.

[0081] Through the above technical solution, this application achieves a comprehensive characterization of the lightning vulnerability of the target area. By collecting and processing multi-dimensional parameters, a target feature description is formed, incorporating key information such as geography, structure, grounding, and exposure. This description method considers both the inherent attributes of the target and environmental factors. Therefore, the early warning system can perform personalized risk assessments based on the unique characteristics of each specific target, improving the targeting and accuracy of the early warning.

[0082] In some of the solutions described above in this application, when estimating storm motion information based on lightning location data and radar echo images at continuous time intervals, relying solely on a single data source or failing to align and verify multi-source data may lead to estimation errors in motion direction and velocity, thereby affecting the accuracy of subsequent risk corridor and target area generation.

[0083] This application further proposes aligning lightning location data with radar echo images at consecutive time points, including unifying the grid and time step. Radar observation results are determined based on correlation matching of radar echoes at adjacent time points, including the radar observation direction and velocity. The displacement direction and velocity of the centroid of lightning activity density at adjacent time points are calculated to obtain location observation results, including the location observation direction and velocity. The radar observation results are compared with the location observation results; when they match, they are merged to obtain the storm movement direction and velocity. Otherwise, the radar observation results are corrected based on the location observation results, and anomalous segments are marked. The spatial expansion trend is determined based on changes in the radar echo contour area and the lightning activity range, and anomalous segments are removed to obtain storm movement information.

[0084] The unified grid is achieved by resampling radar echo images and lightning location data at different resolutions to the same spatial grid, and the unified time step achieves data synchronization through time interpolation or time window aggregation. The radar observation motion direction is obtained by calculating the displacement vector of the echo centroid at adjacent time points, and the radar observation motion velocity is determined by dividing the displacement distance by the time interval. The lightning activity density centroid is calculated by weighting the number and location of lightning events per unit area, and the location observation motion direction and velocity are derived from the centroid displacement vector at adjacent time points. Anomaly segment labeling is based on whether the difference in motion parameters between radar and lightning data exceeds a preset threshold; the correction process uses lightning data to weight and adjust the direction and velocity of radar observation results. The spatial expansion trend is determined by comprehensively calculating the radar echo area growth rate and the lightning activity range expansion rate over consecutive time points.

[0085] Specifically, after processing radar echo images and lightning location data using a unified raster and time step, errors caused by differences in spatial resolution and temporal sampling rate are eliminated. By matching the correlation of radar echoes at adjacent time points, the displacement vector of the echo centroid is extracted to obtain radar observation motion parameters. Simultaneously, the displacement vector of the lightning activity density centroid reflects the actual movement trend of the thunderstorm system. When the two observation results are consistent, merging the data enhances the reliability of the motion parameters. When discrepancies exist, lightning location data is prioritized to correct radar observation results, as lightning activity directly reflects the movement characteristics of the thunderstorm core. The labeling and removal of anomalous segments prevents local anomalies caused by radar ground object interference or lightning location errors from affecting the overall trend judgment. The determination of spatial expansion trends, combined with changes in radar echo contours and lightning activity range, can identify the development intensity and spread rate of the thunderstorm system, providing a dynamic adjustment basis for the width of the risk corridor and the scale of the target area. Through multi-source data cross-validation and anomaly removal, the final output storm motion information has higher accuracy and robustness, laying a reliable foundation for the generation of downstream risk corridors and target areas.

[0086] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0087] The lightning location data at consecutive time points were aligned with the radar echo imagery. Alignment involved standardizing the grid and time step. Specifically, the lightning location data and radar echo imagery were resampled to the same spatial resolution grid, such as a 1km × 1km grid. The time step was standardized to a 5-minute interval.

[0088] Radar observation results are determined based on correlation matching of radar echoes from adjacent time steps. These results include the observed direction and velocity of motion. Further, the displacement between radar echo images from two adjacent time steps is calculated using a cross-correlation method, yielding the motion vector field. Cluster analysis is then performed on the motion vector field to obtain the dominant motion direction and velocity.

[0089] The location observation results are obtained by calculating the displacement direction and velocity of the centroid of lightning activity density at adjacent time steps. These results include the direction and velocity of the observed motion. The lightning activity density is calculated using the kernel density estimation method. The velocity vector is obtained by dividing the difference in the centroid positions of adjacent time steps by the time interval.

[0090] The radar observations are compared with the location observations. When the radar and location observations match, the storm's direction and speed are merged. Otherwise, the radar observations are corrected based on the location observations, and anomalous segments are marked. Specifically, a match is considered to be formed when the directional difference is less than 30 degrees and the speed difference is less than 5 m / s. If they do not match, the location observations are used for correction.

[0091] Spatial expansion trends were determined by analyzing changes in radar echo contour area and lightning activity range, and anomalous segments were removed to obtain storm motion information. The contour area was extracted using a 35 dBZ echo intensity threshold. Lightning activity range was characterized by the area of ​​an ellipse with a 95% confidence interval. The rate of change of area over three consecutive time steps was used as an indicator of spatial expansion trend.

[0092] Through the above technical solutions, this application achieves accurate estimation of storm motion information. Therefore, cross-validation based on multi-source data improves the reliability of storm motion information. Furthermore, by introducing spatial expansion trends, the dynamic characteristics of storm development are captured. The identification and removal of anomalous segments enhances the stability and coherence of storm motion information. This provides more reliable input for risk corridor generation and target area formation, thereby improving the accuracy and timeliness of early warnings.

[0093] In some of the solutions described above in this application, when generating risk corridors based on storm motion information, there is a problem that the risk corridors cannot dynamically adapt to changes in storm motion speed and spatial expansion trends. This results in the size and shape of the risk corridors failing to accurately reflect the actual threat range, potentially leading to deviations where the warning area is too small or too large.

[0094] This application further proposes establishing a risk corridor upstream of each target area along the storm's movement direction. The length is determined based on a predetermined time window and the storm's movement speed, while the width is determined based on spatial expansion trends and historical prediction errors. Uncertainty buffer zones are set on both sides. When the risk corridor intersects with the buffer zone of the target area, a target area is generated around the target area. The target area is an elliptical region with its major axis along the storm's movement direction, where the major axis increases with increasing storm movement speed, and the minor axis increases with increasing spatial expansion trends and the expansion of the uncertainty buffer zone. When the corridor does not intersect, the monitoring status is recorded, but no target area is generated.

[0095] The length of the risk corridor is calculated by multiplying a predetermined time window by the storm's movement speed. For example, if the time window is 30 minutes and the speed is 40 km / h, the corridor length is 20 km. The width is calculated by weighting the linear growth coefficient of the spatial expansion trend with the standard deviation of historical prediction errors. For example, if the expansion trend coefficient is 1.2 and the historical error standard deviation is 5 km, the width is set to 6 km. The uncertainty buffer zone is set as a fixed proportion of the width, for example, 20%. The major and minor axes of the elliptical region are dynamically adjusted using linear functions of speed and expansion trend, respectively. For example, major axis = base length + speed increment × coefficient, minor axis = base width + expansion increment × coefficient.

[0096] Specifically, when generating risk corridors, the corridor length is first calculated based on the storm's movement speed and the warning time window to ensure coverage of the storm's potential movement distance within the time window. The expansion rate of the storm's extent is quantified according to its spatial expansion trend, and the corridor width is determined by combining statistical results of historical prediction errors to cover possible lateral shifts. The setting of an uncertainty buffer zone further increases the tolerance range in the width direction. When the risk corridor intersects with the buffer zone of the target area, it indicates that the storm path may pose a threat to the target; in this case, a target area is generated in an elliptical shape. The buffer zone refers to a preset distance band surrounding the outer boundary of the target area, used to determine the intersection relationship between the spatial risk distribution and the target area. The major axis of the ellipse is aligned with the storm's movement direction, and its length increases with speed to reflect the cumulative effect of the movement distance. The minor axis is superimposed on the buffer zone according to the spatial expansion trend to cover lateral expansion and error range. If the risk corridor does not intersect with the buffer zone, only the monitoring status is recorded to avoid invalid warnings. This scheme achieves real-time matching of risk coverage and storm evolution characteristics through dynamic parameter adjustment and elliptical area generation, improving the spatial accuracy of the warning area.

[0097] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0098] A risk corridor is established upstream of each target area along the storm's movement direction. The length of the risk corridor is determined based on a predetermined time window and the storm's movement speed. For example, when the predetermined time window is 30 minutes and the storm's movement speed is 40 km / h, the risk corridor length is set to 20 km. The width of the risk corridor is determined based on the spatial expansion trend and historical prediction errors. Specifically, the spatial expansion trend can be calculated using the rate of change of the storm boundary in continuous radar echo images. Historical prediction errors are obtained statistically based on the deviation between prediction results and actual observation results over a past period. Uncertainty buffer zones are set on both sides of the risk corridor; the width of the buffer zones can be determined based on the standard deviation of historical prediction errors.

[0099] Furthermore, when the risk corridor intersects with the buffer zone of the target area, a target area is generated around the target area. The target area is an elliptical region with its major axis along the direction of storm movement. The major axis increases with the storm's movement speed, while the minor axis increases with the spatial expansion trend and the uncertainty buffer zone. For example, when the storm's movement speed is 60 km / h, the major axis of the ellipse can be set to 15 km. When the spatial expansion trend is 10 km / h and the uncertainty buffer zone width is 2 km, the minor axis of the ellipse can be set to 8 km.

[0100] Therefore, when the risk corridor and the target area buffer zone do not intersect, the monitoring status is recorded but no target area is generated. In this case, storm movement continues to be monitored until the intersection condition is met or the storm dissipates.

[0101] Through the above technical solution, this application realizes the construction of a dynamic risk corridor based on storm motion characteristics and generates targeted early warning areas according to the location of the target area. Compared with early warning areas of fixed shape and size, this method can more accurately reflect the potential impact of storms on specific targets. At the same time, by introducing an uncertainty buffer zone, the risk of missed warnings due to storm motion prediction errors is reduced. In addition, the dynamic adjustment mechanism of the targeted area improves the spatiotemporal accuracy of the early warning and reduces unnecessary warnings, thereby enhancing the reliability and practicality of the early warning.

[0102] In some of the above-mentioned schemes in this application, when preprocessing multi-source observation data in the target area, the different data sources have differences in spatiotemporal resolution, acquisition frequency and data format, which makes spatiotemporal alignment difficult in the subsequent feature extraction process. In addition, low-quality or missing data may introduce noise interference, affecting the reliability and consistency of unified spatiotemporal data.

[0103] This application further proposes preprocessing multi-source observation data within a target area to obtain unified spatiotemporal data. The multi-source observation data includes radar echo images, lightning location data, camera images, ground electric field observation data, and low-frequency electromagnetic observation data. Ground object obstruction correction and ground clutter suppression are performed on the radar echo images; event deduplication and time window aggregation are performed on the lightning location data; geometric correction and brightness normalization are performed on the camera images; and abnormal pulse removal and baseline correction are performed on the ground electric field observation data and low-frequency electromagnetic observation data. A quality score and missing data marker are generated for each time step, and these are arranged in a fixed channel and time order to form unified spatiotemporal data. When any data source is missing, the missing data marker is used instead of interpolation.

[0104] In this process, ground object occlusion correction of radar echo images is achieved by correcting the terrain occlusion effect on the radar beam path using a digital elevation model, while ground clutter suppression employs a polarization filtering algorithm to remove static ground object reflection interference. Event deduplication of lightning location data is performed by merging duplicate records based on temporal and spatial overlap thresholds, and time window aggregation distributes discrete lightning events according to a fixed time interval statistical density distribution. Geometric correction of camera images eliminates lens distortion through calibration parameters, and brightness normalization uses histogram equalization to eliminate the influence of illumination variations. Abnormal pulse removal from ground electric field observation data and low-frequency electromagnetic observation data filters instantaneous spikes based on statistical thresholds, and baseline correction eliminates low-frequency drift using a moving average method. Quality scoring is calculated comprehensively based on the integrity of the data source, noise level, and processing errors. Missing data markers are used to identify data segments that failed verification or could not be obtained. Fixed channel order is arranged according to priority: radar, lightning, camera, electric field, electromagnetic, and fixed time order is aligned with timestamps at the minute level.

[0105] Specifically, after ground object obstruction correction, radar echo images accurately reflect the true echo intensity distribution within the target area, avoiding false strong echo interference caused by terrain uplift. Ground clutter suppression further eliminates reflection noise from fixed ground objects, improving the detection sensitivity of dynamic storm targets. Lightning location data, through event deduplication and time window aggregation, aggregates discrete lightning points into a continuous density field, reducing the impact of data sparsity on spatiotemporal feature extraction. Geometric correction and brightness normalization of camera images ensure spatial alignment and grayscale consistency of image data under different viewing angles and lighting conditions, facilitating subsequent multi-source data fusion. Abnormal pulse removal and baseline correction of ground electric field and electromagnetic data suppress sensor transient interference and baseline drift, preserving slowly varying signal characteristics related to thunderstorm activity. Quality scoring and missing data marking provide data reliability indicators for downstream processing, avoiding the participation of low-quality or missing data in calculations. Fixed channels and time order ensure strict alignment of multi-source data in the spatiotemporal dimension, and missing data marking replaces missing data to avoid false information introduced by interpolation. Thus, the preprocessed unified spatiotemporal data meets standardization requirements in spatiotemporal alignment, noise suppression, and missing data handling.

[0106] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0107] Multi-source observation data includes radar echo images, lightning location data, camera images, ground electric field observation data, and low-frequency electromagnetic observation data. Ground object obstruction correction and ground clutter suppression are performed on the radar echo images. Event deduplication and time window aggregation are performed on the lightning location data. Geometric correction and brightness normalization are performed on the camera images. Anomaly pulse removal and baseline correction are performed on the ground electric field observation data and low-frequency electromagnetic observation data. A quality score and missing data marker are generated for each time step. A unified spatiotemporal data set is constructed according to a fixed channel order and a fixed time order. When any data source is missing, the missing data marker is used instead of interpolation.

[0108] Specifically, ground feature occlusion correction in radar echo images is achieved by establishing a terrain occlusion model, and ground clutter suppression employs Doppler filtering. Event deduplication in lightning location data is based on a spatiotemporal clustering algorithm, and time window aggregation uses a sliding window method. Geometric correction of camera images utilizes control point registration, and brightness normalization employs histogram equalization. Abnormal pulse removal from ground electric field observation data and low-frequency electromagnetic observation data is based on median filtering, and baseline correction uses piecewise linear fitting. Quality scoring is comprehensively evaluated using indicators such as signal-to-noise ratio and data integrity, and missing data markers use binary encoding. The construction of unified spatiotemporal data adopts a multidimensional array structure to ensure the spatiotemporal consistency and integrity of the data.

[0109] Through the above technical solutions, this application achieves unified preprocessing and spatiotemporal alignment of multi-source heterogeneous data, improving the data quality and reliability of subsequent feature extraction and early warning analysis. Ground object occlusion correction and ground clutter suppression enhance the effective information of radar echo data. Event deduplication and time window aggregation improve the spatiotemporal consistency of lightning location data. Geometric correction and brightness normalization improve the comparability of camera images. Abnormal pulse removal and baseline correction improve the signal quality of electric field and electromagnetic data. The unified spatiotemporal data structure facilitates the fusion and analysis of multi-source data.

[0110] In some of the schemes mentioned above in this application, although the preprocessed multi-source observation data forms unified spatiotemporal data, the features under different time windows and spatial scales cannot be distinguished, which leads to the obscuring of key information when extracting storm features and affects the accuracy of the early warning results.

[0111] This application further proposes to extract storm features from unified spatiotemporal data, including: extracting rapidly changing features within a first time window and continuously changing features within a second time window, where the first time window is shorter than the second time window. Energy distribution within the target area and risk corridor is converged at the first and second spatial scales, respectively, where the first spatial scale is shorter than the second spatial scale. The propagation coherence and boundary steepness along the risk corridor are enhanced. The influence of low-quality and missing information is suppressed based on quality scores and missing data markers, resulting in storm features including local intensity, boundary steepness, propagation coherence, and upstream persistence.

[0112] The first time window was set to 5–10 minutes to capture sudden changes in lightning activity, and the second time window was set to 20–30 minutes to analyze the continuous evolution of the storm system. The first spatial scale used a 500-meter resolution grid to statistically analyze the electric field intensity gradient within the target area, while the second spatial scale used a 2-kilometer resolution grid to calculate the cumulative echo energy within the risk corridor. Propagation coherence was calculated by correlating the lightning density distribution between adjacent time steps, and the steepness of the boundary was quantified by the rate of change of the radar echo gradient. Data channels with quality scores below a preset threshold were replaced with mean filler, and data locations corresponding to missing data markers were directly set to zero.

[0113] Specifically, rapid change characteristics are achieved through short-term abrupt changes in lightning location data, while continuous change characteristics are achieved through trend analysis of radar echo intensity. Energy convergence at different spatial scales is achieved using a sliding window statistical method, with small-scale focusing on the core energy of the target area and large-scale monitoring of the overall situation of the risk corridor. Enhanced propagation coherence employs a spatiotemporal continuity constraint algorithm to interpolate and complete fragmented lightning activity trajectories. Enhanced boundary steepness is achieved by strengthening radar echo front characteristics through edge detection operators. Low-quality data suppression uses a weighted averaging method to reduce its contribution to the feature vector, and feature values ​​at missing data locations are compensated through interpolation of adjacent units. The resulting storm characteristics encompass both early warning signals of short-term abrupt changes and the evolutionary trend of the storm system.

[0114] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0115] Rapidly changing features are extracted from unified spatiotemporal data within a first time window, and continuously changing features are extracted within a second time window. The first time window is set to 5 minutes, and the second time window is set to 30 minutes. Energy distribution within the target area and risk corridor is converged at the first and second spatial scales, respectively. The first spatial scale is set to a 1 km × 1 km grid, and the second spatial scale is set to a 5 km × 5 km grid. The propagation coherence and boundary steepness along the risk corridor are enhanced. The influence of low-quality and missing information is suppressed based on quality scores and missing data markers to obtain storm characteristics. Storm characteristics include local intensity, boundary steepness, propagation coherence, and upstream persistence.

[0116] Specifically, for rapidly changing characteristics, the rate of change of radar echo intensity, lightning frequency, and ground electric field over a 5-minute period is calculated. For continuously changing characteristics, the trend of change of radar echo area and lightning activity range over a 30-minute period is calculated. The radar echo intensity distribution and lightning density distribution within the target area are calculated on a 1 km × 1 km grid. The radar echo intensity distribution and lightning density distribution within the risk corridor are calculated on a 5 km × 5 km grid.

[0117] Furthermore, the correlation coefficients of radar echoes at adjacent times are calculated along the risk corridor to enhance propagation coherence. Gradient values ​​at the edges of radar echoes are calculated to enhance boundary steepness. For data with quality scores below a threshold or missing markers, their weights are reduced during feature extraction. This yields storm features that include local intensity, boundary steepness, propagation coherence, and upstream persistence.

[0118] Through the above technical solution, this application achieves spatiotemporal feature extraction from multi-source observation data within a target area. By setting different time windows and spatial scales, it captures the rapid and continuous changes in storm characteristics, as well as the energy distribution at different spatial scales. Enhanced propagation coherence and boundary steepness improve the ability to characterize storm evolution trends. By suppressing the influence of low-quality and missing information, the reliability of feature extraction is improved. The storm features obtained thus comprehensively reflect the storm's local intensity, boundary characteristics, propagation properties, and persistence.

[0119] In some of the solutions described above in this application, when selecting cross-modal correlations based on target feature descriptions and storm features, there is a problem of unreasonable weight allocation due to the failure to consider the structural attributes of the target area, historical lightning records, and differences in data quality, which may lead to misjudgment or omission.

[0120] This application further proposes determining sensitive directions and sensitive sectors based on structural classification, grounding classification, and windward exposure descriptions in the target feature description. Storm features located in sensitive sectors and varying along the risk corridor are given increased weight, while storm feature data corresponding to locations and time periods marked in historical false alarm and missed alarm records are given decreased weight. Low-quality and missing information is suppressed based on quality scores and missing data markers. Correlation scores for each location and time period are calculated, correlation thresholds are set and ranked, and locations and time periods corresponding to storm feature data exceeding the correlation threshold are retained. The storm feature data for each location and time period are then aggregated to form targeted features for the target area.

[0121] Sensitive sectors are determined through windward exposure descriptions of the target area; for example, the dominant sector marked in the windward exposure description is defined as a sensitive sector. Changes in storm characteristics along the risk corridor are quantified using propagation coherence and upstream persistence indicators; for example, propagation coherence is calculated using the spatial overlap rate of lightning activity within consecutive time steps. Locations and time periods marked in historical false alarm and missed alarm records are linked to current storm characteristic data using timestamps and geographic coordinates; for example, the spatial coordinates corresponding to false alarm records within a certain time period are marked as low-weight areas. Quality scores are generated based on the preprocessing results of multi-source observation data; for example, the signal-to-noise ratio of radar echo data after ground cover correction is used as part of the quality score. Correlation scores are calculated by the matching degree between weighted storm characteristics and target characteristic descriptions; for example, the degree of conductivity exposure in structural grading is multiplied by the local intensity in storm characteristics. The correlation threshold is dynamically adjusted based on the ratio of correct warnings to false alarms in historical data; for example, when the historical false alarm rate exceeds 10%, the correlation threshold is increased by 5%.

[0122] Specifically, the determination of sensitive sectors combines the windward exposure description of the target area. For example, if the dominant sector of a target area is northeast, storm features in the northeast direction are given higher weight. Changes in storm features along the risk corridor are assessed using propagation coherence indices. For instance, when the spatial overlap of lightning activity reaches 80% in adjacent time steps, the propagation coherence score is improved. Storm feature data corresponding to locations and time periods marked in historical false alarm records are weighted less. For example, if a region experiences three false alarms in the past three months, the weight of the current storm feature data for that region is reduced by 30%. Quality scores and missing data markers are used to suppress the impact of low-quality data. For example, if radar echo data at a certain time step has a quality score below the threshold due to ground clutter suppression failure, the corresponding storm feature is removed. Correlation scores are calculated through multi-dimensional matching, such as a linear weighted sum of the height level in structure grading and the boundary steepness in storm features. Storm feature data exceeding the correlation threshold are retained and aggregated; for example, the top 20% of data with the highest correlation scores within a certain time period are selected as target features. Therefore, by dynamically adjusting weights, suppressing the influence of low-quality data and historical misjudgment areas, and combining the matching degree between target attributes and storm characteristics to screen key data, the effectiveness of targeted features can be improved, thereby increasing the accuracy of early warning results.

[0123] As a preferred embodiment, the solution of this application is specifically implemented as follows:

[0124] Sensitive directions and sectors are determined based on the structural classification, grounding classification, and windward exposure description in the target feature description. Storm features located within sensitive sectors and varying along the risk corridor are given increased weight, while storm feature data corresponding to locations and time periods marked in historical false alarm and missed alarm records are given decreased weight. Low-quality and missing information is suppressed based on quality scores and missing data markers. Correlation scores are calculated for each location and time period, correlation thresholds are set, and rankings are performed. Locations and time periods corresponding to storm feature data exceeding the correlation threshold are retained, and the storm feature data for each location and time period are aggregated to form targeted features for the target area.

[0125] Specifically, the sensitive direction and sensitive sector are first determined based on the structural classification, grounding classification, and windward exposure description in the target feature description. For example, for a high-rise building, its structural classification is high, its grounding classification is good, and its windward exposure description shows high exposure in the southwest direction. Therefore, the southwest direction is determined as the sensitive direction, and the 120-degree sector in the southwest direction is the sensitive sector.

[0126] Secondly, increase the weight of storm features located in sensitive sectors and changing along the risk corridor. For example, increase the weight of storm features within sensitive sectors by 50%. Reduce the weight of storm feature data corresponding to locations and time periods marked in historical false alarms and missed alarm records, such as by 30%. At the same time, suppress low-quality and missing information based on quality scores and missing data markers, such as reducing the weight of data with a quality score below 0.6 by 80%.

[0127] Furthermore, the correlation scores between each location and each time period are calculated. The correlation score can be obtained by calculating the similarity between storm features and target feature descriptions. A correlation threshold of 0.7 is set, and the correlation scores are ranked, retaining the locations and time periods corresponding to storm feature data exceeding the threshold.

[0128] Finally, the storm characteristic data for the retained locations and time periods are aggregated to form targeted features for the target region. Targeted features include multi-dimensional feature vectors such as local intensity, boundary steepness, propagation coherence, and upstream persistence.

[0129] Through the above technical solution, this application achieves targeted feature construction for specific objectives. This improves the pertinence and accuracy of early warnings, reduces false alarms and missed alarms, and enhances the interpretability of early warning results. Simultaneously, by comprehensively considering multi-dimensional features, the comprehensiveness and reliability of early warnings are improved. Furthermore, by adjusting the weight of historical records, the early warning system possesses optimization capabilities, continuously improving its effectiveness with use.

[0130] In some of the solutions mentioned above in this application, the fusion results may be inconsistent due to the difference in confidence levels of multi-source data when generating early warning results, and the fixed early warning threshold cannot adapt to the historical false alarms and missed alarms in different target areas, thus affecting the accuracy of early warning.

[0131] This application further proposes generating early warning results within a predetermined time window based on target characteristics. This includes statistically analyzing and ranking the target characteristics, obtaining a preliminary assessment of the probability of lightning strike risk based on local intensity, boundary steepness, propagation coherence, and upstream persistence, determining the arrival time along the risk corridor, and generating a preliminary assessment of the spatial risk distribution within the target area. Fusion weights are determined by combining the confidence and quality scores of each data point from multi-source observation data. Evidence fusion is performed according to consistency correction rules to obtain the fused probability of lightning strike risk, fused arrival time, and fused spatial risk distribution. When determining the early warning threshold, it is corrected based on historical false alarm records and historical missed alarm records of the target area.

[0132] The preliminary assessment uses local intensity to reflect the intensity of lightning activity within the target area, boundary steepness to characterize the abrupt change in risk distribution, propagation coherence to measure the continuity of storm movement along the risk corridor, and upstream persistence to assess the duration of lightning activity in the upstream region. The fusion weights are dynamically adjusted based on the quality score of radar echo images, the event deduplication results of lightning location data, and the baseline correction results of ground electric field observations; for example, the weight of radar echo images is reduced when their quality score is below a preset threshold. Evidence fusion employs Dempster-Shafer theory, transforming the confidence levels of multi-source data into a basic probability allocation function and eliminating conflicting evidence through synthesis rules. The warning threshold correction is dynamically adjusted based on the ratio of false alarms to missed alarms in the target area over the past 30 days; for example, the threshold is increased by 5%–10% when the number of false alarms exceeds the number of missed alarms.

[0133] Specifically, in the preliminary assessment phase, the cumulative energy of lightning activity is calculated by statistically analyzing the local intensity of the target features, the risk spread rate is judged by combining the steepness of the boundary, and the stability of the storm's movement path is predicted using propagation coherence. For example, when the boundary steepness exceeds 45 degrees / km and the propagation coherence lasts for more than 3 time steps, the probability of being struck by lightning in the preliminary assessment increases by 20%. In the fusion phase, if the radar echo image's quality score is below 60 points due to ground cover obstruction, its weight is reduced to 0.3, while the weight of lightning location data is set to 0.5 after time window aggregation. During evidence fusion, to address conflicts in the spatial risk distribution between radar and lightning data, the hit records of lightning location data within the target area are prioritized. The warning threshold is dynamically adjusted based on historical records. For example, if a target area had 2 missed warnings last week, the threshold is reduced by 8% this week; if there were 3 false alarms, it is increased by 12%. Thus, through multi-source data fusion and threshold adaptive mechanisms, the temporal accuracy and spatial hit rate of the warning results are improved.

[0134] As a preferred embodiment, the specific implementation of this application is as follows: Target features are statistically analyzed and sorted. A preliminary assessment value of the lightning risk probability is calculated based on local intensity, boundary steepness, propagation coherence, and upstream persistence. A preliminary assessment value of the arrival time is predicted using a linear regression model along the risk corridor. A preliminary assessment map of the spatial risk distribution is generated within the target area through kernel density estimation. The confidence level of radar echo data is set to 0.9, the confidence level of lightning location data is set to 0.85, and the confidence level of ground electric field observation data is set to 0.75. Data sources with a quality score below 0.6 are excluded. The DS evidence theory is used to fuse multi-source observation data. The fusion weights are adjusted by calculating the conflict factor to generate a fused lightning risk probability, a fused arrival time, and a fused spatial risk distribution. Based on the statistical results of the false alarm rate and missed alarm rate in the target area over the past three years, the warning threshold is dynamically adjusted from the initial value of 0.7 to 0.68. A warning signal is triggered when the fused lightning risk probability exceeds the adjusted threshold.

[0135] Through the above technical solutions, this application solves the problem of assessment bias caused by differences in confidence levels of multi-source data, and improves the credibility of early warning results through evidence theory fusion methods. A dynamic threshold adjustment mechanism based on historical false alarm and missed alarm records reduces the probability of false alarms and missed alarms in the target area caused by fixed thresholds. The spatiotemporal prediction method combining kernel density estimation and linear regression makes the assessment results of spatial risk distribution and arrival time more consistent with the actual storm evolution patterns.

[0136] In some of the schemes mentioned above in this application, the statistical and sorting process of target features may lead to deviations in the preliminary assessment of lightning risk probability, arrival time and spatial risk distribution due to differences in confidence or quality fluctuations in multi-source observation data. At the same time, the static setting of the warning threshold may not be able to adapt to the historical false alarms and missed alarms in different target areas, affecting the accuracy of the warning results.

[0137] This application further proposes to statistically analyze and rank target characteristics, obtain a preliminary assessment of the probability of lightning strike risk based on local intensity, boundary steepness, propagation coherence, and upstream persistence, determine a preliminary assessment of arrival time along the risk corridor, and generate a preliminary assessment of spatial risk distribution within the target area. Fusion weights are determined based on the confidence and quality scores of each data point from multi-source observation data. Evidence fusion is performed according to consistency correction rules to obtain the fused probability of lightning strike risk, fused arrival time, and fused spatial risk distribution. When determining the warning threshold, the warning threshold is corrected based on historical false alarm records and historical missed alarm records of the target area.

[0138] The statistical and ranking process quantifies the storm's energy concentration in the target area through local intensity, reflects the clarity of the storm boundary and its expansion potential through boundary steepness, assesses the storm's movement continuity along the risk corridor, and measures the storm's duration upstream of the target area through upstream persistence. The preliminary lightning risk probability is generated by weighting the above four indicators. The preliminary arrival time assessment is calculated based on the storm's movement speed along the risk corridor and the remaining path length from the target area. The preliminary spatial risk distribution assessment is generated by superimposing the energy distribution of each sub-region within the target area with historical lightning activity density. The confidence level of multi-source observation data is dynamically adjusted based on the inherent error range of the data source and the real-time quality score. For example, when the confidence level of radar echo images is higher than that of ground electric field observation data, their weight ratio is set to 0.6:0.4. Consistency correction rules refer to the priority order and difference threshold rules set for results from multiple sources within the same grid and at the same time step, based on quality scores, time alignment, historical stability, and scene adaptability. These rules are used to resolve conflicts between different data sources in time or space. For example, when the direction of storm movement shown by radar echoes deviates from the direction of lightning location data by more than 15 degrees, a direction consistency check is triggered, and the lightning location data is used preferentially. The warning threshold is dynamically adjusted based on the ratio of false alarms to missed alarms in the target area over the past 30 days. For example, when the number of false alarms exceeds the number of missed alarms, the warning threshold is increased by 10%.

[0139] Specifically, after statistical analysis of the target characteristics, local intensity is calculated using the spatial integration of radar echo reflectivity and lightning activity density; boundary steepness is assessed using the change in radar echo gradient and the sharpness of the lightning activity range edge; propagation coherence is measured by the standard deviation of the storm's movement direction within a continuous time window; and upstream persistence is determined by the proportion of upstream lightning activity in the target area lasting more than 5 minutes. The preliminary assessment of the lightning risk probability is weighted and normalized to a probability value of 0-1 using the four indicators. For example, local intensity has a weight of 0.4, boundary steepness 0.3, propagation coherence 0.2, and upstream persistence 0.1. The preliminary assessment of arrival time is calculated based on the storm's movement speed and the distance to the target area boundary. For example, when the speed is 20 m / s and the distance is 5 km, the arrival time is assessed as 250 seconds. The preliminary assessment of spatial risk distribution is generated by dividing the target area into 100 m × 100 m grids, calculating the product of lightning activity density and radar echo intensity for each grid, and overlaying it with a historical lightning activity heatmap. During the fusion process, the confidence level of multi-source observation data is jointly determined by real-time quality scores and historical error statistics. For example, if the quality score of radar echoes in a certain period is 0.9 and the historical error is 5%, its confidence weight is set to 0.7. If the quality score of ground electric field data is 0.8 but the historical error is 12%, its weight is set to 0.3. Evidence fusion adopts the Dempster-Shafer theory, synthesizing the evaluation results from different data sources as independent evidence. For example, when both radar and lightning data support a high risk of lightning strikes, the fusion probability increases to 0.85. The dynamic correction of the warning threshold is adjusted by the ratio of historical false alarm rate to missed alarm rate. For example, when the false alarm rate is 20% and the missed alarm rate is 10%, the threshold is increased from 0.7 to 0.75 to reduce false alarms.

[0140] As a preferred embodiment, the solution of this application is implemented as follows: When the probability of lightning strike exceeds a dynamically adjusted warning threshold and the spatial risk distribution overlaps with a preset distance zone outside the target area, a targeted warning information generation process is triggered. Risk levels are divided into four levels: Level 1 corresponds to a lightning strike probability of 30%–50% and an arrival time greater than 30 minutes; Level 2 corresponds to a probability of 50%–70% and an arrival time of 15–30 minutes; Level 3 corresponds to a probability of 70%–90% and an arrival time of 5–15 minutes; and Level 4 corresponds to a probability greater than 90% and an arrival time less than 5 minutes. Within a 15-minute time window corresponding to the arrival time, lightning location data is acquired in real time, and lightning events with a distance of less than 1 kilometer from the center point of the target area are selected. If the deviation between the occurrence time of the lightning event and the warning time window is less than 3 minutes, it is determined to be an actual hit. When a hit occurs, the warning is marked as a valid hit event, and statistical updates of historical false alarm records are excluded. If no lightning event meets the distance and time conditions within the time window, the warning data, along with its quality score and missing data marker, is simultaneously stored in the false alarm database. For lightning events that meet the correlation conditions but do not trigger the warning conditions, they are recorded in the missed alarm database and the target area is marked. When the time windows of multiple targeted warning messages overlap and cover the same target area, only the earliest triggered warning is correlated with the actual lightning event for verification; subsequent overlapping warnings will not undergo the same correlation determination process.

[0141] Through the above technical solution, this application solves the problem of statistical distortion in false alarms and missed alarms caused by the lack of correlation determination between early warning results and actual lightning events in existing technologies. It establishes accurate correlation rules by pre-setting dual conditions of distance and time deviation, and avoids duplicate counting by a one-time correlation mechanism, thus achieving accurate attribution of historical false alarms and missed alarms. It also optimizes adaptability to complex scenarios, reducing redundant calculations when multiple early warning signals overlap by using a time-priority principle, thereby improving the processing efficiency and data reliability of the early warning result verification process.

[0142] In summary, this application compresses wide-area thunderstorm activity into the most relevant spatiotemporal window to the target by using upstream risk corridors and target areas, reducing background interference unrelated to the target and lowering false alarms. Within the target area, radar, lightning location, and other accessible observations are unified into unified spatiotemporal data, maintaining data accuracy even under conditions of sensor gaps, noise, and scene switching. By extracting spatiotemporal features across multiple time windows and scales, the application synchronously characterizes sudden enhancement and persistent organizational structure, providing more accurate priors for arrival time and impact range. Using target feature descriptions as queries, the application performs cross-modal correlation selection and weighting of storm features, aligning storm evidence with the target's structure, grounding, and exposure attributes one by one, outputting the probability of lightning exposure and spatial risk distribution, achieving targeted discrimination of the same storm, different targets, and different risks. The application provides manageable early warnings, balancing accuracy and interpretability, improving hit rate and reducing false alarm rate.

[0143] Based on another preferred embodiment described above, see [link to preferred embodiment]. Figure 2 As shown, this embodiment provides a lightning intelligent targeted early warning system for key local targets, used to apply the above-mentioned lightning intelligent targeted early warning method for key local targets, including:

[0144] The acquisition unit is configured to acquire the geographical location, structural parameters, grounding parameters and surrounding terrain parameters of each target area to obtain a target feature description, which is used to characterize lightning vulnerability.

[0145] The first processing unit is configured to estimate storm motion information based on lightning location data and radar echo images at continuous time intervals, and generate a risk corridor in the upstream direction of each target area according to the storm motion information. The risk corridor is combined to form a target area around the target area, and the shape and scale of the target area are determined according to the storm motion information.

[0146] The second processing unit is configured to preprocess multi-source observation data within the target area to obtain unified spatiotemporal data. Spatiotemporal features are extracted from the unified spatiotemporal data to obtain storm features. Driven by the target feature description, cross-modal correlation selection and weighting are performed on the storm features to construct targeted features for the target area.

[0147] The early warning unit is configured to generate early warning results within a predetermined time window based on target characteristics. The early warning results include the probability of lightning strike risk, arrival time, and spatial risk distribution.

[0148] The early warning unit is also configured to generate targeted early warning information when the probability of lightning strike reaches the early warning threshold and the spatial risk distribution intersects with the buffer range of the target area. The targeted early warning information includes the identifier of the target area, the risk level, and the arrival time.

[0149] Specifically, the acquisition unit is configured to collect the geographical location, structural parameters, grounding parameters, and surrounding terrain parameters of the target area, and form a target feature description through time consistency verification, duplicate record cleanup, and outlier removal. The first processing unit is configured to estimate the storm's movement direction and speed based on lightning location data and radar echo images, generate a risk corridor by combining spatial expansion trends, and generate an elliptical target area upstream of the target area, with the major and minor axes dynamically adjusted according to the storm's movement speed and expansion trend, respectively. The second processing unit is configured to preprocess multi-source data such as radar echoes, lightning location, and camera images within the target area, extract spatiotemporal features such as local intensity and boundary steepness, and perform cross-modal correlation weighting by combining sensitive sectors and historical records in the target feature description. The early warning unit is configured to generate early warning information based on the fused lightning risk probability, arrival time, and spatial risk distribution. When the risk probability exceeds a threshold and the spatial distribution intersects with the target buffer range, a targeted early warning is triggered, and the historical records are updated in conjunction with the actual lightning hit results.

[0150] Specifically, the data acquisition unit integrates geographical parameters and historical lightning records of the target area to form structured features characterizing lightning vulnerability. The first processing unit determines the storm's trajectory by fusing radar and lightning location data, dynamically generating a risk corridor that matches the target area's location to ensure the target area covers the potential threat range. The second processing unit employs a multi-scale spatiotemporal feature extraction method, combining target sensitive directions to filter and weight features, enhancing the targeted assessment of lightning risk in the target area. The early warning unit optimizes the early warning threshold and correlates it with actual lightning events through multi-source data fusion and consistency correction rules, reducing the probability of false alarms and missed alarms. For example, when the risk corridor length is calculated to be 5 kilometers based on storm speed and a preset time window, the major axis of the ellipse of the target area expands to 8 kilometers, and the minor axis is adjusted to 3 kilometers based on the spatial expansion trend, ensuring coverage of the buffer zone outside the target area. The early warning unit further correlates risk level with probability and arrival time, triggering an early warning when the spatial risk distribution intersects with the buffer zone, and verifying the accuracy of the early warning through actual lightning hits, dynamically updating historical records to optimize subsequent assessments.

[0151] As a preferred embodiment, the solution of this application is implemented as follows: The system includes a data acquisition unit, a first processing unit, a second processing unit, and an early warning unit. The data acquisition unit obtains the boundary contour through the geographic coordinates of the target area, extracts the elevation by calling the geographic information system, accesses the structural parameter database to read the building height, exposed metal area, and grounding resistance value, and uses a three-dimensional terrain model to calculate the relative height and azimuth distribution of surrounding obstacles, forming a lightning vulnerability description that includes structural classification, grounding classification, and circumferential exposure distribution. The first processing unit receives the time series of lightning location data and the radar reflectivity image sequence, calculates the storm movement vector field using the optical flow method, determines the movement direction and expansion trend based on the principal component analysis of the vector field, and establishes an elliptical target area upstream of the target area whose major axis is proportional to the storm movement speed. Its boundary is fused with the radar echo extrapolation trajectory and the lightning density diffusion range. The second processing unit extracts the waveform features of the lightning electromagnetic pulse signal in the target area, simultaneously fuses the velocity spectrum data of the Doppler radar and the trend change data of the ground electric field instrument, and uses an attention mechanism to dynamically weight the data channels in the sensitive sector to generate a spatiotemporal feature matrix that includes energy gradient propagation and boundary abrupt change features. The early warning unit calculates the probability distribution of lightning events in the target area within the next ten minutes using a time series prediction model. When the probability exceeds the dynamic threshold and there is spatial overlap between the risk hot zone and the target buffer zone, it outputs an early warning message containing the target number, risk quantification value, and estimated arrival time, while updating the historical early warning hit rate statistics table.

[0152] Through the above technical solutions, this application can establish dynamic risk corridors and targeted monitoring areas for specific protection targets. By fusing spatiotemporal features of multi-source observation data and using a target-driven weighting mechanism, interference signals from irrelevant areas are suppressed, improving the accuracy of local lightning risk identification. The system achieves precise matching between early warning triggering conditions and the actual exposure status of targets through real-time correction of storm motion vectors and spatial superposition judgment of buffer zones, reducing the risk of missed reports due to sudden changes in storm paths. The time prediction function of early warning messages provides target units with an operable emergency response window, while the dynamic threshold calculation model is continuously optimized through a hit result feedback mechanism.

[0153] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for intelligent targeted early warning of lightning strikes for key local targets, characterized in that, include: The geographical location, structural parameters, grounding parameters and surrounding terrain parameters of each target area are collected to obtain a target feature description, which is used to characterize lightning vulnerability. Storm motion information is estimated based on lightning location data and radar echo images at continuous time intervals, and a risk corridor is generated in the upstream direction of each target area according to the storm motion information. The risk corridor is combined to form a target area around the target area, and the shape and scale of the target area are determined according to the storm motion information. Multi-source observation data within the target area are preprocessed to obtain unified spatiotemporal data; spatiotemporal features are extracted from the unified spatiotemporal data to obtain storm features; the target feature description is used as a driver to perform cross-modal correlation selection and weighting on the storm features to construct targeted features for the target area. Based on the target features, an early warning result is generated within a predetermined time window. The early warning result includes the probability of lightning strike risk, the arrival time and spatial risk distribution. When the probability of lightning strike reaches the warning threshold and the spatial risk distribution intersects with the buffer zone of the target area, targeted warning information is generated. The targeted warning information includes the identifier of the target area, the risk level, and the arrival time. The buffer zone is a preset distance band outside the target area.

2. The intelligent targeted early warning method for lightning strikes targeting key local areas according to claim 1, characterized in that, When obtaining the target feature description, the following are included: Determine the boundaries, geographical location and altitude of each target area, collect structural parameters, grounding parameters and surrounding terrain parameters, perform time consistency verification, clean up duplicate records and remove outliers, and mark missing data for data that cannot be verified. Based on digital elevation data and the height of nearby facilities, the 360° range is divided into at least four equal-angle sectors. The relative rise and shading angle of each sector are calculated to form a circumferential exposure distribution. Based on the multi-year wind direction statistics and historical lightning records of the target area, the dominant sector is identified and marked as a sensitive sector to form a windward exposure description. The grounding parameters are corrected according to the season and soil moisture; the structural parameters are classified according to height, degree of conductive exposure and down conductor type to obtain structural classification; the grounding parameters are classified according to preset intervals to obtain grounding classification; the historical lightning records are counted in layers according to the number of occurrences and distance to obtain historical record count; and the boundary, geographical location and altitude, structural classification, grounding classification, windward exposure description and historical record count of the target area are combined in sequence to form the target feature description.

3. The intelligent targeted early warning method for lightning strikes targeting key local areas according to claim 1, characterized in that, When estimating storm motion information based on lightning location data and radar echo images at continuous time intervals, the following is included: Align lightning location data at continuous intervals with radar echo images, the alignment including uniform grid and uniform time step; The radar observation results are determined based on the correlation matching of radar echoes at adjacent time points. The radar observation results include the radar observation direction of motion and the radar observation speed. The displacement direction and displacement velocity of the centroid of lightning activity density at adjacent moments are calculated to obtain the positioning observation results, which include the positioning observation motion direction and positioning observation motion velocity. The radar observation results are compared with the location observation results. When the radar observation results and the location observation results are consistent, the storm movement direction and storm movement speed are obtained by merging them; otherwise, the radar observation results are corrected according to the location observation results, and abnormal segments are marked. The spatial expansion trend is determined by the changes in radar echo profile area and lightning activity range, and the abnormal segments are removed to obtain the storm movement information.

4. The intelligent targeted early warning method for lightning strikes targeting key local areas according to claim 3, characterized in that, Based on the storm movement information, a risk corridor is generated upstream of each target area. When the risk corridor forms a target area around the target area, the following is included: A risk corridor is established upstream of each target area along the storm's movement direction. Its length is determined by a predetermined time window and the storm's movement speed, while its width is determined by the spatial expansion trend and historical prediction errors. Uncertainty buffer zones are set on both sides. When the risk corridor intersects with the buffer zone of the target area, a target area is generated around the target area. The target area is an elliptical region with its major axis along the storm's movement direction, where the major axis increases with the storm's movement speed, and the minor axis increases with the spatial expansion trend and the increase of the uncertainty buffer zone. When the corridor does not intersect, the monitoring status is recorded, and the target area is not generated.

5. The intelligent targeted early warning method for lightning strikes targeting key local areas according to claim 1, characterized in that, When preprocessing multi-source observation data within the target area to obtain unified spatiotemporal data, the process includes: The multi-source observation data includes radar echo images, lightning location data, camera images, ground electric field observation data, and low-frequency electromagnetic observation data. The radar echo images are subjected to ground object occlusion correction and ground clutter suppression; the lightning location data are subjected to event deduplication and time window aggregation; the camera images are subjected to geometric correction and brightness normalization; and the ground electric field observation data and low-frequency electromagnetic observation data are subjected to abnormal pulse removal and baseline correction. A quality score and missing measurement marker are generated for each time step, and the unified spatiotemporal data are constructed according to a fixed channel order and a fixed time order. When any data source is missing, the missing measurement marker is used instead of interpolation.

6. The intelligent targeted early warning method for lightning strikes targeting key local areas according to claim 5, characterized in that, When extracting storm features from the unified spatiotemporal data, the process includes: Based on the unified spatiotemporal data, rapid change features are extracted within a first time window, and continuous change features are extracted within a second time window, where the first time window is shorter than the second time window. Energy distribution within the target area and the risk corridor is converged at a first spatial scale and a second spatial scale, respectively, where the first spatial scale is shorter than the second spatial scale. The propagation coherence and boundary steepness along the risk corridor direction are enhanced. Furthermore, the influence of low-quality and missing information is suppressed based on the quality score and missing data markers to obtain the storm features, which include local intensity, boundary steepness, propagation coherence, and upstream persistence.

7. The intelligent targeted early warning method for lightning targeting key local areas according to claim 6, characterized in that, When constructing targeted features for the target region by driving cross-modal correlation selection and weighting on the storm features, using the target feature description as a driving force, the following are included: Sensitive directions and sectors are determined based on the structural classification, grounding classification, and windward exposure description in the target feature description. Storm features located in sensitive sectors and varying along the risk corridor are given increased weight, while storm feature data corresponding to locations and time periods marked in historical false alarm and missed alarm records are given decreased weight. Low-quality and missing information is suppressed based on the quality score and missing data markers. Correlation scores between each location and time period are calculated, correlation thresholds are set and ranked, and locations and time periods corresponding to storm feature data exceeding the correlation threshold are retained. The storm feature data for these locations and time periods are then aggregated to form targeted features for the target area.

8. The intelligent targeted early warning method for lightning strikes targeting key local areas according to claim 1, characterized in that, When generating an early warning result within a predetermined time window based on the target features, the following are included: The target features are statistically analyzed and sorted. A preliminary assessment of the lightning risk probability is obtained based on local intensity, boundary steepness, propagation coherence, and upstream persistence. A preliminary assessment of the arrival time is determined along the risk corridor direction. A preliminary assessment of the spatial risk distribution is generated within the target area. The fusion weight is determined based on the confidence and quality scores of each data point in the multi-source observation data. Evidence fusion is performed according to the consistency correction rule to obtain the fused lightning risk probability, fused arrival time, and fused spatial risk distribution. When determining the warning threshold, the warning threshold is corrected based on the historical false alarm records and historical missed alarm records of the target area.

9. The intelligent targeted early warning method for lightning strikes targeting key local areas according to claim 8, characterized in that, When generating targeted early warning information, the following are included: The risk level is positively correlated with the lightning risk probability, arrival time, and spatial risk distribution. Lightning location data is acquired within a predetermined time window corresponding to the arrival time, and the correlation between the actual lightning location hit result and the target area is determined based on a preset distance range and a preset time deviation. When there is a correlated actual lightning location hit result, the warning is marked as a hit and is not counted as a false alarm in the historical false alarm record. When there is no correlated actual lightning location hit result, the warning is counted in the historical false alarm record and the current quality score and data missing mark are retained. When an actual lightning location hit result correlated with the target area occurs in the same time window without meeting the triggering condition, it is counted in the historical missed alarm record. When multiple targeted warnings cover the same target area and the time windows overlap, the earliest triggered targeted warning information is correlated with the actual lightning location hit result once, and subsequent overlapping warnings are not correlated repeatedly.

10. A lightning intelligent targeted early warning system for key local targets, used to apply the lightning intelligent targeted early warning method for key local targets as described in any one of claims 1-9, characterized in that, include: The acquisition unit is configured to acquire the geographical location, structural parameters, grounding parameters and surrounding terrain parameters of each target area to obtain a target feature description, which is used to characterize lightning vulnerability. The first processing unit is configured to estimate storm motion information based on lightning location data and radar echo images at continuous time intervals, and generate a risk corridor in the upstream direction of each target area according to the storm motion information, and form a target area around the target area by combining the risk corridor, wherein the shape and scale of the target area are determined according to the storm motion information. The second processing unit is configured to preprocess the multi-source observation data within the target area to obtain unified spatiotemporal data; extract spatiotemporal features from the unified spatiotemporal data to obtain storm features; and use the target feature description to drive cross-modal correlation selection and weighting on the storm features to construct targeted features for the target area. The early warning unit is configured to generate an early warning result within a predetermined time window based on the target characteristics. The early warning result includes the probability of lightning strike risk, the time of arrival, and the spatial risk distribution. The early warning unit is further configured to generate targeted early warning information when the probability of lightning strike reaches an early warning threshold and the spatial risk distribution intersects with the buffer range of the target area. The targeted early warning information includes the identifier of the target area, the risk level, and the arrival time.

Images