Hierarchical lightning warning method based on generative network

Abstract

This invention discloses a hierarchical lightning warning method based on generative networks, belonging to the field of short-term lightning warning. To address the problems of existing short-term lightning warnings relying on a single or limited number of products, which makes it difficult to balance accuracy and lead time, lacking multi-source real-time constraint correction and hierarchical trigger release strategies, and being difficult to adaptively configure according to user preferences, this invention acquires multi-source real-time monitoring data, at least three types of short-term forecast products, and geographic information, and performs spatiotemporal alignment. It constructs atomic propositions and their trigger and release logics for the outer, middle, and inner rings, generates candidate parameters, calculates parameter response vectors, and quantizes them to form parameter prototype codes. A trajectory-balanced generative flow network is used to generate and optimize candidate strategies under constraints. The selected target strategy is evaluated, and hierarchical warning information is output based on real-time data. This achieves the technical effects of improving comprehensive warning performance and stability, reducing false alarms and missed alarms, and supporting interpretable and customizable configuration.

Description

Technical Field

[0001] This invention relates to the field of short-term lightning warning, and in particular to a hierarchical lightning warning method based on generative networks. Background Technology

[0002] Lightning is a significant hazardous factor in severe convective weather, characterized by its suddenness, small spatial scale, and rapid evolution. To mitigate the impact of lightning on industries such as power, communications, transportation, aviation, and outdoor operations, operational departments typically implement short-term, imminent lightning warnings. Existing short-term lightning warning technologies have evolved from single-source monitoring to multi-source fusion: early warnings relied primarily on lightning location data and ground-based atmospheric electric field observations; subsequently, indicators such as weather radar echo intensity, echo top height, and storm structure characteristics were introduced to identify and track thunderstorm bodies; in recent years, extrapolation-based thunderstorm cell tracking products, numerical model-based short-term forecast products, and machine learning-based probabilistic forecast products have been continuously enriched. Operationally, multiple short-term forecast products are often combined with multi-source real-time monitoring data to improve warning lead time and coverage, forming a tiered warning and regionalized dissemination application model.

[0003] However, existing technologies still have the following shortcomings, which limit further improvement in comprehensive early warning performance: First, most schemes still rely on single forecast products or simply overlay and vote on a small number of products, making it difficult to achieve both high accuracy and sufficient lead time under different regional, seasonal, and thunderstorm type conditions. This can easily lead to a situation where accuracy and timeliness are mutually constrained.

[0004] Second, there is a lack of systematic synergy, integration, and constraint correction mechanisms between multi-source real-time monitoring data and various short-term forecast products, making it difficult to effectively leverage the complementary advantages of different products and suppress their disadvantages, resulting in insufficient early warning stability and cross-scenario generalization capabilities.

[0005] Third, existing tiered early warning strategies tend to focus on trigger judgments and lack corresponding hierarchical release rules and logical nesting constraints. The early warning status is prone to frequent fluctuations, and it is difficult to perform batch and adaptive parameter configuration and optimization based on different users' preferences for false alarms, missed alarms, and lead time.

[0006] Therefore, a lightning early warning method is needed to address the shortcomings of existing technologies. Summary of the Invention

[0007] One objective of this invention is to propose a hierarchical lightning early warning method based on generative networks. Addressing the problems in existing technologies, such as reliance on simple superposition of single or a few products for short-term lightning warnings, which makes it difficult to balance accuracy and lead time; lack of systemic collaborative constraints for correction between multi-source real-time data and multiple types of short-term forecasts; and lack of paired trigger and de-activation decisions for hierarchical warnings, making adaptive optimization based on user preferences difficult, this invention proposes a method that acquires multi-source real-time monitoring data, at least three types of short-term forecast products, and geographic information, performs quality control and spatiotemporal alignment, constructs a three-layered structure (outer, middle, and inner circles) and atomic propositions, forms trigger and de-activation logic for each circle, enumerates or samples candidate parameters, calculates parameter response vectors on historical samples, and generates parameter prototype codes through finite quantization, uses a trajectory balancing generation flow network to generate candidate strategies under circle constraints and logical nesting constraints, iteratively optimizes candidate strategies based on warning scores, and finally selects the target strategy for real-time sample output of hierarchical warning information. This invention has the technical effects of improving comprehensive warning performance and stability, reducing false alarms and missed alarms, and supporting interpretable and customizable configurations for different user preferences.

[0008] This invention provides a hierarchical lightning early warning method based on generative networks, comprising: S1. Acquire multi-source real-time monitoring data, at least three types of short-term forecast products, and geographic information for the area to be warned. Preprocess the data to obtain historical and real-time spatiotemporal sample sequences. S2. Set up three concentric circles (outer, middle, and inner circles). Construct atomic propositions containing multi-source real-time monitoring data and / or short-term forecast products and their threshold parameters. Logically combine these atomic propositions to obtain the triggering and deactivation logic for each circle, forming a set of strategy templates and / or a strategy space for generating candidate strategies. S3. Under the premise of satisfying the concentric circle constraints and logical nesting constraints, use the radius of the three concentric circles, the configurable threshold parameters of the atomic propositions, and the selection and logical combination method of the atomic propositions as candidate parameters. Generate candidate strategies according to preset enumeration rules and / or sampling rules. S4. Select a parameter set, calculate the parameter response vector of each group of candidate parameters in the candidate parameter set on the historical spatiotemporal sample sequence, and perform finite-level quantization to form a parameter prototype code set, and establish a mapping relationship between the parameter prototype code and the candidate parameters; S5. Under the constraints of concentric circles and logical nesting, use a trajectory balancing generation flow network to generate candidate strategies on the parameter prototype code set, and use the warning score on the historical spatiotemporal sample sequence as a reward to update the trajectory balancing generation flow network; S6. Perform a final evaluation of the candidate strategies, select the target strategy based on the warning score, and restore the currently effective strategy according to the mapping relationship; S7. Execute the currently effective strategy on the real-time spatiotemporal sample sequence, and output the graded warning information and corresponding graded warning signals for the outer circle, middle circle, and inner circle.

[0009] Optionally, S1 includes: Collect radar data, lightning data, and atmospheric electric field data covering the target area as multi-source real-time monitoring data; collect short-term forecast products from at least three different sources or with different mechanisms; and collect geographic information of the target area. The short-term forecast products include thunderstorm body identification and tracking information; The multi-source real-time monitoring data and short-term forecast products are subjected to quality control processes including missing data identification, outlier removal, and consistency checks. The data processed through quality control is converted to a spatial coordinate reference consistent with the geographic information of the target area, and time synchronization is performed according to a preset time step. The spatial location is mapped to a preset spatial grid of the target area, and the spatiotemporal alignment of the multi-source real-time monitoring data, short-term forecast products and geographic information of the target area is completed. Based on the spatiotemporally aligned data, a spatiotemporal sample sequence containing multi-source features at each time step is constructed in chronological order. Historical spatiotemporal sample sequences and real-time spatiotemporal sample sequences are obtained according to the historical sample time range and the real-time business time range, respectively.

[0010] Optionally, S2 includes: Based on historical spatiotemporal sample sequences, the initial value ranges of the radii of the outer, middle, and inner circles are determined. Within these initial value ranges, the radii of the outer, middle, and inner circles are initialized, ensuring that the radius of the outer circle is greater than that of the middle circle, and the radius of the middle circle is greater than that of the inner circle. The outer, middle, and inner circles are buffer zones generated in a geographic coordinate system with reference to the target area boundary and / or warning reference points within the area to be warned. The radii of the outer, middle, and inner circles are the horizontal distances in the geographic coordinate system. Based on historical spatiotemporal sample sequences, real-time atomic propositions, single-time warning atomic propositions, and adjacent-time warning atomic propositions are defined, where real-time atomic propositions are used to characterize multiple events in the historical spatiotemporal sample sequences. Whether the source real-time monitoring data meets the real-time conditions, the single-time early warning atomic proposition is used to characterize whether the short-term forecast products in the historical spatiotemporal sample sequence meet the early warning conditions in a single time period, and the adjacent-time early warning atomic proposition is used to characterize whether the short-term forecast products in the historical spatiotemporal sample sequence meet the persistence conditions in adjacent time periods, and each atomic proposition includes at least one configurable threshold parameter; the single-time early warning atomic proposition and the adjacent-time early warning atomic proposition include early warning conditions set for the at least three types of short-term forecast products respectively, and the early warning conditions of the at least three types of short-term forecast products are fused through at least one fusion method of logical OR operation, logical AND operation, and weighted voting fusion method to determine the single Whether the time-based early warning atomic proposition and the adjacent time-based early warning atomic propositions are satisfied; for the outer circle, middle circle and inner circle respectively, at least one atomic proposition is selected from the real-time atomic proposition, single-time early warning atomic proposition and adjacent time-based early warning atomic proposition, and corresponding triggering logic expression and deactivation logic expression are constructed through logical operations to form the strategy template set and / or strategy space; the trigger determination of the outer circle, middle circle and inner circle includes: based on the spatiotemporally aligned multi-source real-time monitoring data and / or short-term forecast products, a thunderstorm indication binary field corresponding to each atomic proposition is generated, the thunderstorm indication binary field is used to characterize the spatial area within the target area that satisfies the constraint condition of the atomic proposition; when the atomic proposition When generating adjacent time-series warning atomic propositions, firstly, based on the short-term forecast products, single-time thunderstorm indicator binary fields are generated for at least two adjacent time-series. Then, a logical AND operation is performed point-by-point on the single-time thunderstorm indicator binary fields according to their spatial location on the same spatial grid, and / or the single-time thunderstorm indicator binary fields are accumulated point-by-point within a preset time window and compared with a persistence threshold to obtain the thunderstorm indicator binary fields corresponding to the adjacent time-series warning atomic propositions. Morphological processing is performed on each thunderstorm indicator binary field to eliminate isolated noise and fill voids. Connectivity analysis is then performed on each morphologically processed thunderstorm indicator binary field to mark spatially continuous regions as the thunderstorm indicator feature regions corresponding to the respective atomic propositions.For each atomic proposition, if the corresponding thunderstorm indicator feature region has a non-empty intersection with the spatial region of the corresponding concentric circle, the atomic proposition is determined to be satisfied, and the determination result of the atomic proposition is recorded as 1; otherwise, it is recorded as 0. The determination results of all the atomic propositions are substituted into the triggering logic expression of the corresponding concentric circle in the strategy template set and / or strategy space for Boolean operation to obtain the comprehensive logic operation result. When the comprehensive logic operation result is 1, the corresponding concentric circle is determined to satisfy the triggering condition.

[0011] Optionally, S3 includes: For each set of candidate parameters generated under the premise of satisfying the concentric circle constraints and logical nesting constraints, which consists of the radius of each circle, the configurable threshold parameters of the atomic propositions, and the selection and logical combination of the atomic propositions, triggering and de-activation judgments are performed on the outer circle, middle circle, and inner circle respectively based on the historical spatiotemporal sample sequence to obtain the triggering time sequence and de-activation time sequence corresponding to each circle. The parameter response vector is calculated based on the triggering time sequence and de-activation time sequence. The advancing relationship of the outer circle, middle circle, and inner circle is characterized by the order of the triggering times of the outer circle, middle circle, and inner circle, as well as the time difference between adjacent triggering times. The sequential correspondence between forecast entry into the circle and actual enhancement is characterized by the time difference between the moment when the short-term forecast product meets the entry condition and the moment when the multi-source actual monitoring data meets the actual enhancement condition. The synergistic enhancement relationship between lightning data, radar data, and atmospheric electric field data is characterized by the number of times the lightning data, radar data, and atmospheric electric field data simultaneously satisfy the corresponding atomic propositions within a preset time window. The continuity relationship between adjacent time intervals is characterized by the duration of consecutive time intervals satisfying the corresponding atomic propositions. The parameters are then fine-tuned. The state maintenance relationship is characterized by the consistency of the judgment results before and after the parameter fine-tuning of the candidate parameters with a preset step size. The parameter response vector also includes a proposition selection mode component and a logic structure mode component. The proposition selection mode component is encoded by the identifier set of the atomic propositions selected in the candidate parameters, and the logic structure mode component is encoded by the number of the logic expression structure template corresponding to the logic combination method in the candidate parameters, forming a parameter response set. Each scalar component of the parameter response vector in the parameter response set is subjected to scalar quantization processing with a finite number of quantization levels according to a preset quantization interval, and the quantized scalar components are combined to form a parameter prototype code. Multiple candidate parameters with the same parameter prototype code are merged into the same parameter prototype code to form a parameter prototype code set. At the same time, a mapping relationship between the parameter prototype code and the candidate parameter set is established. For each parameter prototype code, a representative candidate parameter is selected in its corresponding candidate parameter set according to a preset representative rule. The preset representative rule includes: selecting the candidate parameter with the smallest distance between the parameter response vector and the center of the corresponding quantization interval of the parameter prototype code, and / or selecting the candidate parameter with the median value of each parameter. Furthermore, the parameter response vector also includes a product divergence component, which is characterized by the spatial consistency index among the thunderstorm indication binary fields of the at least three types of short-term forecast products at the same time. The product divergence component is then involved in the scalar quantization process of the finite quantization series so that the parameter prototype code represents the multi-product consistency case and the multi-product divergence case. Furthermore, the parameter response vector also includes a missing detection robustness component and a computational cost component. The missing detection robustness component is characterized by the consistency between the trigger judgment result after performing a preset missing detection simulation on the multi-source real-time monitoring data and the trigger judgment result before the missing detection simulation. The computational cost component is characterized by the unit time step consumption of performing the trigger judgment and de-judgment judgment in a preset hardware environment. The missing detection robustness component and the computational cost component are then involved in the scalar quantization processing of the finite quantization level.

[0012] Optionally, S4 includes: Based on the parameter prototype code set, the parameter prototype code is determined as the parameter selection unit of the trajectory balancing generation flow network. The trajectory balancing generation flow network selects the parameter prototype codes of the corresponding circles layer by layer in the order of outer circle, middle circle and inner circle, and then combines them to obtain each complete strategy in the candidate strategy set. Each parameter prototype code represents at least the radius parameter of the corresponding layer, the selection result of the atomic proposition, the logical combination method, and the threshold parameter. The layer constraints include the outer circle's radius being greater than the middle circle's radius and the middle circle's radius being greater than the inner circle's radius. The logical nesting constraints include the middle circle's triggering logic expression being allowed to be satisfied only when the outer circle is in a triggered state, and the inner circle's triggering logic expression being allowed to be satisfied only when the middle circle is in a triggered state. For each complete strategy, after mapping the parameter prototype code to representative candidate parameters according to the mapping relationship, triggering is performed on the outer circle, middle circle, and inner circle respectively according to the historical spatiotemporal sample sequence. The system determines and de-determines to obtain tiered early warning results, calculates an early warning score based on the tiered early warning results, and uses the early warning score as a trajectory reward for the trajectory balancing generation flow network, so that the trajectory balancing generation flow network outputs the candidate strategy set; the early warning score includes at least two of the following: hit reward, missed report penalty, false alarm penalty, and early warning lead time reward, and uses configurable weights to weight and combine the included items; it also receives early warning performance preference parameters from target users, the early warning performance preference parameters including accuracy weight and timeliness weight, and determines the configurable weights based on the accuracy weight and timeliness weight.

[0013] Optionally, S5 includes: For each complete strategy in the candidate strategy set, trigger and de-trigger judgments are performed in the outer, middle, and inner circles according to historical spatiotemporal sample sequences to obtain corresponding graded early warning results. Based on the graded early warning results, an early warning score is calculated for each complete strategy, and complete strategies whose early warning scores satisfy the preset score constraints are selected. When multiple complete strategies satisfy the preset score constraints, the complete strategy with the highest early warning score is selected as the target strategy. For the target strategy, according to the mapping relationship, the parameter prototype code in the target strategy is mapped to its corresponding representative candidate parameters. The representative candidate parameters include the radius values ​​of the outer, middle, and inner circles, the configurable threshold parameter values ​​of atomic propositions, and the selection results and logical combination methods of atomic propositions, thereby parameterizing the target strategy into the currently effective strategy. The preset score constraints include at least one of the following: early warning score not less than a preset score threshold, missed detection index not greater than a preset missed detection threshold, false alarm index not greater than a preset false alarm threshold, and early warning lead time not less than a preset lead time threshold.

[0014] Optionally, S6 includes: Based on the real-time spatiotemporal sample sequence, extract the real-time features corresponding to the outer circle, middle circle and inner circle, and according to the trigger logic expression and deactivation logic expression of the outer circle, the trigger logic expression and deactivation logic expression of the middle circle and the trigger logic expression and deactivation logic expression of the inner circle in the current effective strategy, execute the trigger judgment and deactivation judgment respectively for the outer circle, middle circle and inner circle to update the outer circle state, middle circle state and inner circle state. When the outer ring is triggered, it outputs a warning message at the attention level; when the middle ring is triggered, it outputs a warning message at the alert level; and when the inner ring is triggered, it outputs a warning message at the alarm level. Based on the attention-level graded early warning information, the alert-level graded early warning information, and the alarm-level graded early warning information, the alarm device is driven to issue attention-level graded early warning signals, alert-level graded early warning signals, and alarm-level graded early warning signals, respectively. Furthermore, it also includes: estimating the speed and direction of thunderstorm movement based on the positional changes of thunderstorm cells in the short-term forecast products used in the current effective strategy at different forecast lead times, and calculating the expected entry time of the warning reference point based on the radius of the outer circle, the radius of the middle circle, the radius of the inner circle, and the speed of thunderstorm movement, and outputting the expected entry time along with the attention-level warning information, the alert-level warning information, and the alarm-level warning information.

[0015] The beneficial effects of this invention are: 1. By performing quality control and spatiotemporal alignment of multi-source real-time monitoring data with at least three types of short-term forecast products, and using atomic propositions and their logical combinations as a unified expression framework, it is possible to achieve complementary advantages and suppression of disadvantages among multiple products under different regions, seasons and thunderstorm types, thereby improving the accuracy and generalization stability of early warning.

[0016] 2. By setting up three concentric circles (outer, middle, and inner) and constructing triggering and deactivation logic for each circle, while introducing nested logic constraints and persistence determination, the system can utilize the evolutionary pattern of thunderstorms from far to near and from weak to strong to reduce false alarms and missed alarms, reduce frequent fluctuations in warning status, and improve the timeliness and reliability of warnings.

[0017] 3. By quantifying the response of candidate parameters on historical samples to form parameter prototype codes, and using a trajectory balancing generation flow network to automatically generate and optimize strategies based on early warning scores under constraints, and setting score weights according to user preferences, the batch adaptive configuration of circle radius, threshold and logical combination is realized, reducing deployment and maintenance costs and improving strategy interpretability and reusability. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 The flowchart shows a hierarchical lightning early warning method based on generative networks. Figure 2 This is a flowchart of step S3 of the present invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figures 1-2 A hierarchical lightning early warning method based on generative networks includes: S1. Acquire multi-source real-time monitoring data, at least three types of short-term forecast products, and geographic information for the area to be warned. Preprocess the data to obtain historical and real-time spatiotemporal sample sequences. S2. Set up three concentric circles (outer, middle, and inner circles). Construct atomic propositions containing multi-source real-time monitoring data and / or short-term forecast products and their threshold parameters. Logically combine these atomic propositions to obtain the triggering and deactivation logic for each circle, forming a set of strategy templates and / or a strategy space for generating candidate strategies. S3. Under the premise of satisfying the concentric circle constraints and logical nesting constraints, use the radius of the three concentric circles, the configurable threshold parameters of the atomic propositions, and the selection and logical combination method of the atomic propositions as candidate parameters. Generate candidate strategies according to preset enumeration rules and / or sampling rules. S4. Select a parameter set, calculate the parameter response vector of each group of candidate parameters in the candidate parameter set on the historical spatiotemporal sample sequence, and perform finite-level quantization to form a parameter prototype code set, and establish a mapping relationship between the parameter prototype code and the candidate parameters; S5. Under the constraints of concentric circles and logical nesting, use a trajectory balancing generation flow network to generate candidate strategies on the parameter prototype code set, and use the warning score on the historical spatiotemporal sample sequence as a reward to update the trajectory balancing generation flow network; S6. Perform a final evaluation of the candidate strategies, select the target strategy based on the warning score, and restore the currently effective strategy according to the mapping relationship; S7. Execute the currently effective strategy on the real-time spatiotemporal sample sequence, and output the graded warning information and corresponding graded warning signals for the outer circle, middle circle, and inner circle.

[0021] In this specific embodiment, S1 includes: In the area to be warned, radar data, lightning data, and atmospheric electric field data are deployed and integrated to form a multi-source real-time monitoring dataset. Simultaneously, three short-term forecast products and geographic information of the target area are also integrated. The radar data consists of volume scan reflectivity data, from which two types of radar features—combined reflectivity and echo top height—expressed in a spatial grid are calculated. The lightning data consists of lightning location data, aggregated at a preset time step to form lightning frequency features and the time difference between the most recent lightning strike and the current moment on a spatial grid. The atmospheric electric field data consists of a sequence of ground-based electric field meter stations, and inverse distance-weighted interpolation is used to generate electric field intensity features on a spatial grid, with the interpolation constant exponent being... And the maximum distance from the site participating in the interpolation to the grid point is limited to... The three short-term forecast products are thunderstorm body identification and tracking information product, radar extrapolation-based future reflectivity product, and numerical model-based lightning occurrence probability product. All three output spatial range or spatial grid field corresponding to the forecast lead time and are converted into short-term forecast features on the same spatial grid. The geographic information of the target area includes the boundary vector of the area to be warned, spatial grid definition parameters, and terrain elevation raster. The boundary vector is rasterized into a grid mask layer for subsequent sample extraction by region. Quality control processing was performed on the aforementioned multi-source real-time monitoring data and short-term forecast products. This quality control process included two parts: missing data identification and outlier removal, and consistency checking. Missing data identification was defined as a percentage of valid grid points within a single time step being less than [percentage missing]. The feature at that time step was identified as missing and marked as missing for subsequent model training and real-time service skipping or downweighting. Outlier removal limited the range of values ​​for the radar combined reflectivity. Grid values ​​exceeding the range are marked as missing measurements, and the range for echo top height is limited. Grid values ​​outside the range are marked as missing measurements, and the range of electric field strength values ​​is limited. The system sets the station values ​​that are out of range as missing before performing interpolation. The lightning count feature is limited to non-negative integers and records with negative numbers or reversed timestamps are removed. For consistency checks, if the maximum time difference between the radar volume scan completion time, the lightning location time, and the electric field sampling time within the same time step is greater than 3 minutes, the three types of real-time features of that time step are uniformly marked as missing to avoid cross-time mixing. After quality control is completed, spatiotemporal alignment is performed. Spatial alignment uses the WGS-84 geographic coordinate datum and defines the latitude and longitude range with the bounding rectangle of the area to be warned as the boundary. Within this range, a fixed-resolution spatial grid is constructed with the longitude resolution set to [value missing]. And latitude resolution Radar features, lightning features, electric field features, and three short-term forecast features are all resampled onto this spatial grid. Point data is assigned to the grid using nearest neighbor attribution, and surface data is assigned to the grid using raster coverage. Time alignment uses a fixed time step. Generate a unified timeline and take the data within 3 minutes before and after the center time of each time step as the data source for that time step, and data outside the window will not participate in the alignment of that time step; After spatiotemporal alignment, for each time step on the unified timeline A spatiotemporal sample containing multi-source features at each time step is constructed. The spatiotemporal sample is stored in the form of a multi-channel grid tensor and concatenated by channel. ; in Indicates the first The spatiotemporal sample feature tensor at each time step, Concat This represents the concatenation operator based on the channel dimension. Indicates the first A set of radar feature grid channels for each time step, where each channel is a one-dimensional array consistent with the spatial grid. Indicates the first A set of lightning feature grid channels at each time step. Indicates the first A set of atmospheric electric field characteristic grid channels at each time step. Indicates the first The short-term forecast feature grid channel set is obtained by converting thunderstorm body identification and tracking information products at each time step. Indicates the first The short-term forecast feature grid channel set obtained by converting radar extrapolated future reflectivity products at each time step. Indicates the first The set of short-term forecast feature grid channels obtained by converting the numerical model lightning occurrence probability product at each time step. A time-independent set of geographic information grid channels for a target area, including at least two types of channels: boundary mask and terrain elevation. Arranged in chronological order A spatiotemporal sample sequence is constructed and divided according to time range to obtain a historical spatiotemporal sample sequence and a real-time spatiotemporal sample sequence. The historical spatiotemporal sample sequence covers two consecutive years of historical business data, and the time step and spatial grid parameters are consistent with the alignment rules mentioned above. The real-time spatiotemporal sample sequence covers data for 120 minutes consecutively prior to the current business time and is updated in a streaming manner when a new time step arrives. The earliest time step is discarded to keep the sequence length constant.

[0022] In this specific embodiment, S2 includes: Based on the early warning reference points in the geographic information of the target area, the outer, middle, and inner rings are defined as geographic buffer layers centered on the early warning reference points and rasterized onto the spatial grid described in S1 to form a layer mask. The radius of the outer ring is initialized to... The center circle radius is initialized to The inner radius is initialized to Furthermore, in the subsequent candidate strategy generation, the range of values ​​for the outer radius is fixed at [value missing]. The range of values ​​for the center circle radius is fixed. The range of values ​​for the inner radius is fixed. This satisfies the concentric circle constraint that the outer circle radius is greater than the middle circle radius and the middle circle radius is greater than the inner circle radius; Atomic propositions are defined within the same spatiotemporal grid system, and the spatial determination results of propositions are unified using the binary field of thunderstorm indication. The real-time atomic propositions are constructed from multi-source real-time monitoring data and used to characterize "real-time enhancement." These real-time atomic propositions include three categories: radar enhancement propositions, electric field enhancement propositions, and lightning enhancement propositions. Radar enhancement propositions are satisfied with a combined reflectivity of not less than 40 dBZ and an echo top height of not less than 10 km, and are generated by thresholding point-by-point across the entire grid based on these conditions. Lightning enhancement propositions are determined by a time step... The condition that the number of internal lightning strikes is not less than 3 is met, and a binary field for generating a lightning storm indicator is generated accordingly. The electric field enhancement proposition requires that the absolute value of the electric field strength is not less than 3. To satisfy the conditions and generate a binary field for electric field thunderstorm indication, the "real-world enhancement" criterion is defined as at least two of the above three types of enhancement propositions being satisfied within the same time step; The single-time early warning atomic proposition is constructed from at least three types of short-term forecast products and used to characterize "forecast entry into the circle". Among them, the thunderstorm body identification and tracking information product is rasterized to generate the first forecast thunderstorm indication binary field by rasterizing the thunderstorm body coverage area corresponding to the 30-minute forecast lead time. The radar extrapolated future reflectivity product is thresholded to generate the second forecast thunderstorm indication binary field by thresholding the region where the extrapolated combined reflectivity is not less than 35 dBZ corresponding to the 30-minute forecast lead time. The numerical model lightning occurrence probability product is thresholded to generate the third forecast thunderstorm indication binary field by thresholding the region where the probability is not less than 0.40 corresponding to the 30-minute forecast lead time. The fusion of the three types of short-term forecast products adopts weighted voting fusion to generate a fused forecast thunderstorm indicator binary field corresponding to the single-time warning atomic proposition. The weighted voting fusion is performed as follows: ; in Indicates time step The fused forecast thunderstorm indicator binary field has each grid point taking a value of 0 or 1. This represents an indicator function that outputs 1 if the condition within the parentheses is true, and 0 otherwise. This indicates that the voting weight of the second-value field for the first forecast thunderstorm indicator is 0.34. This indicates that the voting weight of the second forecast thunderstorm indicator binary field is 0.33. This indicates that the voting weight of the second-valued field for the third-forecast thunderstorm indicator is 0.33. Indicates time step The first forecast thunderstorm indicator binary field generated by thunderstorm body identification and tracking information products, Indicates time step The second forecast thunderstorm indication binary field is generated from radar extrapolated future reflectivity products. Indicates time step The third forecast thunderstorm indicator binary field is generated from the lightning occurrence probability product of the numerical model. This represents the weighted voting threshold, which is set to 0.50. Adjacent time-series warning atomic propositions are used to characterize "persistence," and they are generated by performing a sequence of two consecutive time steps on the same spatial grid. Perform logical AND operations point by point to obtain a persistent thunderstorm indicator binary field and require that the field continuously satisfy a length of 2 time steps; After obtaining the thunderstorm indicator binary field corresponding to each atomic proposition, morphological opening and closing operations are sequentially performed on each thunderstorm indicator binary field to eliminate isolated noise and fill voids. The structuring element employs... The square structural element is slid across the mesh point by point to perform erosion and dilation operations; For each morphologically processed binary field of thunderstorm indication, a connected component analysis was performed, and the spatially continuous region was marked as the thunderstorm indication feature region using the 8-neighborhood connectivity criterion. At the same time, thunderstorm indication feature regions with an area of ​​less than 9 grid points were deleted to suppress residual noise. For each sphere and each atomic proposition, find the intersection between the thunderstorm indicator feature region of the atomic proposition and the spatial region corresponding to the mask of the sphere. If the intersection is not empty, determine that the atomic proposition is satisfied in the sphere and record it as 1; otherwise, record it as 0. Based on the binary judgment results of the atomic propositions within the aforementioned layers, triggering and de-activation logic expressions for each layer are constructed, forming a set of strategy templates. The outer layer triggering logic expression is defined as "adjacent time-based warning atomic propositions within the outer layer are satisfied," and the outer layer de-activation logic expression is defined as "a single time-based warning atomic proposition within the outer layer is not satisfied for three consecutive time steps, and the real-world enhancement within the outer layer is not satisfied for three consecutive time steps." The middle layer triggering logic expression is defined as "the outer layer is in a triggered state, and adjacent time-based warning atomic propositions within the middle layer are satisfied, or the real-world enhancement within the middle layer is satisfied," and the middle layer de-activation logic expression is defined as... The logic expression for triggering the inner circle is defined as "the single-time warning atomic proposition within the middle circle is not satisfied for 3 consecutive time steps and the real-time enhancement within the middle circle is not satisfied for 2 consecutive time steps". The logic expression for triggering the inner circle is defined as "the middle circle is in the triggered state and the single-time warning atomic proposition within the inner circle is satisfied and the real-time enhancement within the inner circle is satisfied". The logic expression for releasing the inner circle is defined as "the real-time enhancement within the inner circle is not satisfied for 2 consecutive time steps". The nested logic constraint is solidified so that the output of the middle circle trigger judgment is allowed to be satisfied only when the outer circle is in the triggered state and the output of the inner circle trigger judgment is allowed to be satisfied only when the middle circle is in the triggered state, thereby ensuring the consistency of the business in the hierarchical advancement.

[0023] In this specific embodiment, S3 includes: Satisfying the concentric circle constraint Under the premise of logical nesting constraints, a candidate parameter set is constructed and its response is calculated and prototype quantized. This represents the outer radius, with units of km. This indicates the radius of the center circle, with the unit being km. This represents the inner radius in km, and the candidate radius values ​​are fixed. And only combinations that satisfy strict inequalities are retained; Candidate threshold parameters and persistence parameters are generated by discrete enumeration and correspond one-to-one with the atomic propositions in S2. The combined reflectivity threshold of the radar enhancement proposition is denoted as... And the unit is dBZ and taken The echo peak height threshold of the radar enhancement proposition is denoted as... And the unit is km and takes The threshold for the number of lightning strikes in the lightning enhancement proposition is denoted as... And the unit is and take and Given a time step of S1, the threshold value of the absolute value of the electric field intensity in the electric field enhancement proposition is denoted as . And the unit is and take The threshold for extrapolating future reflectivity products from radar is denoted as... And the unit is dBZ and taken as The threshold for the numerical model lightning occurrence probability product is denoted as... And take The length of consecutive satisfaction of adjacent early warning atomic propositions is denoted as . And the unit is step and take This corresponds to the continuity definition of "logical AND of two consecutive time steps" in S2; The proposition selection mode is fixed as three layers, which are selected from the set of actual atomic propositions and the set of early warning atomic propositions respectively, and encoded into a binary mask string of length 10. The first to sixth bits indicate whether to select the radar enhancement proposition, lightning enhancement proposition, electric field enhancement proposition, thunderstorm body identification and tracking information product proposition, radar extrapolation proposition, and numerical mode probability proposition, respectively. A value of 1 indicates selection and a value of 0 indicates non-selection. The seventh to tenth bits indicate whether to enable adjacent early warning atomic propositions in the outer ring triggering logic, whether to enable actual atomic propositions in the middle ring triggering logic, whether to enable actual atomic propositions in the inner ring triggering logic, and whether to enable continuous non-satisfaction counting in the release logic. The logical structure pattern is fixed at three structural templates, each numbered and encoded as an integer. ,in This indicates that the outer ring triggers the warning atomic propositions in adjacent time intervals, and the release is achieved through a conjunction structure of "continuous failure of warning and continuous failure of actual condition". This indicates that the middle ring triggering adopts a disjunction structure of "adjacent time-warning atomic propositions or actual enhancement" and is subject to the nesting constraints of the outer ring triggering state. This indicates that the inner circle triggering adopts a conjunction structure of "single-time early warning atomic proposition and real-world enhancement" and is subject to the nested constraints of the middle circle triggering state; For each set of candidate parameters in the candidate parameter set, perform time steps on the historical spatiotemporal sample sequence. From 1 to Stepwise execute the same thunderstorm indicator binary field construction process as S2, replacing only the thresholds used with those corresponding to the candidate parameter set. and ,in S1 represents the total number of time steps in the historical spatiotemporal sample sequence, and the input feature tensor for each time step is defined as follows: In each time step, morphological opening and closing operations, connected domain analysis, and spatial intersection determination with the layer mask are completed to obtain the atomic proposition satisfaction marks of the outer, middle, and inner circles and substitute them into the corresponding logical structure template to complete the trigger determination and release determination, thereby forming the trigger time sequence and release time sequence of each layer. The parameter response vector is calculated based on the trigger and deactivation time sequences, forming a parameter response set. The outer, middle, and inner circle progression relationship components are calculated as the time difference between the first trigger of the outer circle and the first trigger of the middle circle, and the time difference between the first trigger of the middle circle and the first trigger of the inner circle, expressed in minutes. The prediction entry into the circle and the actual enhancement sequence correspondence components are calculated as the time difference between the first time the outer circle satisfies the single-time warning atomic proposition and the first time the outer circle satisfies the actual enhancement proposition, expressed in minutes. The synergistic enhancement relationship components of lightning data, radar data, and atmospheric electric field data are calculated by averaging the number of time steps in which the radar enhancement proposition, lightning enhancement proposition, and electric field enhancement proposition are simultaneously satisfied using a 30-minute sliding time window over the entire sequence. The adjacent time continuity component is calculated as the fused prediction thunderstorm indicator binary field. The maximum length of continuous satisfaction within the layer, measured in time steps, and the state-preserving relation components after parameter fine-tuning are obtained by performing fixed-step fine-tuning on the same candidate parameter group and comparing the consistency of trigger judgments, with the fine-tuning step size fixed at [value missing]. Increase , Increase , Add 1, Increase , Increase by 5dBZ After increasing by 0.10 and replaying the historical sequence before and after the fine-tuning to obtain the three-layer trigger state sequence, the proportion of completely consistent time steps is calculated. The product divergence component is calculated using the spatial consistency index of the binary fields of thunderstorm indications in the three types of short-term forecast products at the same time step and characterized by the average value of historical series. The binary fields of the three types of short-term forecast products are defined by S2. Product divergence over time Record as And calculate using the following formula: ; in and This indicates the short-term forecast product number participating in the consensus calculation and whose value belongs to the set. , Indicates the index of a grid point on a spatial grid. Represents the set of spatial grid point indices for the target region. Indicates time step First Short-term forecast products at grid points The thunderstorm indicator is a binary value that takes either 0 or 1. This represents the operation of finding the maximum value point by point. This represents a constant that avoids a denominator of zero and takes... And define the product divergence component as the whole sequence. ; The missing-test robustness component is obtained through deterministic missing-test simulation, where the time step indices in the historical sequence satisfy... The radar signature channel is set as a missing detection and will meet the following requirements. The lightning characteristic channel is set as a missing test and will satisfy The electric field characteristic channel is set as a missing measurement, and after the missing measurement simulation, a three-layer trigger state sequence is generated according to the same trigger judgment and de-judgment process as before the missing measurement simulation. The proportion of the two being completely consistent at each time step is calculated as the missing measurement robustness component. The computational cost component was measured under a fixed hardware environment defined as a single-core Intel Xeon Gold 6230 processor with 64GB of memory and parallel computing disabled. The measurement method involved performing a complete three-layer trigger and de-trigger determination on the historical sequence for each set of candidate parameters, recording the total time, and then dividing by 1 / 2. Get the time taken per unit time step and the unit is ; Each of the above scalar components is quantized at a finite number of levels according to a preset quantization interval, and the quantization level sequence is concatenated according to a fixed field order to form the parameter prototype code. The time difference component, which represents the progression relationship and the sequential correspondence relationship, is used... The minute-level boundary forms a 5-level quantization, and the collaborative enhancement frequency component adopts... The boundary is divided into 5 levels of quantization, and the duration component is used. The boundary is divided into 5 levels of quantization. Both the state-preserving relation component and the missing-measure robust component after parameter fine-tuning are adopted. The boundary is quantified into 5 levels, and the product divergence is weighted using... The boundary is divided into 5 levels of quantization, and the cost component is calculated using... The boundary is divided into 5 levels of quantification, and the proposition selection mode is encoded and the logical structure mode is numbered. As a discrete field, it is directly incorporated into the parameter prototype code; Multiple candidate parameters with the same parameter prototype code are grouped into the same parameter prototype code to form a parameter prototype code set. A mapping relationship from "parameter prototype code to candidate parameter set" is established for subsequent strategy generation. At the same time, for each parameter prototype code, a representative candidate parameter is selected from its corresponding candidate parameter set. The selection rule for the representative candidate parameter is defined as follows: calculate the center point of each quantization interval corresponding to the parameter prototype code, calculate the L1 distance between the unquantized scalar component of the candidate parameter and the center point, and take the smallest distance. When there are ties for the smallest distance, take the combination of the median of each threshold parameter and the radius of the three concentric layers as the representative candidate parameter.

[0024] In this specific embodiment, S4 includes: The parameter prototype code set is used as the discrete action space of the trajectory balancing generation flow network, and the policy generation process is modeled as a three-step decision-making process in a layered order, with the number of decision steps fixed. And corresponding to the outer circle, middle circle, and inner circle in that order; Define "partial strategy" as a state and denote it as... ,in Indicates the current decision step and takes a value from 1 to... This indicates an empty state that does not contain any selection results for any layer. This indicates the state after the outer ring parameter prototype code has been selected. This indicates the state after selecting the outer and middle circle parameter prototype codes. This indicates the termination state after selecting the prototype codes for the outer, middle, and inner circle parameters; "in the first" Step: Select a parameter prototype code, define it as an action, and record it as... ,and Select from the subset of available parameter prototype codes on the outer ring. From satisfaction The middle circle can be selected using a subset of parameter prototype codes. From satisfaction The inner ring can be selected using a subset of parameter prototype codes, where These represent the outer circle radius, middle circle radius, and inner circle radius obtained by mapping the selected parameter prototype code, respectively, with units of km; The nested logical constraints are implemented by a forced execution method, that is, in the evaluation of candidate strategies and subsequent real-time execution, the "middle circle trigger judgment is only allowed to be satisfied when the outer circle is in the trigger state" and "inner circle trigger judgment is only allowed to be satisfied when the middle circle is in the trigger state" are fixed, so that the generation network is only responsible for selecting the parameter prototype code of each circle without destroying the business consistency of hierarchical advancement. The trajectory balancing generation flow network adopts a forward policy network. Output the action distribution and combine it with an action mask to achieve constrained sampling, where Indicates that the parameter is The forward policy network in state Select action The probability is calculated by the input of the forward policy network as a state encoding vector, which is composed of "embedding vectors of selected parameter prototype codes concatenated by time step" and "embedding of the current layer number". The embedding dimension of the parameter prototype code is fixed at 32 and implemented through a learnable embedding table. The embedding dimension of the layer number is fixed at 8. After concatenation, the vectors are input into two fully connected networks to obtain the logarithmic scores of each selectable parameter prototype code. The probability distribution is obtained through softmax. The hidden layer width of the two fully connected networks is fixed at 128 and the activation function is ReLU. Reverse probability of trajectory balancing generation flow network Implemented using deterministic reverse rules, i.e. Removing the prototype code of the last selected parameter uniquely yields And set the probability of this unique reverse transition to 1, thereby avoiding the introduction of reverse network parameters and ensuring that trajectory balance training is reproducible; For each complete trajectory sampled from the forward policy network First, based on the mapping relationship established in S3, the prototype codes of parameters of each circle are mapped to the corresponding representative candidate parameters and combined to obtain a complete strategy. Then, the triggering and de-triggering process of S2 is replayed on the historical spatiotemporal sample sequence to obtain the graded warning results and calculate the warning score as the trajectory reward. The truth value of the historical lightning event is defined as the time step when the "lightning enhancement proposition in the inner circle is satisfied" is the time when the event occurs. The hit judgment is defined as the inner circle first entering the triggering state within 60 minutes before the event occurs and the triggering state lasts for no less than 2 time steps. The missed judgment is defined as the inner circle never entering the triggering state within 60 minutes before the event occurs. The false alarm judgment is defined as the inner circle entering the triggering state but no lightning event occurs within 60 minutes thereafter. The warning advance is defined as the time difference between the time when the event occurs and the time when the inner circle first triggers, in minutes, and negative values ​​are truncated to 0. The early warning score is obtained by weighted summation of hit reward, missed warning penalty, false alarm penalty, and early warning lead time reward, and then converted into a positive reward through exponential transformation to meet the positivity requirement of the generative flow network. The user-input early warning performance preference parameter is composed of accuracy weights. With timeliness weight Constitutes and satisfies Based on this, the weight of the hit reward item is set to... The weight of the underreporting penalty item is set to The weight of the false alarm penalty item is set to The weight of the early warning reward item is set to... To achieve stronger penalties for underreporting; The trajectory balancing generation flow network uses trajectory balancing target pairs for forward policy network parameters. With normalization constant The update is performed, and the mean squared error is used as the training loss. The trajectory balance constraint is written as follows: ; in This represents the global normalization constant of the generating flow network and is a learnable scalar. Indicates the decision step index. Represents the total number of decision steps and takes Indicates the forward policy network in state Select action The probability, This represents the subordinate state given by the reverse rule. Revert to state The probability, Representing the trajectory The corresponding positive reward of the complete policy on historical spatiotemporal sample sequences; use Optimizer and Joint training and learning rate The batch size is fixed at 64 trajectories and the number of training rounds is 20,000. In each round, the in-batch trajectory is sampled from the current forward policy network, and the corresponding reward and trajectory balance residual are calculated. Then, the parameters are updated by backpropagation. After training, the forward policy network outputs the outer circle, middle circle and inner circle parameter prototype codes according to the maximum probability selection method and combines them to form a candidate policy set.

[0025] In this specific embodiment, S5 includes: Let the set of candidate strategies be denoted as... Furthermore, its elements are complete strategies and are composed of outer circle parameter prototype codes, middle circle parameter prototype codes and inner circle parameter prototype codes in a layered order. At the same time, the historical spatiotemporal sample sequence is divided into optimization set and evaluation set according to time sequence. The optimization set covers the first 18 months and is used for the reward update of the trajectory balance generation flow network in S4. The evaluation set covers the last 6 months and is used for the final evaluation in S5 to avoid overfitting to the optimization set. against For each complete strategy, based on the mapping relationship from "parameter prototype code to representative candidate parameter", the outer circle parameter prototype code, middle circle parameter prototype code, and inner circle parameter prototype code are respectively restored to the representative candidate parameters of the corresponding circle, thus obtaining the parameterized expression of the complete strategy and serving as the strategy to be evaluated. The representative candidate parameters include at least the circle radius. and threshold parameters , With duration It includes proposition selection patterns and logical structure patterns; The above-mentioned strategies to be evaluated are replayed in full time on the evaluation set according to the thunderstorm indicator binary field construction, morphological processing, connected component analysis, layer space intersection determination, trigger determination and de-determination process described in S2, and the outer circle state sequence, middle circle state sequence and inner circle state sequence are obtained, and the corresponding graded warning results are obtained. On the evaluation set, the true value of a lightning event is defined as the "time step in which the proposition of lightning enhancement within the inner circle is satisfied" as the time of the event occurrence. The time of the event occurrence is used as the anchor point to count hits, misses, and false alarms. A hit is defined as the inner circle first entering the trigger state within 60 minutes before the event occurs and the trigger state lasts for no less than 2 time steps. A miss is defined as the inner circle never entering the trigger state within 60 minutes before the event occurs. A false alarm is defined as the inner circle entering the trigger state but no lightning event occurs within the following 60 minutes. The warning lead time is defined as the time difference between the time of the event occurrence and the time of the first trigger of the inner circle, expressed in minutes, and negative values ​​are truncated to 0. Based on the above statistical results, calculate the early warning score for each complete strategy group. and underreporting indicators False alarm indicators Compared with the average warning lead time ,in Describe a complete set of policies and Following the approach described in S4, the reward for hits, the penalty for missed reports, the penalty for false alarms, and the reward for advance warning are weighted according to accuracy. With timeliness weight The score definition is obtained by weighting with defined configurable weights and recalculated on the evaluation set, underreporting indicators. Defined as the ratio of the number of missed alarms to the total number of lightning events, the false alarm index. Defined as the ratio of the number of false alarms in the assessment cluster to the number of triggering segments in the inner circle, the average early warning lead time. Defined as the arithmetic mean of the advance warning amounts of all hit events in the evaluation set; Fix the preset scoring constraints as follows and and and And based on this The complete set of strategies that meet the preset scoring constraints is obtained through screening. ,when When empty, relax the constraints to retain only. And within this relaxed set, the optimal selection described later will continue to be performed; exist When not in an empty space, the target strategy is selected based on maximizing the early warning score, and in the case of a tie, the smaller score is used in order of priority. With smaller As a deterministic disambiguation criterion, the selection of the target strategy is written as: ; in Indicate the target strategy, The operator that takes the independent variable that maximizes the objective function. This represents the complete set of strategies that satisfy the preset scoring constraints. Representation Strategy Early warning score on the evaluation set; For the target strategy Based on the mapping relationship described in S3, the outer circle parameter prototype codes, middle circle parameter prototype codes, and inner circle parameter prototype codes are mapped to representative candidate parameters and merged to obtain the current effective strategy. The current effective strategy is stored in a parameter table and contains at least... , , The proposition selection pattern and logical structure pattern are combined with the solidified logical nesting constraints.

[0026] In this specific embodiment, S6 includes: The system uses time steps The real-time spatiotemporal sample sequence is continuously updated, and a corresponding feature tensor is generated at each new time step. ,in This represents the index of the current time step and increments as real-time data progresses. The channel structure is consistent with S1 and includes radar features, lightning features, electric field features, three types of short-term forecast features, and geographic information channels; At each time step Read the radius of the third concentric circle from the currently effective strategy. and threshold parameters With duration On the same unified spatial grid as S2, the thunderstorm indicator binary field corresponding to each atomic proposition is generated sequentially, and morphological opening and closing operations, as well as 8-neighborhood connected component analysis and small region deletion are performed. The morphological structuring element is fixed as follows: The deletion threshold for square and small regions is fixed at an area of ​​less than 9 grid points. Subsequently, the processed thunderstorm indicator feature region is spatially intersected with the outer, middle, and inner ring masks, respectively. The emptiness of the intersection is then converted into a binary judgment result of the atomic proposition in each ring. This binary judgment result is then substituted into the outer ring triggering logic expression, outer ring deactivation logic expression, middle ring triggering logic expression, middle ring deactivation logic expression, and inner ring triggering logic expression in the current effective strategy to perform Boolean operations and update the three-ring state. The outer ring state is denoted as... The state of the middle circle is recorded as The inner circle state is recorded as And forcefully solidify the nested logical constraints so that they only apply when... The center circle trigger determination result is allowed to be set to satisfy only when... Only when the inner circle trigger judgment result is set to satisfied can the consistency of graded advancement be guaranteed and cross-level jumps be avoided. when When the value changes from 0 to 1 or remains at 1, a level-based early warning message is generated. A warning level is generated when the value changes from 0 to 1 or remains at 1. When the value changes from 0 to 1 or remains at 1, an alarm-level graded early warning message is generated, and the graded early warning message is uniformly encapsulated into a structured message containing the early warning level, release time, impact circle, circle radius parameter, trigger basis atomic proposition identifier, and corresponding threshold parameter value. Simultaneously, the current time step is read based on the thunderstorm identification and tracking information product used in the currently effective strategy. Forecast lead time and The spatial polygonal boundary of the corresponding thunderstorm cell is determined, and its geometric centroid is calculated in the geographic coordinate system. The displacement distance of the thunderstorm cell's centroid from 0 min to 30 min is calculated using the WGS-84 ellipsoidal geodesic distance, and the moving velocity is obtained by dividing by 30 min. And the unit is The direction of movement is determined by the azimuth angle of the line connecting the two centroids for display purposes; Based on the early warning reference point and the specific thunderstorm cell... The geodetic distance of the centroid of time was obtained The unit is km, and the outer, middle, and inner circles are respectively defined by the circle radius. Calculate the estimated entry time of the warning reference point And the unit is The calculation is as follows: ; in Indicates at time step The corresponding radius is The expected entry time into the circle is Indicates time step The distance from the centroid of a thunderstorm cell to the warning reference point. Indicates the outer radius or center circle radius or inner radius Indicates time step Single-target movement speed during thunderstorms This operation represents taking the larger of the two values ​​and is used to truncate negative estimated times to 0. Will , The expected entry time field and the corresponding movement direction field are written together into the graded early warning information and displayed on the output terminal; Finally, the attention-level, alert-level, and alarm-level warning information are sent to the alarm device through preset communication interfaces to drive the issuance of attention-level, alert-level, and alarm-level warning signals, respectively. The attention-level warning signal corresponds to a low-frequency audible and visual alert and lasts for 10 seconds; the alert-level warning signal corresponds to a medium-frequency audible and visual alert and lasts for 30 seconds; and the alarm-level warning signal corresponds to a high-frequency audible and visual alert and lasts for 60 seconds, and is repeatedly triggered at 60-second intervals before the inner circle state is released.

[0027] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

[0028] This invention performs quality control and alignment of multi-source real-time monitoring data with at least three types of short-term forecast products on a unified spatiotemporal grid. It constructs triggering and deactivation logics for three concentric circles (outer, middle, and inner circles) using atomic propositions and their logical combinations. This allows warning determination to simultaneously utilize the sequential correspondence between forecast entry and real-time enhancement, the synergistic enhancement relationship between lightning, radar, and electric fields, and the persistence characteristics of adjacent time intervals. This enables graded advancement and stable deactivation during the evolution of thunderstorms from far to near and from weak to strong, reducing false alarms and missed warnings and suppressing warning status fluctuations. Furthermore, it uniformly scores candidate strategies based on historical spatiotemporal sample sequences and weights indicators such as hit rate, missed warnings, false alarms, and lead time according to user preferences. This allows the strategies to achieve a controllable optimization trade-off between accuracy and timeliness, thereby improving overall warning performance and cross-scenario stability.

[0029] In terms of algorithm structure, this invention addresses the problems of "large product differences and difficulty in coordination, large strategy space and difficulty in automatic optimization, and difficulty in balancing timeliness and reliability" by making business-oriented structural improvements: On the one hand, the trigger-deactivation response of candidate parameters on historical samples is extracted into parameter response vectors and quantized at a finite level to form parameter prototype codes and establish mapping relationships. This transforms the high-dimensional continuous and combinatorial strategy search into a controllable discrete prototype selection, significantly reducing search complexity and enhancing strategy interpretability and reusability. On the other hand, under the constraints of concentric circle radius and logical nesting, a trajectory balancing generation flow network is introduced to generate strategies in concentric circle order and uses the early warning score as the trajectory reward for iterative updates. This makes the generation process naturally satisfy the hierarchical advancement logic and can suppress false alarms and missed alarms in a targeted manner. At the same time, factors such as multi-product divergence, lack of test robustness, and computational cost can be further incorporated into the prototype representation and optimization objectives to better adapt to the incompleteness of multi-source data and the real-time computing requirements, further improving the stability of early warning and deployment efficiency.

Claims

1. A hierarchical lightning early warning method based on generative networks, characterized in that, include: S1. Obtain multi-source real-time monitoring data, at least three types of short-term forecast products and geographic information of the area to be warned, and preprocess them to obtain historical spatiotemporal sample sequences and real-time spatiotemporal sample sequences; S2. Set up three concentric circles: outer circle, middle circle, and inner circle. Construct atomic propositions that include multi-source real-time monitoring data and / or short-term forecast products and their threshold parameters. The triggering logic and deactivation logic of each circle are obtained by logical combination of atomic propositions, forming a set of strategy templates and / or strategy space for generating candidate strategies. The outer, middle, and inner circles are buffer zones generated in a geographic coordinate system with reference to the boundary of the target area and / or the warning reference points within the area to be warned. The radii of the outer, middle, and inner circles are the horizontal distances in the geographic coordinate system. S3. Under the premise of satisfying the concentric circle constraints and logical nesting constraints, the candidate parameters are the three concentric circle radii, the configurable threshold parameters of the atomic propositions, and the selection and logical combination of the atomic propositions. A set of candidate parameters is generated according to the preset enumeration rules and / or sampling rules. The parameter response vectors of each group of candidate parameters in the candidate parameter set are calculated on the historical spatiotemporal sample sequence and quantized at a finite number of levels to form a parameter prototype code set. The mapping relationship between the parameter prototype code and the candidate parameters is established. S4. Under the constraints of concentric circles and logical nesting, a trajectory balancing generation flow network is adopted to generate candidate strategies on the parameter prototype code set, and the warning score on the historical spatiotemporal sample sequence is used as a reward to update the trajectory balancing generation flow network. S5. Conduct a final evaluation of the candidate strategies, select the target strategy based on the warning score, and restore the currently effective strategy based on the mapping relationship; S6. Execute the currently effective strategy on the real-time spatiotemporal sample sequence and output the hierarchical early warning information and corresponding hierarchical early warning signals for the outer circle, middle circle, and inner circle.

2. The hierarchical lightning early warning method based on generative networks according to claim 1, characterized in that, S1 includes: Collect radar data, lightning data, and atmospheric electric field data covering the target area as multi-source real-time monitoring data; collect short-term forecast products from at least three different sources or with different mechanisms; and collect geographic information of the target area. The short-term forecast products include thunderstorm body identification and tracking information; The multi-source real-time monitoring data and short-term forecast products are subjected to quality control processes including missing data identification, outlier removal, and consistency checks. The data processed through quality control is converted to a spatial coordinate reference consistent with the geographic information of the target area, and time synchronization is performed according to a preset time step. The spatial location is mapped to a preset spatial grid of the target area, and the spatiotemporal alignment of the multi-source real-time monitoring data, short-term forecast products and geographic information of the target area is completed. Based on the spatiotemporally aligned data, a spatiotemporal sample sequence containing multi-source features at each time step is constructed in chronological order. Historical spatiotemporal sample sequences and real-time spatiotemporal sample sequences are obtained according to the historical sample time range and the real-time business time range, respectively.

3. The hierarchical lightning early warning method based on generative networks according to claim 2, characterized in that, S2 include: The initial value range of the radius of the outer circle, the radius of the middle circle, and the radius of the inner circle are determined based on the historical spatiotemporal sample sequence, and the radius of the outer circle, the radius of the middle circle, and the radius of the inner circle are initialized within the initial value range, and the radius of the outer circle is greater than the radius of the middle circle, and the radius of the middle circle is greater than the radius of the inner circle. Based on historical spatiotemporal sample sequences, we define real-time atomic propositions, single-time early warning atomic propositions, and adjacent-time early warning atomic propositions. Real-time atomic propositions characterize whether multi-source real-time monitoring data in historical spatiotemporal sample sequences meet real-time conditions. Single-time early warning atomic propositions characterize whether short-term forecast products in historical spatiotemporal sample sequences meet early warning conditions in a single time period. Adjacent-time early warning atomic propositions characterize whether short-term forecast products in historical spatiotemporal sample sequences meet persistence conditions in adjacent time periods. Each atomic proposition includes at least one configurable threshold parameter. The real-time conditions, early warning conditions, or persistence conditions characterized by each atomic proposition, along with the corresponding configurable threshold parameter, are used as the atomic proposition constraints. The single-time early warning atomic proposition and the adjacent-time early warning atomic proposition include early warning conditions set for the at least three types of short-term forecast products respectively. The early warning conditions of the at least three types of short-term forecast products are fused through at least one fusion method among logical OR operation, logical AND operation, and weighted voting fusion method to determine whether the single-time early warning atomic proposition and the adjacent-time early warning atomic proposition are satisfied. For the outer ring, middle ring and inner ring respectively, at least one atomic proposition is selected from the real-time atomic proposition, single-time warning atomic proposition and adjacent-time warning atomic proposition, and corresponding triggering logic expression and deactivation logic expression are constructed through logical operations to form the strategy template set and / or strategy space; The trigger determination for the outer, middle and inner circles includes: based on spatiotemporally aligned multi-source real-time monitoring data and / or short-term forecast products, generating a corresponding thunderstorm indication binary field for each of the atomic propositions, wherein the thunderstorm indication binary field is used to characterize the spatial region within the target area that satisfies the atomic proposition constraint conditions; When the atomic proposition is an adjacent time-based warning atomic proposition, firstly, based on the short-term forecast product, generate single-time thunderstorm indication binary fields for at least two adjacent time periods, and perform logical AND operations on the single-time thunderstorm indication binary fields point by point according to spatial location on the same spatial grid, and / or accumulate the single-time thunderstorm indication binary fields point by point within a preset time window and compare them with the persistence threshold to obtain the thunderstorm indication binary field corresponding to the adjacent time-based warning atomic proposition; Morphological processing is performed on each of the aforementioned thunderstorm indicator binary fields to eliminate isolated noise and fill voids; For each of the thunderstorm indicator binary fields after morphological processing, a connected component analysis is performed to mark the spatially continuous regions as the thunderstorm indicator feature regions corresponding to the corresponding atomic propositions; For each atomic proposition, if the corresponding thunderstorm indicator feature region has a non-empty intersection with the spatial region of the corresponding layer, the atomic proposition is determined to be satisfied, and the determination result of the atomic proposition is recorded as 1; otherwise, it is recorded as 0. Substitute the judgment results of each of the atomic propositions into the trigger logic expression of the corresponding layer in the strategy template set and / or strategy space and perform Boolean operation to obtain the comprehensive logic operation result; When the result of the comprehensive logic operation is 1, it is determined that the corresponding layer meets the triggering condition.

4. The hierarchical lightning early warning method based on generative networks according to claim 1, characterized in that, S3 include: For each set of candidate parameters generated under the premise of satisfying the concentric circle constraints and logical nesting constraints, which consists of the radius of each circle, the configurable threshold parameters of the atomic propositions, and the selection and logical combination of the atomic propositions, trigger judgment and release judgment are performed on the outer circle, middle circle and inner circle respectively according to the historical spatiotemporal sample sequence to obtain the trigger time sequence and release time sequence corresponding to each circle. The parameter response vector is calculated based on the trigger time sequence and the release time sequence. The advancing relationship of the outer circle, middle circle and inner circle is characterized by the order of the trigger times of the outer circle, the middle circle and the inner circle and the time difference between adjacent trigger times. The sequential correspondence between forecast entry into the circle and actual enhancement is characterized by the time difference between the time when the short-term forecast product meets the entry condition and the time when the multi-source actual monitoring data meets the actual enhancement condition. The synergistic enhancement relationship between lightning data, radar data and atmospheric electric field data is characterized by the number of times the lightning data, radar data and atmospheric electric field data simultaneously meet the corresponding atomic proposition within a preset time window. The continuity relationship between adjacent time periods is characterized by the length of time when adjacent time periods continuously meet the corresponding atomic proposition. The state maintenance relationship after parameter fine-tuning is characterized by the consistency of the trigger judgment results before and after parameter fine-tuning of the candidate parameters with a preset step size. The parameter response vector also includes a proposition selection mode component and a logic structure mode component. The proposition selection mode component is encoded by the identifier set of the atomic propositions selected in the candidate parameters, and the logic structure mode component is encoded by the number of the logic expression structure template corresponding to the logic combination method in the candidate parameters, thus forming a parameter response set. Each scalar component of the parameter response vector in the parameter response set is subjected to scalar quantization processing with a finite number of quantization levels according to a preset quantization interval. The quantized scalar components are combined to form a parameter prototype code. Multiple candidate parameters with the same parameter prototype code are merged into the same parameter prototype code and form a parameter prototype code set. At the same time, a mapping relationship between the parameter prototype code and the candidate parameter set is established. For each parameter prototype code, a representative candidate parameter is selected from its corresponding candidate parameter set according to a preset representative rule. The preset representative rule includes: selecting the candidate parameter whose distance between the parameter response vector and the center of the quantization interval corresponding to the parameter prototype code is the smallest, and / or selecting the candidate parameter whose value is the median of each parameter.

5. The hierarchical lightning early warning method based on generative networks according to claim 1, characterized in that, S4 includes: Based on the parameter prototype code set, the parameter prototype code is determined as the parameter selection unit of the trajectory balancing generation flow network. The trajectory balancing generation flow network selects the parameter prototype codes of the corresponding circles layer by layer in the order of outer circle, middle circle and inner circle, and then combines them to obtain each complete strategy in the candidate strategy set. Each parameter prototype code at least represents the radius parameter of the corresponding layer, the selection result of the atomic proposition, the logical combination method, and the threshold parameter; The concentric constraints include the outer circle's radius being greater than the middle circle's radius and the middle circle's radius being greater than the inner circle's radius. The nested logical constraints include the middle circle's triggering logic expression being allowed to be satisfied only when the outer circle is in a triggered state, and the inner circle's triggering logic expression being allowed to be satisfied only when the middle circle is in a triggered state. For each complete strategy, after mapping the parameter prototype code to representative candidate parameters according to the mapping relationship, triggering and de-triggering judgments are performed on the outer circle, middle circle and inner circle respectively according to the historical spatiotemporal sample sequence to obtain graded warning results. The warning score is calculated based on the graded warning results and used as the trajectory reward of the trajectory balancing generation flow network so that the trajectory balancing generation flow network outputs the candidate strategy set. The early warning score includes at least two of the following: hit reward, missed report penalty, false alarm penalty, and early warning lead time reward, and uses configurable weights to weight and combine the included items. The system receives warning performance preference parameters from the target user, which include accuracy weight and timeliness weight, and determines the configurable weight based on the accuracy weight and timeliness weight.

6. The hierarchical lightning early warning method based on generative networks according to claim 1, characterized in that, S5 include: For each complete policy in the candidate policy set, triggering and de-triggering decisions are performed in the outer, middle, and inner circles respectively based on historical spatiotemporal sample sequences to obtain corresponding graded early warning results; Based on the tiered early warning results, the early warning score for each complete strategy is calculated, and complete strategies whose early warning scores satisfy the preset score constraints are selected according to the preset score constraints. When there are multiple sets of complete strategies whose warning scores meet the preset scoring constraints, the set of complete strategies with the highest warning scores is selected as the target strategy. For the target strategy, based on the mapping relationship, the parameter prototype code in the target strategy is mapped to its corresponding representative candidate parameters. The representative candidate parameters include the radius values ​​of the outer circle, middle circle and inner circle, the configurable threshold parameter values ​​of the atomic propositions, and the selection results and logical combination methods of the atomic propositions, thereby parameterizing the target strategy into the currently effective strategy. The preset scoring constraints include at least one of the following: the early warning score is not less than a preset scoring threshold, the missed reporting index is not greater than a preset missed reporting threshold, the false alarm index is not greater than a preset false alarm threshold, and the early warning lead time is not less than a preset lead time threshold.

7. The hierarchical lightning early warning method based on generative networks according to claim 1, characterized in that, S6 include: Based on the real-time spatiotemporal sample sequence, extract the real-time features corresponding to the outer circle, middle circle and inner circle, and according to the trigger logic expression and deactivation logic expression of the outer circle, the trigger logic expression and deactivation logic expression of the middle circle and the trigger logic expression and deactivation logic expression of the inner circle in the current effective strategy, execute the trigger judgment and deactivation judgment respectively for the outer circle, middle circle and inner circle to update the outer circle state, middle circle state and inner circle state. When the outer ring is triggered, it outputs a warning message at the attention level; when the middle ring is triggered, it outputs a warning message at the alert level; and when the inner ring is triggered, it outputs a warning message at the alarm level. Based on the attention-level graded early warning information, the alert-level graded early warning information, and the alarm-level graded early warning information, the alarm device is driven to issue attention-level graded early warning signals, alert-level graded early warning signals, and alarm-level graded early warning signals, respectively.

8. The hierarchical lightning early warning method based on generative networks according to claim 4, characterized in that, The parameter response vector also includes a product divergence component, which is characterized by the spatial consistency index among the thunderstorm indication binary fields of the at least three types of short-term forecast products at the same time. The product divergence component is then involved in the scalar quantization process of the finite quantization series so that the parameter prototype code represents the multi-product consistency case and the multi-product divergence case.

9. The hierarchical lightning early warning method based on generative networks according to claim 4, characterized in that, The parameter response vector also includes a missing detection robustness component and a computational cost component. The missing detection robustness component is characterized by the consistency between the trigger judgment result after performing a preset missing detection simulation on the multi-source real-time monitoring data and the trigger judgment result before the missing detection simulation. The computational cost component is characterized by the unit time step consumption of performing the trigger judgment and de-judgment judgment in a preset hardware environment. The missing detection robustness component and the computational cost component are then involved in the scalar quantization processing of the finite quantization level.

10. The hierarchical lightning early warning method based on generative networks according to claim 7, characterized in that, Also includes: Based on the positional changes of thunderstorm cells in the short-term forecast products used in the current effective strategy at different forecast lead times, the speed and direction of thunderstorm movement are estimated. The estimated entry time of the warning reference point is calculated based on the radius of the outer circle, the radius of the middle circle, the radius of the inner circle, and the speed of thunderstorm movement. The estimated entry time is output together with the attention-level warning information, the alert-level warning information, and the alarm-level warning information.

Images