Artificial intelligence storm prediction model construction method based on multi-source heterogeneous data fusion

Abstract

This invention discloses a method for constructing an artificial intelligence-based rainstorm prediction model based on multi-source heterogeneous data fusion, relating to the field of rainstorm prediction technology. It involves collecting multi-source meteorological data of the area to be predicted, including satellite remote sensing data, weather radar data, ground meteorological station data, and numerical weather prediction model data. Based on preset dominant factor discrimination rules, the area to be predicted is divided into different dominant type zones, and virtual cloud entities are constructed based on these zones. By establishing a motion model of the virtual cloud entities, their life trajectory is generated, thereby constructing a cloud motion map reflecting the dynamic evolution of the virtual cloud entities. The system generates intervention commands based on the cloud motion map analysis, projects these commands into the cloud motion map, and finally outputs a rainstorm prediction report containing deterministic warnings and potential risk assessments. This invention constructs a virtual cloud entity system with dynamic perception and proactive intervention capabilities, achieving accurate prediction of the rainstorm formation process.

Description

Technical Field

[0001] This invention relates to the field of rainstorm prediction technology, specifically to a method for constructing an artificial intelligence rainstorm prediction model based on the fusion of multi-source heterogeneous data. Background Technology

[0002] The current mainstream rainstorm prediction models mainly adopt numerical weather prediction models and extrapolation methods. Numerical weather prediction models use supercomputers to solve atmospheric dynamics and thermodynamic equations for prediction, while extrapolation methods use image recognition and trajectory tracking technology based on radar, satellite and other observation data to predict the movement and development of rainfall.

[0003] However, existing technologies have limitations. In terms of multi-source data fusion, existing technologies mostly adopt static integration methods, failing to fully consider the spatiotemporal matching and physical consistency constraints between different types of data. This results in the complementary advantages between data not being fully utilized, reducing the accuracy of rainstorm forecasts. At the same time, in terms of forecasting mechanisms, existing systems are mainly based on fixed algorithm processes, lacking active verification and exploration of key uncertainties in the forecasting process. They cannot effectively identify and correct uncertainties in the forecasting process. This limitation makes it difficult for existing models to guarantee the accuracy and timeliness of forecasts when dealing with sudden and localized severe convective weather.

[0004] Therefore, it is of great significance to develop an artificial intelligence rainstorm prediction model based on the fusion of multi-source heterogeneous data. Summary of the Invention

[0005] The purpose of this invention is to provide a method for constructing an artificial intelligence-based rainstorm prediction model based on the fusion of multi-source heterogeneous data, so as to solve the problems in the background technology.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for constructing an artificial intelligence-based rainstorm prediction model based on multi-source heterogeneous data fusion, comprising:

[0007] Collect multi-source meteorological data for the area to be predicted, including satellite remote sensing data, weather radar data, ground meteorological station data, and numerical weather prediction model data;

[0008] Based on the preset dominant factor discrimination rules, the region to be predicted is divided into different dominant type regions, and virtual cloud entities are constructed based on the dominant type regions;

[0009] Construct a motion model of a virtual cloud entity, and generate the life development trajectory of the virtual cloud entity within a preset time based on the motion model;

[0010] Construct a cloud movement map of the region to be predicted based on virtual cloud entities and life development trajectories;

[0011] Based on cloud motion maps, generate changes in multi-source meteorological data for the area to be predicted, and generate intervention instructions based on these changes.

[0012] Intervention commands are projected into cloud motion maps to generate rainstorm forecast reports.

[0013] In a preferred embodiment, the step of collecting multi-source meteorological data of the area to be predicted, including satellite remote sensing data, weather radar data, ground meteorological station data, and numerical weather prediction model data, is as follows:

[0014] Multi-source meteorological data of the area to be predicted are collected at a preset period, and spatiotemporal labels are marked for each type of data. The multi-source meteorological data are uniformly remapped to a preset uniform spatial resolution grid for spatiotemporal alignment.

[0015] Among them, satellite remote sensing data includes infrared brightness temperature, water vapor channel brightness temperature and visible light albedo of the area to be predicted;

[0016] Weather radar data includes the combined reflectivity, basic radial velocity, and vertically integrated liquid water content of the area to be predicted;

[0017] Ground meteorological station data includes temperature, humidity, air pressure, wind speed, wind direction, and minute-level precipitation for the area to be predicted;

[0018] Numerical weather prediction model data includes the convective available potential energy field, water vapor flux field, and vertical velocity field of the region to be predicted.

[0019] In a preferred embodiment, the step of dividing the region to be predicted into different dominant type regions based on a preset dominant factor discrimination rule, and constructing a virtual cloud entity based on the dominant type regions, is as follows:

[0020] A uniform spatial resolution grid is obtained, and multi-source meteorological data within each grid area are used as the data to be analyzed.

[0021] Dominant factors are generated based on the dominant factor discrimination rules for the data to be analyzed;

[0022] Taking the dominant factor as the root system, the non-dominant multi-source meteorological data is absorbed into the surrounding grid range in a root-like manner based on the principle of physical consistency, and the boundary of the virtual cloud entity is dynamically generated.

[0023] A preset volume threshold for virtual cloud entities is set. If the volume of a virtual cloud entity exceeds the volume threshold, root-based absorption will stop.

[0024] In a preferred embodiment, the step of generating dominant factors from the data to be analyzed based on the dominant factor discrimination rule is as follows:

[0025] The dominant factors include cloud top development, precipitation particle dominance, energy field dominance, and water vapor transport dominance.

[0026] When the infrared brightness temperature of satellite remote sensing data within the grid area is lower than the first preset threshold and its local gradient exceeds the second preset threshold, the dominant factor within the grid area is determined to be the dominant factor for cloud top development.

[0027] When the combined reflectivity of weather radar data within the grid area is higher than the third preset threshold and its spatial continuity meets the fourth preset condition, the dominant factor within the grid area is determined to be the precipitation particle dominant factor.

[0028] When the convective effective potential energy of the numerical weather prediction model data within the grid range is higher than the fifth preset threshold, the dominant factor within that grid range is determined to be the dominant factor of the energy field.

[0029] When the lower-level water vapor flux of the numerical weather prediction model data within the grid range is higher than the sixth preset threshold, the dominant factor within that grid range is determined to be the dominant factor for water vapor transport.

[0030] If multiple types of dominant factors exist within the same grid range, then the grid range is taken as a candidate region, and the multiple types of dominant factors within the candidate region are taken as candidate dominant factors.

[0031] For each candidate factor in the candidate region, the strength of evidence is measured. The strength of evidence includes the magnitude of exceeding a preset threshold, the spatiotemporal variation trend of the physical parameters of the candidate factor, and the uniformity of its spatial distribution.

[0032] Simultaneously, the physical consistency between candidate factors and non-dominant multi-source meteorological data within the candidate region is evaluated, where physical consistency refers to whether the non-dominant multi-source meteorological data supports the meteorological conditions represented by the candidate factors.

[0033] Based on the evaluation results of the evidence strength and physical consistency of each candidate factor, a pre-defined classifier outputs the unique dominant factor of the candidate region.

[0034] In a preferred embodiment, the step of constructing a motion model of the virtual cloud entity and generating a life development trajectory of the virtual cloud entity within a preset time based on the motion model is as follows:

[0035] Collect the spatial location sequence of each virtual cloud entity within a continuous preset period of history;

[0036] The optical flow method is used to calculate the movement vector of each virtual cloud entity, which includes the movement speed and direction;

[0037] By combining environmental wind field data from numerical weather prediction models to correct moving vectors, a motion model of virtual cloud entities is generated.

[0038] The motion path within a preset period is predicted using extrapolation based on the motion model as the trajectory of life development.

[0039] In a preferred embodiment, the step of constructing a cloud motion map of the region to be predicted based on virtual cloud entities and life development trajectories is as follows:

[0040] The spatial distribution of the virtual cloud entities at the current moment is superimposed to generate a spatial layer;

[0041] The life development trajectory of each virtual cloud entity is superimposed to generate a trajectory layer, and different colors are used to distinguish the movement trajectory and life development trajectory of the virtual cloud entity.

[0042] The spatial layer and trajectory layer are registered and fused to generate a cloud motion map.

[0043] In a preferred embodiment, the step of generating changes in multi-source meteorological data of the area to be predicted based on cloud motion maps, and generating intervention instructions based on these changes, is as follows:

[0044] Multiple standard rainstorm models are predefined, and each standard rainstorm model is composed of a specific combination of multiple virtual cloud entities in space and time.

[0045] Real-time monitoring of cloud movement maps, and chain matching of the current and predicted states of virtual cloud entities with life development trajectories with standard rainstorm patterns;

[0046] When the match degree with any standard rainstorm pattern exceeds the match threshold, the matching pattern and the corresponding confidence level are output.

[0047] Intervention instructions are generated based on matching patterns and corresponding confidence levels. These intervention instructions include pattern catalysis instructions and potential exploration instructions.

[0048] A pre-set confidence range is used to generate a model catalysis instruction when the confidence of the matching pattern is in the medium range. The model catalysis instruction is a test cloud entity used to verify the sensitivity of the region to develop into a full-scale rainstorm standard pattern.

[0049] If a certain area does not match the standard rainstorm pattern, a potential exploration command is generated. The potential exploration command is a probabilistic cloud cluster used to simulate the probability that the area will develop into the standard rainstorm pattern under random perturbation.

[0050] In a preferred embodiment, the step of projecting intervention commands into the cloud motion map to generate a rainstorm forecast report is as follows:

[0051] The generated intervention commands are projected onto the target area of ​​the cloud motion map;

[0052] For a certain region, in response to the pattern catalysis command, a test cloud entity is injected to perform virtual cloud entity movement simulation in that region. Based on whether a complete rainstorm standard pattern is catalyzed, the environmental sensitivity coefficient of that region is output.

[0053] For a certain area, in response to the potential exploration command, a probability cloud is injected to analyze the combination of virtual cloud entities in that area, and the combination and proportion of the standard rainstorm pattern are output as the probability of potential rainstorm.

[0054] The cloud cluster motion map is updated based on the high-confidence matching pattern, environmental sensitivity coefficient, and latent rainstorm probability, and a rainstorm forecast report is generated.

[0055] The rainstorm forecast report includes a list of high-risk rainstorm standard models, the distribution of environmental sensitivity coefficients for each region, and a probability distribution map of potential rainstorms.

[0056] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0057] 1. This invention constructs a virtual cloud entity system with dynamic interactive capabilities, enabling refined simulation and forward-looking prediction of the rainstorm formation process. Compared with traditional static analysis models, this scheme introduces an intervention command mechanism, which can actively detect key uncertainties in the meteorological system. When the system identifies a preliminary standard rainstorm pattern but lacks sufficient confidence, it automatically injects optimized test cloud entities for catalytic verification. This dynamic detection method is equivalent to conducting a controllable meteorological experiment in virtual space. By observing the interaction between the test cloud and existing cloud, the system can accurately assess the sensitivity and probability of the region developing into a rainstorm. This active detection mechanism significantly improves the accuracy and reliability of short-term rainstorm prediction, enabling forecasters to identify potential risks earlier and make accurate judgments.

[0058] 2. This invention constructs a more complete knowledge system for rainstorm identification through deep fusion and intelligent analysis of multi-source heterogeneous data. It adopts a root-based data absorption method, radiating outwards from the dominant factor as the core, to achieve organic integration of multi-source meteorological data. This data fusion method not only retains the characteristic advantages of each data source, but also ensures the scientific nature of data fusion through the principle of physical consistency. In the process of constructing virtual cloud entities, the system fully considers the complementarity and synergy between different meteorological types, forming a cloud movement map with clear physical meaning. The multi-level and multi-dimensional data fusion mechanism enables the model to have a more comprehensive understanding of the evolution law of rainstorms, providing a solid data foundation for accurate prediction, while enhancing the model's adaptability and robustness under complex meteorological conditions. Attached Figure Description

[0059] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0060] Figure 1 This is a flowchart of the method of the present invention.

[0061] Figure 2 This is a logic block diagram of the present invention. Detailed Implementation

[0062] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0063] Example 1, please refer to Figure 1 As shown in this embodiment, the method for constructing an artificial intelligence-based rainstorm prediction model based on multi-source heterogeneous data fusion includes:

[0064] S1. Collect multi-source meteorological data for the area to be predicted, including satellite remote sensing data, weather radar data, ground meteorological station data, and numerical weather prediction model data;

[0065] S2. Based on the preset dominant factor discrimination rules, the region to be predicted is divided into different dominant type regions, and virtual cloud entities are constructed based on the dominant type regions;

[0066] S3. Construct a motion model of the virtual cloud entity, and generate the life development trajectory of the virtual cloud entity within a preset time based on the motion model;

[0067] S4. Construct a cloud movement map of the region to be predicted based on virtual cloud entities and life development trajectories;

[0068] S5. Generate changes in multi-source meteorological data for the area to be predicted based on cloud motion maps, and generate intervention instructions based on the changes.

[0069] S6. Project the intervention command into the cloud movement map to generate a rainstorm forecast report;

[0070] As described in steps S1-S6 above, the current mainstream rainstorm prediction models mainly adopt numerical weather prediction models and extrapolation methods. Numerical weather prediction models use supercomputers to solve atmospheric dynamics and thermodynamic equations for prediction, while extrapolation methods use image recognition and trajectory tracking technology based on radar, satellite and other observation data to predict the movement and development of rainfall.

[0071] However, existing technologies have limitations. In terms of multi-source data fusion, existing technologies mostly adopt static integration methods, failing to fully consider the spatiotemporal matching and physical consistency constraints between different types of data. This results in the complementary advantages between data not being fully utilized, reducing the accuracy of rainstorm forecasts. At the same time, in terms of forecasting mechanisms, existing systems are mainly based on fixed algorithm processes, lacking active verification and exploration of key uncertainties in the forecasting process. They cannot effectively identify and correct uncertainties in the forecasting process. This limitation makes it difficult for existing models to guarantee the accuracy and timeliness of forecasts when dealing with sudden and localized severe convective weather.

[0072] This invention constructs a virtual cloud entity system with dynamic interactive capabilities, enabling refined simulation and forward-looking prediction of the rainstorm formation process. Compared with traditional static analysis models, this scheme introduces an intervention command mechanism, which can actively detect key uncertainties in the meteorological system. When the system identifies a preliminary standard rainstorm pattern but lacks sufficient confidence, it automatically injects optimized test cloud entities for catalytic verification. This dynamic detection method is equivalent to conducting a controlled meteorological experiment in virtual space. By observing the interaction between the test cloud and existing cloud, the system can accurately assess the sensitivity and probability of the region developing into a rainstorm. This active detection mechanism significantly improves the accuracy and reliability of short-term rainstorm prediction, enabling forecasters to identify potential risks earlier and make accurate judgments.

[0073] Through deep fusion and intelligent analysis of multi-source heterogeneous data, a more complete knowledge system for rainstorm identification has been constructed. A root-based data absorption method is adopted, radiating outwards from the dominant factor as the core, achieving organic integration of multi-source meteorological data. This data fusion approach not only preserves the characteristic advantages of each data source but also ensures the scientific nature of the data fusion through the principle of physical consistency. In the construction of virtual cloud entities, the system fully considers the complementarity and synergy between different meteorological types, forming cloud movement maps with clear physical meaning. This multi-level, multi-dimensional data fusion mechanism enables the model to more comprehensively understand the evolution of rainstorms, providing a solid data foundation for accurate prediction, while also enhancing the model's adaptability and robustness under complex meteorological conditions.

[0074] In one embodiment, step S1, which involves collecting multi-source meteorological data of the area to be predicted, including satellite remote sensing data, weather radar data, ground meteorological station data, and numerical weather prediction model data, includes:

[0075] S11. Collect multi-source meteorological data of the area to be predicted at a preset period, and label each type of data with a spatiotemporal label. Then, remap the multi-source meteorological data to a preset uniform spatial resolution grid for spatiotemporal alignment.

[0076] S12. Among them, satellite remote sensing data includes infrared brightness temperature, water vapor channel brightness temperature and visible light albedo of the area to be predicted;

[0077] S13. Weather radar data includes the combined reflectivity, basic radial velocity, and vertical integrated liquid water content of the area to be predicted.

[0078] S14. Ground meteorological station data includes temperature, humidity, air pressure, wind speed, wind direction, and minute-level precipitation for the area to be predicted.

[0079] S15. Numerical prediction model data includes the convective available potential energy field, water vapor flux field, and vertical velocity field of the region to be predicted.

[0080] As described in steps S11-S15 above, multi-source meteorological data is periodically collected by connecting to the system through a standardized data interface. The system acquires multi-band observation data from meteorological satellites in real time through satellite data receiving stations, collects basic data-level scanning data by connecting to weather radar networks, acquires minute-level ground observation data by connecting to meteorological station networks, and obtains high-resolution model output from numerical weather prediction centers. After parsing and verification, all types of data are uniformly stored in the meteorological data center. For the multi-source meteorological data center, a time stamp accurate to the second and geographic coordinate information are added to each data unit through a spatiotemporal label generation unit. In the data remapping process, a bilinear interpolation algorithm is used to downscale satellite remote sensing data and numerical model data to a uniform one-kilometer grid resolution. At the same time, Kriging spatial interpolation technology is used to upscale discretely distributed ground station observation data to a uniform spatial resolution grid. Furthermore, Upon receiving multi-source meteorological data, optimization measures are implemented to improve the accuracy of rainstorm forecasts. The steps include: identifying and removing anomalous radar echoes using a peak detection algorithm; eliminating abnormal values ​​in station observations using a threshold filtering method based on statistical distribution characteristics; and ultimately generating a multi-dimensional meteorological data cube with a completely unified spatiotemporal reference. Satellite data processing employs an atmospheric correction model to eliminate interference from surface reflection and atmospheric radiation, and an optimized cloud detection algorithm accurately extracts the physical parameters of effective cloud areas. Radar data processing utilizes a velocity defuzzification algorithm to correct the radial velocity field. Ground station data undergoes an adaptive quality control process to identify and correct observation jumps, and a moving average method based on time series analysis is used to smooth high-frequency fluctuation signals. Numerical model data is fused with measured observation data using three-dimensional variational assimilation technology, and an accurate thermodynamic parameter field is calculated through a comprehensive physical parameterization scheme.

[0081] In one embodiment, step S2, which divides the region to be predicted into different dominant type regions based on a preset dominant factor discrimination rule and constructs a virtual cloud entity based on the dominant type regions, includes:

[0082] S21. Obtain a uniform spatial resolution grid and acquire multi-source meteorological data within each grid area as data to be analyzed;

[0083] S22. Based on the dominant factor discrimination rule, the dominant factors are generated from the data to be analyzed.

[0084] S23. Taking the dominant factor as the root system, and based on the principle of physical consistency, the non-dominant multi-source meteorological data is absorbed into the surrounding grid range in a root-like manner to dynamically generate the boundary of the virtual cloud entity.

[0085] S24. Preset the volume threshold of the virtual cloud entity. If the volume of the virtual cloud entity exceeds the volume threshold, root absorption will stop.

[0086] As described in steps S21-S24 above, during the virtual cloud entity construction stage, the system first reads preprocessed standard grid data. Based on a uniform spatial resolution grid, the area to be predicted is divided into analysis units of grid size. The data to be analyzed in each analysis unit is processed by the feature extraction module to generate a standardized feature vector containing physical parameters. A dominant factor discrimination rule is implemented for the standardized feature vector in each analysis unit, and a multi-threshold joint judgment algorithm is used to generate the dominant factor. Once the dominant factor in the analysis unit is generated, the system begins to absorb multi-source meteorological data from other analysis units, with the dominant factor as the root system. The steps are as follows: [The text abruptly ends here, likely due to an incomplete sentence or a missing section.] The sub-grid serves as the main body of the growing root system. Grids adjacent to the main root system that do not contain dominant factors are considered candidate grids. A physical consistency evaluation model is used to calculate the correlation between the sub-grid and its neighboring grids. This model constructs a three-dimensional evaluation system encompassing spatial gradient, physical field coordination, and evolutionary trend. The spatial gradient module calculates the standardized differences between the main root system and candidate grids in various physical parameter fields, ensuring a smooth transition between the candidate grids and the overall system. The physical field coordination analysis examines the coupling relationships between different parameter fields, verifying whether the candidate grids meet the physical environment configuration required by the main root system. The evolutionary trend characteristics compare the parameter change trajectories of the candidate grids and the main root system, requiring… Its development direction is consistent with the evolution law of the root system. The evaluation results of these three dimensions are fused through an adaptive weight allocation mechanism to finally generate a physical consistency comprehensive score. When the physical consistency comprehensive score exceeds a preset threshold, root-based absorption is triggered to generate a new root system, and the new root system is used as a new starting point to iteratively absorb surrounding candidate units. When the volume of the new root system exceeds a preset volume or there are no candidate grids, root-based absorption stops, and the generated root system is treated as a virtual cloud entity. For example, if the radar combined reflectivity of a certain area analysis unit is as high as 50 dBZ, far exceeding the preset threshold, then the dominant factor of this analysis unit is precipitation particle dominant. The analysis unit serves as the root system entity for constructing the virtual cloud entity. Starting from this root system entity, the system examines the six adjacent grid units. The system first scans the grid to its east and finds that the location has a reflectivity of 42 dBZ, a brightness temperature of -62 degrees Celsius, and a convective effective potential energy of 2000 J / kg. The physical properties of this candidate unit are close to those of the root system entity, i.e., the physical consistency score is high. Therefore, the analysis unit is absorbed into the root system entity to generate a new root system entity boundary. The system continues to probe on the newly formed boundary. The candidate grid on the southeast side triggers a gradient alarm due to a sharp drop in reflectivity to 28 dBZ. In addition, the candidate grid is accompanied by discontinuous thermodynamic field characteristics, which causes the score to drop sharply and it is rejected for absorption.Although the reflectivity of the grid on the north side is only 38 dBZ, its liquid water content gradient is stable and it is located in the environmental wind convergence zone. Therefore, it was included in the main root system area with a good score. After multiple rounds of iterative growth, the system encountered candidate grids with discontinuous physical characteristics in all directions, ultimately forming a virtual cloud entity in three-dimensional space that conforms to meteorological principles.

[0087] In one embodiment, step S22, which generates dominant factors from the data to be analyzed based on the dominant factor discrimination rule, includes:

[0088] S221, the dominant factors include cloud top development dominant factors, precipitation particle dominant factors, energy field dominant factors and water vapor transport dominant factors.

[0089] S222. When the infrared brightness temperature of satellite remote sensing data within the grid area is lower than the first preset threshold and its local gradient exceeds the second preset threshold, the dominant factor within the grid area is determined to be the dominant factor for cloud top development.

[0090] S223. When the combined reflectivity of weather radar data within the grid range is higher than the third preset threshold and its spatial continuity meets the fourth preset condition, the dominant factor within the grid range is determined to be the precipitation particle dominant factor.

[0091] S224. When the convective effective potential energy of the numerical weather prediction model data within the grid range is higher than the fifth preset threshold, the dominant factor of the grid range is determined to be the dominant factor of the energy field.

[0092] S225. When the lower-level water vapor flux of the numerical weather prediction model data within the grid range is higher than the sixth preset threshold, the dominant factor within the grid range is determined to be the dominant factor for water vapor transport.

[0093] S226. If there are multiple dominant factors of the same grid range, then the grid range shall be regarded as a candidate region, and the multiple dominant factors of the candidate region shall be regarded as candidate dominant factors.

[0094] S227. Measure the strength of evidence for each candidate factor in the candidate region. The strength of evidence includes the magnitude of exceeding a preset threshold, the spatiotemporal variation trend of the physical parameters of the candidate factor, and the uniformity of its spatial distribution.

[0095] S228. At the same time, assess the physical consistency between candidate factors and non-dominant multi-source meteorological data within the candidate region, where physical consistency refers to whether the non-dominant multi-source meteorological data supports the meteorological conditions represented by the candidate factors.

[0096] S229. Based on the evaluation results of the evidence strength and physical consistency of each candidate factor, output the unique dominant factor of the candidate region using a pre-defined classifier.

[0097] As described in steps S221-S229 above, during the dominant factor discrimination process, the system employs a multi-level decision tree structure to perform refined classification of each grid cell. The first threshold is used for infrared brightness temperature judgment, and a baseline value is determined by percentile analysis of cloud top temperatures from historical rainstorm cases, followed by vertical correction based on the real-time ambient temperature field. The second threshold serves as the criterion for judging the brightness temperature gradient, and is adaptively set by calculating the spatial variation coefficient of the brightness temperature field around the target area. For precipitation particle dominant factors, the third threshold determines the reflectance threshold based on the local climatological statistical characteristics of different types of precipitation systems, and performs vertical gradient correction with altitude. The fourth threshold serves as a spatial continuity condition, determined through morphological operations. The minimum connected region area is used for quantification to ensure the identification of complete precipitation entities with meteorological significance; the fifth threshold of the dominant energy field factor is based on climatological statistical analysis of sounding data and is dynamically updated according to seasonal characteristics and weather system type using a moving percentile method; the sixth threshold of the dominant water vapor transport factor is determined by calculating the climatological average and coefficient of variation of lower-level water vapor flux, and is adjusted in real time considering the influence of weather system scale and geographic latitude; when a grid cell simultaneously meets multiple factor conditions, the system initiates a multi-factor arbitration mechanism; this mechanism first quantifies the evidence strength of each candidate factor, including calculating the relative magnitude of exceeding the threshold, analyzing the spatiotemporal variation trend of physical parameters, and assessing the average spatial distribution. Regarding uniformity, in terms of exceeding threshold amplitude quantification, the system employs a standardized relative difference calculation method. For each candidate factor, it calculates the relative proportion by which its core physical parameter exceeds the corresponding threshold. For example, for the reflectivity factor, it calculates the percentage difference exceeding the third threshold, and for the energy factor, it calculates the standardized distance by which its convective effective potential energy exceeds the fifth threshold. This relative quantification method eliminates the influence of different physical dimensions, making the comparison between different factors physically meaningful. In terms of spatiotemporal trend analysis, the system constructs a multi-timescale trend assessment model. By extracting the target parameter sequences from the six most recent observation periods, it calculates the changes in short-term (e.g., the last two periods) and long-term (e.g., the last six periods) values. To assess the trend, a weighted regression algorithm is used to calculate the trend slope, and the persistence of the direction of change is statistically tested to ensure the reliability of the trend analysis. In terms of spatial distribution uniformity assessment, the system uses spatial statistical methods to establish a local assessment window centered on the target grid, calculates the coefficient of variation and autocorrelation characteristics of the core physical parameters in space, and also examines the morphological characteristics of high reflectivity areas for precipitation particle factors, including their boundary regularity and internal uniformity. The quantitative results of these three dimensions are fused through predefined weight coefficients, with the magnitude of exceeding the threshold accounting for the main weight, and spatiotemporal trends and spatial characteristics serving as important supplements, ultimately generating a comprehensive evidence strength score between 0 and 1.In the physical consistency verification stage, a multi-source data cross-validation mechanism was established. This mechanism pre-sets corresponding verification rule sets for each candidate factor type. For example, for precipitation particle-dominated factors, it verifies whether satellite infrared brightness temperature synchronously shows low-temperature cloud top characteristics and whether numerical weather prediction models indicate that the lifting condensation height is below the zero-degree layer height; for energy field-dominated factors, it verifies whether ground stations observe a decreasing trend in temperature-dew point difference and whether satellite data detects cloudless or low-cloud areas conducive to energy accumulation; for water vapor transport-dominated factors, it checks whether the radar radial velocity field exhibits low-level convergence characteristics and whether satellite water vapor channel images show that the wet area matches the transport path. Each verification rule generates a 0-1 physical consistency score through similarity calculation and logical judgment. Finally, the geometric mean of the scores of each rule is taken as the comprehensive physical consistency index of the factor. This process is used to obtain evidence strength and physical consistency. After evaluating the results, a multi-class classifier based on the random forest algorithm is used for final decision-making. The input feature vector of this classifier consists of three parts: the quantification results of the three dimensions of evidence strength, the evaluation results of the physical consistency test, and the type encoding of each dominant factor. The pre-trained random forest model calculates the winning probability of each candidate factor through a voting mechanism of multiple decision trees, and outputs its classification confidence. Furthermore, a dual decision rule is set: first, the probability of the winning factor must exceed a preset minimum confidence threshold; second, a sufficient probability difference must be maintained between the winning factor and the second-ranked factor. If both conditions are met, the winning factor is determined as the dominant factor; otherwise, the grid is marked as undetermined and re-evaluated in the next cycle. This mechanism ensures the accuracy of the decision and avoids errors caused by forced classification when evidence is insufficient.

[0098] In one embodiment, step S3, which involves constructing a motion model of a virtual cloud entity and generating a life trajectory of the virtual cloud entity over a preset time period based on the motion model, includes:

[0099] S31. Collect the spatial location sequence of each virtual cloud entity within a continuous preset period in history;

[0100] S33. Use optical flow to calculate the movement vector of each virtual cloud entity. The movement vector includes the movement speed and direction.

[0101] S34. Correct the moving vector by combining the environmental wind field data in the numerical weather prediction model to generate a motion model of the virtual cloud entity;

[0102] S35. Based on the extrapolation method of the motion model, predict the motion path within a preset period as the life development trajectory.

[0103] As described in steps S31-S35 above, during the virtual cloud entity motion modeling process, the system first obtains the spatial position data of each virtual cloud entity in a continuous time series using a centroid tracking algorithm. This algorithm calculates the geometric center based on the boundary polygon of the virtual cloud entity and establishes a temporal position sequence. Subsequently, the dense optical flow method is used to analyze the displacement changes of the virtual cloud entity's morphological field at adjacent time points. The motion vector of each pixel is obtained by solving the constant brightness equation. Then, cluster analysis is performed on the motion vectors of all pixels within the virtual cloud entity region, and the vectors and intensities of the main cluster centers are taken as the moving speed and direction of the virtual cloud entity. In the motion vector correction stage, the system extracts environmental wind field data of different pressure layers from numerical weather prediction models and uses the vertical weighted average method to calculate the guiding airflow that matches the height of the virtual cloud entity. By establishing a vector similarity model, the moving vector calculated by the optical flow method is compared and analyzed with the environmental wind field. When the directional deviation between the two exceeds a threshold, an adaptive correction mechanism is activated. The optical flow vector is deflected and its velocity is adjusted according to the type and development stage of the virtual cloud entity, ultimately generating a motion model that conforms to the principles of atmospheric dynamics. Based on the corrected motion model, the system uses an extrapolation algorithm with acceleration compensation to predict the future trajectory. This algorithm not only considers the current moving vector but also introduces the motion change trend in the most recent period as an acceleration parameter. By solving the motion differential equation, a continuous and smooth prediction path is obtained. Furthermore, the system monitors the intensity change characteristics of the virtual cloud entity. When a rapid development or decay trend is detected, the confidence interval of the prediction path is dynamically adjusted, ultimately generating a three-dimensional spatial trajectory containing a timestamp as the life development trajectory.

[0104] In one embodiment, step S4, which involves constructing a cloud motion map of the region to be predicted based on virtual cloud entities and life development trajectories, includes:

[0105] S41. Overlay the spatial distribution of the virtual cloud entities at the current moment to generate a spatial layer;

[0106] S42. Overlay the life development trajectory of each virtual cloud entity to generate a trajectory layer, and use different colors to distinguish the movement trajectory and life development trajectory of the virtual cloud entity.

[0107] S43. Register and fuse the spatial layer and trajectory layer to generate a cloud motion map;

[0108] As described in steps S41-S43 above, during the construction of the 3D cloud motion map, the system fuses the 3D spatial distribution of virtual cloud entities with their life development trajectories using a spatial registration algorithm. The spatial layer contains the 3D bounding box and volumetric features of each virtual cloud entity, while the trajectory layer provides the historical and predicted paths of each virtual cloud entity in 3D space. The system establishes a unified 3D spatial coordinate system and uses point cloud registration technology to ensure that the virtual cloud entities in the spatial layer and the path points in the trajectory layer correspond precisely in 3D space. The spatial index structure accelerates query processing, enabling rapid association and matching between virtual cloud entities and trajectories. Furthermore, during the fusion process, the system maintains the topological relationship between virtual cloud entities and trajectories, establishes a corresponding trajectory reference for each virtual cloud entity, and associates the trajectory points with the state of the virtual cloud entity they belong to.

[0109] In one embodiment, step S5, which generates changes in multi-source meteorological data of the area to be predicted based on cloud motion maps and generates intervention instructions based on these changes, includes:

[0110] S51. Predefine multiple rainstorm standard modes, each of which is composed of a specific combination of multiple virtual cloud entities in time and space.

[0111] S52. Real-time monitoring of cloud cluster movement maps, and chain matching of the current and predicted status of virtual cloud cluster entities with life development trajectories with standard rainstorm patterns;

[0112] S53. When the matching degree with any rainstorm standard pattern is detected to exceed the matching threshold, output the matching pattern and the corresponding confidence level.

[0113] S54. Generate intervention instructions based on matching patterns and corresponding confidence levels. Intervention instructions include pattern catalysis instructions and potential exploration instructions.

[0114] S55. Preset confidence range. When the confidence of the matching pattern is in the medium range, generate a pattern catalysis instruction. The pattern catalysis instruction is a test cloud entity used to verify the sensitivity of the region to develop into a full-scale rainstorm standard pattern.

[0115] S56. When a certain area does not match the standard rainstorm pattern, a potential exploration command is generated. The potential exploration command is a probabilistic cloud cluster, which is used to simulate the probability that the area will develop into the standard rainstorm pattern under random perturbation.

[0116] As described in steps S51-S56 above, in the definition of the standard rainstorm model, the system constructs a variety of combination paradigms of virtual cloud entities based on meteorological principles. Each standard rainstorm model is essentially an organic combination of a group of virtual cloud entities with specific dominant factors in three-dimensional space and time. The system uses a graph structure to formally express these models, where nodes represent virtual cloud entities with different dominant factors, and edges represent the spatial association and temporal evolution relationship between entities. For example, a typical standard rainstorm model may require energy field-dominated virtual cloud entities, water vapor transport-dominated virtual cloud entities, and cloud top development-dominated virtual cloud entities to form a coordinated configuration within a specific spatial range, and their life development trajectory shows a movement pattern of convergence towards key areas. In the real-time matching process, the system evaluates the degree of agreement between the current cloud movement map and the predefined standard rainstorm model through multi-dimensional feature similarity calculation. The matching algorithm first... First, the distribution patterns of virtual cloud entities with different dominant factors in three-dimensional space are identified, and it is detected whether they meet the spatial proximity and topological structure specified by the standard rainstorm model. Then, the life development trajectory of these virtual cloud entities is analyzed to verify whether their movement trend conforms to the temporal development requirements of the model. When the matching degree exceeds the dynamic threshold, the system generates corresponding intervention instructions based on the confidence level. For the matching model catalysis instruction with medium confidence, a test virtual cloud entity is generated. This entity has the characteristics of the relatively weak dominant factor in the current combination. By introducing this test virtual cloud entity, the sensitivity of the region to develop into a complete standard rainstorm model is verified. For unmatched regions, the potential exploration instruction generates multiple probabilistic virtual cloud entities with different dominant factors, simulates their interaction under various initial conditions, and statistically analyzes the probability distribution of their evolution into an effective standard rainstorm model, thereby comprehensively assessing the rainstorm potential of the region.

[0117] In one embodiment, step S6, which projects intervention commands into cloud motion maps to generate a rainstorm forecast report, includes:

[0118] S61. Project the generated intervention command onto the target area of ​​the cloud motion map;

[0119] S62. For a certain region, in response to the mode catalysis command, a test cloud entity is injected to perform virtual cloud entity motion simulation in that region. Based on whether a complete rainstorm standard mode is catalyzed, the environmental sensitivity coefficient of that region is output.

[0120] S63. For a certain area, the potential exploration command is injected with a probability cloud to analyze the combination of virtual cloud entities in that area, and the combination and proportion of the standard rainstorm pattern are output as the probability of potential rainstorm.

[0121] S64. Based on the high-confidence matching pattern, environmental sensitivity coefficient, and latent rainstorm probability, update the cloud cluster motion map and generate a rainstorm forecast report;

[0122] S65, List of high-risk rainstorm standard models confirmed in the rainstorm forecast report, distribution of environmental sensitivity coefficients in various regions, and probability distribution map of latent rainstorms.

[0123] As described in steps S61-S65 above, the system first uses a spatial positioning engine to precisely project intervention commands onto the target area in the cloud motion map. For mode catalysis commands, the system generates a test virtual cloud entity with specific physical parameters. The parameters of the test virtual cloud entity are optimized based on the environmental field characteristics of the target area, including reasonable water vapor content, energy value, and disturbance characteristics. The system then starts a fast numerical simulator, injects the test virtual cloud entity into the current environmental field, observes the interaction process between the test virtual cloud entity and the existing virtual cloud entity, and records in detail whether the test virtual cloud entity can catalyze the formation of a complete rainstorm standard model. Based on the catalytic experiment results, the system uses a logistic regression model to calculate the environmental sensitivity coefficient of the region. This coefficient comprehensively considers factors such as the number of successful catalysis attempts, the intensity of the catalytic process, and the time required for catalysis. For example, in region A, the existence of an energy-dominated virtual cloud entity indicates that the region has sufficient atmospheric instability and energy. Simultaneously, the existence of a water vapor-dominated virtual cloud entity indicates that the region has abundant water vapor supply. Furthermore, their life trajectories show that they will overlap in one hour. The system matches a standard "energy-water vapor coupled" rainstorm model, but with a confidence level of 65%, which is considered moderate, due to the lack of a key "dynamic lift" signal, i.e., cloud top development-dominated. Therefore, the system generates a model catalytic command, which creates a test virtual cloud entity whose main... The guiding factor is designed as "cloud-top development-driven," simulating a virtual disturbance that can provide initial dynamic lift. The system projects this test virtual cloud entity between the energy field and water vapor field entities in region A and initiates a rapid simulation. Two scenarios exist: first, the test virtual cloud entity triggers the aggregation of energy and water vapor, rapidly evolving into a deep cloud with a highly reflective core, generating a complete standard rainstorm pattern; second, the test virtual cloud entity dissipates quickly after injection, failing to effectively organize energy and water vapor, and the cloud combination in region A gradually weakens. Based on the generated scenario output environmental sensitivity coefficient, if scenario one occurs, the system outputs an environmental sensitivity coefficient of 0.9 (relatively high) for region A, meaning that the probability of rainstorms in region A is extremely high; if scenario two occurs, the system outputs an environmental sensitivity coefficient of 0 for region A.A value of 2 (lower) means that the probability of heavy rain in region A is very low. For potential exploration commands, the system generates a large number of probabilistic virtual cloud entities using the Monte Carlo method. These probabilistic virtual cloud entities have random initial positions and physical parameters. While keeping the current environmental field constant, the system runs multiple sets of parallel simulation experiments to observe the evolution of these probabilistic virtual cloud entities combined with existing virtual cloud entities. Each simulation experiment records whether a standard heavy rain pattern occurs. Finally, the proportion of effective combinations is statistically analyzed as the potential heavy rain probability for the region. Furthermore, the system uses cluster analysis to identify high-incidence pattern types and calculates the probability of occurrence for each standard heavy rain pattern individually. In the data fusion stage, the system... A multi-dimensional update mechanism integrates high-confidence matching patterns, environmental sensitivity coefficients, and latent storm probabilities into cloud motion maps. Hierarchical rendering technology is used to overlay these assessment results in three-dimensional space with varying transparency. Confirmed high-risk areas are highlighted with bold warning colors, while potential risk areas use gradient colors to represent probability levels. The final storm forecast report contains three core components: a detailed list of validated high-risk storm standard patterns, a spatial distribution map of environmental sensitivity coefficients for each region, and an isosurface distribution map of the overall latent storm probability. Furthermore, the report provides interactive query functionality, allowing users to delve into the supporting evidence and uncertainty analysis for each forecast conclusion.

[0124] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for constructing an artificial intelligence-based rainstorm prediction model based on multi-source heterogeneous data fusion, characterized in that, Collect multi-source meteorological data for the area to be predicted, including satellite remote sensing data, weather radar data, ground meteorological station data, and numerical weather prediction model data; Based on the preset dominant factor discrimination rules, the region to be predicted is divided into different dominant type regions, and virtual cloud entities are constructed based on the dominant type regions; A uniform spatial resolution grid is obtained, and multi-source meteorological data within each grid area are used as the data to be analyzed. Dominant factors are generated based on the dominant factor discrimination rules for the data to be analyzed; The dominant factors include cloud top development, precipitation particle dominance, energy field dominance, and water vapor transport dominance. Taking the dominant factor as the root system, the non-dominant multi-source meteorological data is absorbed into the surrounding grid range in a root-like manner based on the principle of physical consistency, and the boundary of the virtual cloud entity is dynamically generated. A preset volume threshold for virtual cloud entities is set. If the volume of a virtual cloud entity exceeds the volume threshold, root-like absorption will stop. Construct a motion model of a virtual cloud entity, and generate the life development trajectory of the virtual cloud entity within a preset time based on the motion model; Construct a cloud movement map of the region to be predicted based on virtual cloud entities and life development trajectories; Based on cloud motion maps, generate changes in multi-source meteorological data for the area to be predicted, and generate intervention instructions based on these changes. Multiple standard rainstorm models are predefined, and each standard rainstorm model is composed of a specific combination of multiple virtual cloud entities in space and time. Real-time monitoring of cloud movement maps, and chain matching of the current and predicted states of virtual cloud entities with life development trajectories with standard rainstorm patterns; When the match degree with any standard rainstorm pattern exceeds the match threshold, the matching pattern and the corresponding confidence level are output. Intervention instructions are generated based on matching patterns and corresponding confidence levels. These intervention instructions include pattern catalysis instructions and potential exploration instructions. Intervention commands are projected into cloud motion maps to generate rainstorm forecast reports.

2. The method for constructing an artificial intelligence-based rainstorm prediction model based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The steps for collecting multi-source meteorological data of the area to be predicted, including satellite remote sensing data, weather radar data, ground meteorological station data, and numerical weather prediction model data, are as follows: Multi-source meteorological data of the area to be predicted are collected at a preset period, and spatiotemporal labels are marked for each type of data. The multi-source meteorological data are uniformly remapped to a preset uniform spatial resolution grid for spatiotemporal alignment. Among them, satellite remote sensing data includes infrared brightness temperature, water vapor channel brightness temperature and visible light albedo of the area to be predicted; Weather radar data includes the combined reflectivity, basic radial velocity, and vertically integrated liquid water content of the area to be predicted; Ground meteorological station data includes temperature, humidity, air pressure, wind speed, wind direction, and minute-level precipitation for the area to be predicted; Numerical weather prediction model data includes the convective available potential energy field, water vapor flux field, and vertical velocity field of the region to be predicted.

3. The method for constructing an artificial intelligence-based rainstorm prediction model based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The steps for generating dominant factors from the data to be analyzed based on the dominant factor discrimination rule are as follows: When the infrared brightness temperature of satellite remote sensing data within the grid area is lower than the first preset threshold and its local gradient exceeds the second preset threshold, the dominant factor within the grid area is determined to be the dominant factor for cloud top development. When the combined reflectivity of weather radar data within the grid area is higher than the third preset threshold and its spatial continuity meets the fourth preset condition, the dominant factor within the grid area is determined to be the precipitation particle dominant factor. When the convective effective potential energy of the numerical weather prediction model data within the grid range is higher than the fifth preset threshold, the dominant factor within that grid range is determined to be the dominant factor of the energy field. When the lower-level water vapor flux of the numerical weather prediction model data within the grid range is higher than the sixth preset threshold, the dominant factor within that grid range is determined to be the dominant factor for water vapor transport. If multiple types of dominant factors exist within the same grid range, then the grid range is taken as a candidate region, and the multiple types of dominant factors within the candidate region are taken as candidate dominant factors. For each candidate factor in the candidate region, the strength of evidence is measured. The strength of evidence includes the magnitude of exceeding a preset threshold, the spatiotemporal variation trend of the physical parameters of the candidate factor, and the uniformity of its spatial distribution. Simultaneously, the physical consistency between candidate factors and non-dominant multi-source meteorological data within the candidate region is evaluated, where physical consistency refers to whether the non-dominant multi-source meteorological data supports the meteorological conditions represented by the candidate factors. Based on the evaluation results of the evidence strength and physical consistency of each candidate factor, a pre-defined classifier outputs the unique dominant factor of the candidate region.

4. The method for constructing an artificial intelligence-based rainstorm prediction model based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The steps for constructing a motion model of a virtual cloud entity and generating a life development trajectory of the virtual cloud entity within a preset time based on the motion model are as follows: Collect the spatial location sequence of each virtual cloud entity within a continuous preset period of history; The optical flow method is used to calculate the movement vector of each virtual cloud entity, which includes the movement speed and direction; By combining environmental wind field data from numerical weather prediction models to correct the moving vector, a motion model of the virtual cloud entity is generated. The motion path within a preset period is predicted using extrapolation based on the motion model as the trajectory of life development.

5. The method for constructing an artificial intelligence-based rainstorm prediction model based on multi-source heterogeneous data fusion according to claim 4, characterized in that, The steps for constructing a cloud movement map of the region to be predicted based on virtual cloud entities and life development trajectories are as follows: The spatial distribution of the virtual cloud entities at the current moment is superimposed to generate a spatial layer; The life development trajectory of each virtual cloud entity is superimposed to generate a trajectory layer, and different colors are used to distinguish the movement trajectory and life development trajectory of the virtual cloud entity. The spatial layer and trajectory layer are registered and fused to generate a cloud motion map.

6. The method for constructing an artificial intelligence-based rainstorm prediction model based on multi-source heterogeneous data fusion according to claim 1, characterized in that, The steps for generating changes in multi-source meteorological data of the area to be predicted based on cloud motion maps, and generating intervention instructions based on these changes, are as follows: A pre-set confidence range is used to generate a model catalysis instruction when the confidence of the matching pattern is in the medium range. The model catalysis instruction is a test cloud entity used to verify the sensitivity of the region to develop into a full-scale rainstorm standard pattern. If a certain area does not match the standard rainstorm pattern, a potential exploration command is generated. The potential exploration command is a probabilistic cloud cluster used to simulate the probability that the area will develop into the standard rainstorm pattern under random perturbation.

7. The method for constructing an artificial intelligence-based rainstorm prediction model based on multi-source heterogeneous data fusion according to claim 6, characterized in that, The steps for projecting intervention commands into cloud motion maps to generate rainstorm forecast reports are as follows: The generated intervention commands are projected onto the target area of ​​the cloud motion map; For a certain region, in response to the pattern catalysis command, a test cloud entity is injected to perform virtual cloud entity motion simulation in that region. Based on whether a complete rainstorm standard pattern is catalyzed, the environmental sensitivity coefficient of that region is output. For a certain area, in response to the potential exploration command, a probability cloud is injected to analyze the combination of virtual cloud entities in that area, and the combination and proportion of the standard rainstorm pattern are output as the probability of potential rainstorm. The cloud cluster motion map is updated based on the high-confidence matching pattern, environmental sensitivity coefficient, and latent rainstorm probability, and a rainstorm forecast report is generated. The rainstorm forecast report includes a list of high-risk rainstorm standard models, the distribution of environmental sensitivity coefficients for each region, and a probability distribution map of potential rainstorms.

Images