An application method for mountainous area near weather forecast based on AI technology

Abstract

The application provides an application method for mountainous area near weather forecasting based on AI technology, integrates satellite, radar, numerical model, terrain and geographical multi-dimensional multi-source data, realizes heterogeneous data standardization fusion through space-time grid pretreatment, makes up for the defects of sparse ground observation sites and local data loss in mountainous areas, breaks the limitations of traditional single data source forecasting, and provides more accurate data basis for weather research and judgment under complex terrain in mountainous areas. Based on the radar radial velocity data and the VAD algorithm, the application completes fine wind field inversion, generates a high-credibility regional wind field map, can accurately identify local wind shear, airflow convergence, vertical lifting and other mesoscale and small-scale convective triggering features in mountainous areas, solves the problems of traditional sounding data scarcity and mountainous area wind field observation blank, and provides core dynamic field basis for strong convective weather prediction.

Description

Technical Field

[0001] This invention relates to the field of nowcasting, and more particularly to a method for applying AI technology to nowcasting weather in mountainous areas. Background Technology

[0002] Due to their complex terrain, mountainous areas are prone to frequent severe convective weather events such as short-duration heavy rainfall and thunderstorms, often triggering serious meteorological disasters. Therefore, conducting nowcasting (0-2 hours) for this region is of significant importance for disaster prevention and mitigation. However, existing nowcasting technologies still face the following prominent shortcomings in complex terrain applications: First, the ability to predict precipitation from topography is significantly insufficient. Current operational radar quantitative precipitation estimation mainly relies on traditional ZR relationship transformation and extrapolation algorithms. These methods do not consider the enhancing effects of topographic uplift and local convergence on precipitation. At the same time, numerical models have limited ability to characterize the dynamic effects of small- to medium-scale topography, making it difficult to accurately simulate the impact of complex topography on local circulation and precipitation distribution. This often leads to a systematic underestimation of precipitation on windward slopes in mountainous areas.

[0003] Second, the early warning system for thunderstorms and strong winds in mountainous areas relies on manual identification, resulting in a low level of automation. The identification of characteristics of severe convective storms such as bow echoes and mesocyclones still mainly depends on the subjective experience of forecasters. This is not only inefficient but also has a high rate of missed and false alarms, which seriously restricts the ability to respond quickly to sudden strong wind weather. There is still a lack of effective automated intelligent identification methods.

[0004] Third, the application of radar wind field inversion technology is insufficient. Radar radial velocity data contains rich information on wind field structure, which is of great value for improving the accuracy and timeliness of short-term weather forecasts. However, most municipal meteorological stations currently only use it for qualitative display and have not yet used mature methods such as velocity-azimuth display (VAD) to invert gridded wind fields and integrate them into the actual forecasting process, resulting in the forecasting potential of this key observational information not being effectively released.

[0005] Fourth, there is a disconnect between early warning products and geographic information. Existing nowcasting conclusions are mostly described in the form of regional areal data, and risk warnings cannot be accurately applied to the township level or even smaller geographic units. The spatial resolution is insufficient, making it difficult to meet the actual needs of grassroots disaster prevention decision-making for refined location information, thus weakening the operability and support effectiveness of early warnings.

[0006] Fifth, the comprehensive utilization of multi-source data is low, and the collaborative analysis capability is weak. Satellite, radar, numerical model, and topographic geographic information data are often acquired and processed independently in current operations, lacking a unified spatiotemporal matching and collaborative fusion mechanism. In particular, municipal-level stations have not yet formed an integrated comprehensive analysis technology framework, making it difficult for various data advantages to complement and coordinate, thus hindering the overall improvement of the ability to provide refined forecasts of complex weather processes.

[0007] In summary, existing nowcasting technologies have significant shortcomings in quantitative precipitation estimation, automatic wind identification, wind field inversion application, precise positioning and early warning, and multi-source data fusion. There is an urgent need to develop intelligent and collaborative new forecasting systems to effectively improve the accuracy of early warnings and operational service capabilities for severe convective weather in complex terrain areas. Summary of the Invention

[0008] To address the aforementioned technical problems, this invention provides a method for applying AI technology to near-term weather forecasting in mountainous areas.

[0009] To achieve the above objectives, the technical solution of this invention is as follows: In a first aspect, the present invention provides a method for applying AI technology to near-term weather forecasting in mountainous areas, the method comprising: Multi-source data for predicting mountainous areas is acquired, and the multi-source data is preprocessed to obtain spatiotemporally matched multi-source gridded data. The multi-source data includes meteorological satellite observation data, radar base data and radar combined reflectivity map, numerical model wind field data, specific humidity data, topographic data and administrative division boundary vector data. Based on the numerical model wind field data, the specific humidity data, and the topographic data, the topographic vertical velocity is calculated; based on the topographic vertical velocity and the specific humidity data, combined with the horizontal divergence discrimination condition, the topographic precipitation contribution is calculated, and the topographic precipitation contribution is fused with the precipitation field obtained based on the radar combined reflectivity image conversion to generate a refined precipitation forecast map. Radial velocity data is decoded from the radar base data, projected onto a latitude and longitude grid, and then inverted and plotted based on the VAD algorithm to generate a VAD wind field inversion map. A multimodal image sequence comprising the radar combined reflectivity map, the VAD wind field inversion map, the radial velocity data, and the meteorological satellite observation data is constructed and input into a pre-trained multimodal large language model; the multimodal large language model identifies strong convection features and outputs structured strong convection forecast results; Based on the refined precipitation forecast map and the severe convection forecast results, areas with precipitation intensity exceeding the standard and severe convection feature points are extracted as meteorological risk points. The meteorological risk points are matched with the administrative boundaries of townships using a spatial connectivity algorithm to generate a township-level risk warning list, which is then overlaid onto the geographic information base map to generate a visualized warning image.

[0010] In some embodiments, calculating the terrain vertical velocity specifically includes: calculating the terrain vertical velocity based on the horizontal wind speed component in the numerical model wind field data and the terrain height spatial gradient determined based on the terrain data; the formula for calculating the terrain vertical velocity is: In the formula, Vertical velocity relative to the terrain; Ground density; It is the acceleration due to gravity; This is the horizontal wind speed vector; The terrain height spatial gradient; the horizontal divergence discrimination condition specifically includes: retaining only the horizontal divergence. Regional orographic precipitation contribution, eliminating spurious precipitation signals outside convergence zones; where, horizontal divergence u and v represent the zonal and radial wind speeds at ground level, respectively; x and y represent the spatial distances in the zonal and longitudinal directions, respectively.

[0011] In some embodiments, generating a VAD wind field inversion map specifically includes: extracting preset key parameters from the radar base data based on the radar base data encoding format, and restoring the radial velocity data according to the corresponding encoding restoration formula; within a local grid window, based on the physical relationship between the radial velocity data and the horizontal wind field in polar coordinates, solving a system of equations using the least squares method to invert and obtain the east-west wind speed component and the north-south wind speed component; performing Gaussian smoothing on the east-west wind speed component and the north-south wind speed component, and synthesizing a high-resolution two-dimensional wind field based on the smoothed wind speed component; and drawing the VAD wind field inversion map by superimposing the administrative division boundary vector data on the high-resolution two-dimensional wind field.

[0012] In some embodiments, the multimodal large language model identifies severe convective features and outputs structured severe convective weather forecast results, including: automatically identifying bow echoes, gust fronts, mesocyclones, velocity pairs, and three-body scattering features in images using the multimodal large language model to determine the potential location of severe convective weather; and outputting structured data in JSON format containing a wind gap area field, wherein the wind gap area field contains a list of township names whose wind speeds reach a preset threshold, and data on the maximum wind speed range containing a gust intensity field.

[0013] In some embodiments, generating a township-level risk warning list includes: for lightning warnings, selecting strong echo points with a radar combined reflectivity of not less than 65 dBZ as lightning risk points; for heavy precipitation warnings, statistically analyzing grid points in the refined precipitation forecast map where precipitation intensity exceeds 50 mm / h as heavy precipitation risk points; for thunderstorm and gale warnings, extracting grid points with wind speeds exceeding a preset gale threshold from the VAD wind field inversion map, and verifying them based on the velocity fuzzy features in the radial velocity data, and then using these as thunderstorm and gale risk points; using a geospatial connectivity algorithm, performing spatial correlation analysis between the lightning risk points, the heavy precipitation risk points, and the thunderstorm and gale risk points and the township administrative division boundary vector data respectively, determining the township area to which each risk point falls, and classifying and statistically analyzing them according to risk type to generate the township-level risk warning list.

[0014] In some embodiments, the method further includes: configuring the main control program to cyclically call the independent executable scripts corresponding to the above steps at preset time intervals to achieve fully automated operation; the preset time interval is 10 minutes; automatically storing the generated refined precipitation forecast map, the VAD wind field inversion map, the severe convection forecast results, and the township-level risk warning list to a specified directory according to the generation date, and generating a dynamically refreshed webpage index file to achieve unattended operational operation.

[0015] This invention provides an AI-based method for near-term weather forecasting in mountainous areas. It integrates multi-dimensional, multi-source data from satellites, radar, numerical models, and topographic geography. Through spatiotemporal point-based preprocessing, it achieves standardized fusion of heterogeneous data, overcoming the limitations of sparse ground observation stations and missing local data in mountainous areas. This breaks the limitations of traditional single-source forecasting and provides a more accurate data foundation for weather analysis in complex mountainous terrain. By quantifying topographic vertical velocity and combining it with horizontal divergence dynamics, it accurately calculates the contribution of topographic precipitation and integrates topographic precipitation with radar-derived precipitation fields. This fully adapts to the modulation effect of mountainous terrain on precipitation, accurately depicting the localized differential precipitation characteristics in mountainous areas and significantly improving the accuracy of short-term precipitation forecasts. Furthermore, this invention uses radar radial velocity data combined with the VAD algorithm to complete refined wind field inversion, generating highly reliable regional wind field maps. These maps can accurately identify small- and medium-scale convective triggering features such as local wind shear, airflow convergence, and vertical lift in mountainous areas, solving the problems of scarce traditional radiosonde data and gaps in mountainous wind field observations. This provides core dynamic field evidence for severe convective weather forecasting.

[0016] Furthermore, this invention constructs a multi-dimensional meteorological image sequence and automatically fuses multi-modal features of cloud images, echoes, wind fields, and speed using a pre-trained multi-modal large language model. This intelligently identifies latent precursor signals of severe convection, replacing traditional manual experience-based judgment methods. This effectively improves the timeliness and accuracy of severe convection identification, solving the problems of sudden onset of severe convection in mountainous areas, delayed manual predictions, and high false alarm rates. This invention uses a spatial connectivity algorithm to accurately match precipitation and severe convection risk points to township administrative units, generating standardized warning lists and visualized warning images. This achieves a closed-loop transformation from meteorological element forecasting to localized risk warnings, refining forecast results to the township level. This aligns perfectly with the practical application needs of meteorological disaster prevention, warning issuance, and emergency response at the grassroots level in mountainous areas, demonstrating strong practicality and applicability. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the method for applying AI technology to near-term weather forecasting in mountainous areas, as provided in an embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings. The described embodiments should not be regarded as limitations on 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.

[0019] In the following description, references to "some embodiments" refer to a subset of all possible embodiments; however, it is understood that "some embodiments" may be the same or different subsets of all possible embodiments and may be combined with each other without conflict. Unless otherwise defined, all technical and scientific terms used in the embodiments of the invention have the same meaning as commonly understood by one of ordinary skill in the art to which the embodiments of the invention pertain. The terminology used in the embodiments of the invention is for the purpose of describing the embodiments of the invention only and is not intended to limit the invention.

[0020] The following describes an exemplary application of AI-based near-term weather forecasting equipment for mountainous areas according to embodiments of the present invention. This AI-based near-term weather forecasting equipment for mountainous areas provided in the embodiments of the present invention can be implemented as a terminal or as a server. In one implementation, the AI-based near-term weather forecasting equipment for mountainous areas provided in the embodiments of the present invention can be implemented as various types of terminals such as laptops, tablets, desktop computers, and mobile devices. In another implementation, the AI-based near-term weather forecasting equipment for mountainous areas provided in the embodiments of the present invention can also be implemented as a server. The server can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks (CDN), and big data and artificial intelligence platforms. The terminal and the server can be directly or indirectly connected via wired or wireless communication, and no limitations are imposed in the embodiments of the present invention. The following will describe an exemplary application when the AI-based near-term weather forecasting equipment for mountainous areas is used as a server.

[0021] This invention provides a method for applying AI technology to near-term weather forecasting in mountainous areas. (See also...) Figure 1 , Figure 1 This is a flowchart illustrating a method for applying AI technology to near-term weather forecasting in mountainous areas, provided by an embodiment of the present invention. Figure 1 The steps shown are explained.

[0022] S110, acquire multi-source data for the predicted mountainous area, preprocess the multi-source data to obtain spatiotemporally matched multi-source gridded data; the multi-source data includes meteorological satellite observation data, radar base data and radar combined reflectivity map, numerical model wind field data, specific humidity data, topographic data and administrative division boundary vector data.

[0023] Here, meteorological satellite observation data refers to all-weather, large-scale observation data obtained by meteorological satellites through onboard detection equipment to remotely sense the Earth's atmosphere, cloud system, and underlying surface. It covers core observation elements such as cloud top temperature, cloud cover, water vapor distribution, infrared radiation, and visible light reflectance. It has the characteristics of wide spatial coverage and no ground blind spots, which can make up for the deficiency of sparse ground observation stations in mountainous areas.

[0024] Here, radar base data refers to the raw detection data after weather radar scanning and detection, without secondary processing or preprocessing. It is the original data source for all radar-derived meteorological products. It includes the raw echo information corresponding to each sampling distance, azimuth angle, and elevation angle within the radar detection range, the raw sampled value of radial velocity, spectral width data, etc. It has the characteristics of high time resolution and high local detection accuracy, and can accurately capture the subtle changes of small-scale convective systems in mountainous areas.

[0025] Here, the radar composite reflectivity map refers to a two-dimensional visualized raster image generated by vertically integrating and fusing radar reflectivity factors at multiple elevation angles based on radar base data. It can intuitively reflect the comprehensive echo intensity distribution of water vapor condensates such as clouds, raindrops, and ice crystals in the atmosphere, serving as visualized data for judging the strength of convective cloud clusters and precipitation potential. For example, a higher composite reflectivity value indicates more vigorous local convection development, and a higher probability and intensity of precipitation.

[0026] Here, numerical model wind field data refers to gridded wind field meteorological data obtained through simulation calculations based on atmospheric dynamics and thermodynamic equations using atmospheric numerical forecasting models. It mainly includes two core elements: horizontal wind direction and wind speed, covering upper-air and near-surface wind field information at different altitudes and times, and can reflect large-scale atmospheric horizontal motion patterns. For example, hourly gridded wind field data output from WRF numerical models and ECMWF numerical models can be used.

[0027] Here, the gridded atmospheric specific humidity meteorological parameter output by the specific humidity data index model is defined as the mass of water vapor contained in a unit mass of moist air, usually expressed in g / kg. It is a core physical quantity characterizing atmospheric water vapor content and measuring precipitation water vapor conditions. The higher the specific humidity value, the more abundant the atmospheric water vapor reserves, and the easier it is for precipitation to form under conditions of orographic lifting and convection triggering.

[0028] Here, topographic data refers to gridded geographic data that characterizes the topographic features of the underlying surface in mountainous areas. Its core elements include altitude, slope, aspect, and topographic relief. It accurately depicts the complex topographic features of mountainous areas, including their undulating terrain and numerous gullies, and is the core foundational geographic data for calculating topographically forced vertical motion and analyzing the effects of orographic precipitation. For example, 30m resolution DEM digital elevation topographic data can be used.

[0029] Here, administrative boundary vector data refers to geographic boundary data stored in vector format, which accurately records the boundary outlines, ranges and attribute information of administrative regions at the national, provincial, municipal and township levels. It has the characteristics of accurate boundaries and spatial matching, and is used to accurately transfer meteorological risks, precipitation and convection characteristics to the corresponding administrative units to achieve refined early warning at the township level.

[0030] Here, spatiotemporally matched multi-source gridded data refers to the standardized data product obtained after uniform preprocessing of the aforementioned multi-source data. Multi-source data suffers from inconsistent spatial resolution, projected coordinate systems, and temporal resolution. This invention addresses this by performing preprocessing operations such as resampling, coordinate transformation, temporal interpolation, spatial registration, missing value imputation, and outlier removal to uniformly map all data to the same latitude and longitude grid and the same time scale, forming a spatially aligned, temporally synchronized, and format-uniform gridded dataset.

[0031] S120, based on the numerical model wind field data, the specific humidity data, and the topographic data, calculate the topographic vertical velocity; based on the topographic vertical velocity and the specific humidity data, combined with the horizontal divergence discrimination condition, calculate the topographic precipitation contribution, and fuse the topographic precipitation contribution with the precipitation field obtained based on the radar combined reflectivity image conversion to generate a refined precipitation forecast map.

[0032] Here, topographic vertical velocity refers to the vertical velocity of the atmosphere caused by the obstruction, lifting, and bypassing effects of mountainous terrain during horizontal airflow. It is a core physical quantity characterizing the intensity of topographically forced vertical atmospheric motion. Complex mountainous terrain forces near-surface airflow to rise and fall, directly affecting water vapor condensation, cloud development, and precipitation formation, and is the core driving force behind orographic precipitation formation.

[0033] Here, the horizontal divergence discrimination condition is a preset atmospheric dynamic field discrimination threshold condition in this invention, constructed based on the convergence and divergence characteristics of the horizontal wind field. Horizontal divergence is a physical quantity characterizing the degree of convergence or divergence of horizontal airflow in the atmosphere. Airflow convergence (negative divergence) triggers air lifting and water vapor accumulation, which are important dynamic conditions for precipitation and convection. This condition is used to screen topographic movement areas with precipitation generation potential, eliminate invalid vertical topographic movement interference, and improve the accuracy of topographic precipitation calculation.

[0034] Here, orographic precipitation contribution refers to the precipitation component contributed solely by the orographic forcing effect in mountainous areas. It is distinct from precipitation from large-scale weather systems and is the local precipitation formed by the upward airflow from the orographic system and the promotion of water vapor condensation. It is an important component of refined precipitation in mountainous areas and can accurately depict the differentiated precipitation characteristics caused by local topography in mountainous areas.

[0035] Here, the refined precipitation forecast map refers to a high spatial and temporal resolution mountain precipitation forecast visualization product generated by integrating radar-retrieved precipitation fields and topographic precipitation contributions. This forecast map fully considers both large-scale precipitation systems and local topographic precipitation effects in mountainous areas, solving the problems of low accuracy and lack of detail in traditional precipitation forecasts in areas with complex mountain terrain, and accurately reflecting the differences in local precipitation in mountainous areas.

[0036] S130, the radial velocity data is decoded from the radar base data, the radial velocity data is projected onto the latitude and longitude grid, and then inverted and plotted based on the VAD algorithm to generate a VAD wind field inversion map.

[0037] Here, radial velocity data refers to the raw wind field detection data extracted from radar base data. It represents the velocity of the detected air particles relative to the radial direction of the weather radar. The velocity is negative when moving towards the radar and positive when moving away from the radar. It is the core data for inverting the atmospheric horizontal wind field.

[0038] Here, the VAD wind field inversion map refers to the gridded wind field visualization image generated after calculating and inverting radar radial velocity data through the VAD algorithm. It intuitively displays the spatial distribution of wind direction and wind speed at different heights and locations within the detection area, and can accurately capture convection triggering characteristics such as wind field convergence and shear in small and medium-scale mountainous areas.

[0039] S140, construct a multimodal image sequence containing the radar combined reflectivity map, the VAD wind field inversion map, the radial velocity data, and the meteorological satellite observation data, and input it into a pre-trained multimodal large language model; the multimodal large language model identifies strong convection features and outputs structured strong convection forecast results.

[0040] Here, multimodal image sequences refer to ordered sequences formed by combining meteorological images and data products from different data sources and with different representational dimensions according to temporal and spatial association rules. Specifically, this invention includes four types of modal data: radar combined reflectivity maps, VAD wind field inversion maps, radial velocity data visualization images, and meteorological satellite observation images. It also covers multi-dimensional features such as convective echoes, wind fields, cloud systems, and water vapor, providing complete input for comprehensive analysis using multimodal large-scale language models.

[0041] Here, severe convective characteristics refer to various meteorological features that can characterize the occurrence, development, and evolution of severe convective weather, including small- and medium-scale meteorological anomalies such as high echo intensity of convective cloud clusters, high echo top height, wind field convergence and shear, radial velocity anomalies, sudden drop in cloud top temperature, and rapid accumulation of water vapor. These are the basis for identifying severe convective weather such as thunderstorms, strong winds, and hail.

[0042] Here, structured severe convection forecast results refer to standardized and structured forecast data output by multimodal large language models. Unlike disordered image features, they contain quantified and regular forecast fields such as the location, time, intensity level, impact range, and evolution trend of severe convection, which can be directly used for subsequent risk extraction and early warning generation.

[0043] S150, based on the refined precipitation forecast map and the severe convection forecast results, areas with precipitation intensity exceeding the standard and severe convection feature points are extracted as meteorological risk points respectively; the meteorological risk points are matched with township administrative boundaries using a spatial connection algorithm to generate a township-level risk warning list, which is then overlaid onto the geographic information base map to generate a visualized warning image.

[0044] Here, the township-level risk warning list refers to the structured warning ledger generated after spatial matching. It is summarized by township administrative unit and includes core warning information such as risk type (precipitation risk, severe convective weather risk), risk level, risk location, expected occurrence time, and impact range for each township, providing accurate data support for the issuance of meteorological warnings at the grassroots level.

[0045] Here, the visualized early warning image refers to the intuitive and visualized image generated by overlaying township-level meteorological risk points, risk areas, and early warning level information onto the geographic information base map. It can intuitively display the distribution and intensity differences of meteorological risks in all townships across the region, and realize the visualization and intuitive judgment of risks.

[0046] This invention provides an AI-based method for near-term weather forecasting in mountainous areas. It integrates multi-dimensional, multi-source data from satellites, radar, numerical models, and topographic geography. Through spatiotemporal point-based preprocessing, it achieves standardized fusion of heterogeneous data, overcoming the limitations of sparse ground observation stations and missing local data in mountainous areas. This breaks the limitations of traditional single-source forecasting and provides a more accurate data foundation for weather analysis in complex mountainous terrain. By quantifying topographic vertical velocity and combining it with horizontal divergence dynamics, it accurately calculates the contribution of topographic precipitation and integrates topographic precipitation with radar-derived precipitation fields. This fully adapts to the modulation effect of mountainous terrain on precipitation, accurately depicting the localized differential precipitation characteristics in mountainous areas and significantly improving the accuracy of short-term precipitation forecasts. Furthermore, this invention uses radar radial velocity data combined with the VAD algorithm to complete refined wind field inversion, generating highly reliable regional wind field maps. These maps can accurately identify small- and medium-scale convective triggering features such as local wind shear, airflow convergence, and vertical lift in mountainous areas, solving the problems of scarce traditional radiosonde data and gaps in mountainous wind field observations. This provides core dynamic field evidence for severe convective weather forecasting.

[0047] Furthermore, this invention constructs a multi-dimensional meteorological image sequence and automatically fuses multi-modal features of cloud images, echoes, wind fields, and speed using a pre-trained multi-modal large language model. This intelligently identifies latent precursor signals of severe convection, replacing traditional manual experience-based judgment methods. This effectively improves the timeliness and accuracy of severe convection identification, solving the problems of sudden onset of severe convection in mountainous areas, delayed manual predictions, and high false alarm rates. This invention uses a spatial connectivity algorithm to accurately match precipitation and severe convection risk points to township administrative units, generating standardized warning lists and visualized warning images. This achieves a closed-loop transformation from meteorological element forecasting to localized risk warnings, refining forecast results to the township level. This aligns perfectly with the practical application needs of meteorological disaster prevention, warning issuance, and emergency response at the grassroots level in mountainous areas, demonstrating strong practicality and applicability.

[0048] In some embodiments, calculating the terrain vertical velocity specifically includes: calculating the terrain vertical velocity based on the horizontal wind speed component in the numerical model wind field data and the terrain height spatial gradient determined based on the terrain data; the formula for calculating the terrain vertical velocity is: In the formula, Vertical velocity relative to the terrain; Ground density; It is the acceleration due to gravity; This is the horizontal wind speed vector; The terrain height spatial gradient; the horizontal divergence discrimination condition specifically includes: retaining only the horizontal divergence. Regional orographic precipitation contribution, eliminating spurious precipitation signals outside convergence zones; where, horizontal divergence u and v represent the zonal and radial wind speeds at ground level, respectively; x and y represent the spatial distances in the zonal and longitudinal directions, respectively.

[0049] In some embodiments, generating a VAD wind field inversion map specifically includes: extracting preset key parameters from the radar base data based on the radar base data encoding format, and restoring the radial velocity data according to the corresponding encoding restoration formula; within a local grid window, based on the physical relationship between the radial velocity data and the horizontal wind field in polar coordinates, solving a system of equations using the least squares method to invert and obtain the east-west wind speed component and the north-south wind speed component; performing Gaussian smoothing on the east-west wind speed component and the north-south wind speed component, and synthesizing a high-resolution two-dimensional wind field based on the smoothed wind speed component; and drawing the VAD wind field inversion map by superimposing the administrative division boundary vector data on the high-resolution two-dimensional wind field.

[0050] Here, preset key parameters refer to the core, mandatory parameter fields used for parsing radial velocity data, pre-selected and determined according to the radar base data's proprietary encoding format. These parameters form the foundational set for decoding the original radar data. These parameters are standard scan output parameters for weather radar, corresponding to fixed spatial location and physical quantity information detected by the radar, requiring no additional custom configuration. For example, preset key parameters may include radar scan azimuth, scan elevation, detection range library number, original radial velocity code value, data quality marker, timestamp, etc. All parameters together constitute a complete data source for radial velocity reconstruction.

[0051] Here, radar base data encoding format refers to the standardized binary encoding specification preset by weather radar equipment manufacturers for storing and transmitting raw radar detection data. Different models of weather radar are equipped with their own fixed encoding formats to encapsulate massive amounts of scanned data. Radar base data is not plaintext data, but is stored in binary encrypted encoding form, with various meteorological parameters and spatial parameters encapsulated according to fixed bytes and a fixed order.

[0052] Here, the encoding restoration formula refers to the mathematical conversion formula configured for the corresponding radar base data encoding format, used to convert the original radar encoded values ​​into actual physical observation values. To compress storage space and standardize data format, radars do not directly store actual wind speed values; instead, they store normalized encoded values. These values ​​need to be inversely calculated using a dedicated restoration formula to obtain the physically meaningful true radial velocity values. Different radar models and parameters correspond to specific restoration formulas, which can eliminate data deviations caused by encoding compression and ensure the authenticity of the original data.

[0053] The local grid window in this invention is a grid area of ​​a preset fixed size, used to limit the data calculation range of a single wind field inversion. Only the radial velocity sample data within the window is called to solve the equation system, avoiding interference from mixed global data.

[0054] Here, the physical relationship of the horizontal wind field refers to the inherent mathematical and physical correlation between the radar radial velocity and the horizontal two-dimensional wind field (east-west wind speed and north-south wind speed) in the polar coordinate system in atmospheric dynamics. The physical principle is that the radial velocity observed by radar is the projection component of the horizontal wind field in the radar radial direction. The coupling equations of radial velocity, east-west wind component, and north-south wind component can be established through trigonometric functions, which is the theoretical basis for VAD wind field inversion.

[0055] Here, the east-west wind speed component is one of the orthogonal decomposition components of a two-dimensional horizontal wind field, specifically referring to the velocity of atmospheric air particles moving along the east-west direction. Eastward is defined as positive and westward as negative. It is a core physical quantity characterizing the horizontal zonal motion of the horizontal wind field, and together with the north-south wind speed component, it constitutes a complete two-dimensional horizontal wind field.

[0056] Here, the north-south wind speed component is another orthogonal component of the two-dimensional horizontal wind field, specifically referring to the velocity of atmospheric air particles moving along the north-south direction. Northward is defined as positive and southward as negative, and it is orthogonal to the east-west wind speed component, making it an essential parameter for completely reconstructing the regional horizontal wind field structure.

[0057] In this embodiment, preset key parameters are extracted based on the radar-specific encoding format, and radial velocity data is reconstructed using a dedicated encoding restoration formula. This method overcomes the shortcomings of traditional general decoding methods that ignore radar encoding differences and directly coarseen the decoding. It can accurately adapt to the storage rules of different radar models, effectively avoiding problems such as original data packet loss, code value conversion distortion, and parameter missingness. It ensures the authenticity, integrity, and validity of radial velocity data from the data source, providing a reliable raw data foundation for subsequent wind field inversion. In addition, an equation system is constructed based on the physical relationship of the wind field within a local grid window. Combined with the least squares method to optimally solve for the wind velocity components, it can adapt to the local wind field heterogeneity of complex underlying surfaces in mountainous areas, effectively reducing the interference of observational random errors, accurately restoring the small-scale wind field structure, and significantly improving the refinement of wind field inversion under complex mountainous terrain.

[0058] In some embodiments, the multimodal large language model identifies severe convective features and outputs structured severe convective weather forecast results, including: automatically identifying bow echoes, gust fronts, mesocyclones, velocity pairs, and three-body scattering features in images using the multimodal large language model to determine the potential location of severe convective weather; and outputting structured data in JSON format containing a wind gap area field, wherein the wind gap area field contains a list of township names whose wind speeds reach a preset threshold, and data on the maximum wind speed range containing a gust intensity field.

[0059] In some embodiments, generating a township-level risk warning list includes: for lightning warnings, selecting strong echo points with a radar combined reflectivity of not less than 65 dBZ as lightning risk points; for heavy precipitation warnings, statistically analyzing grid points in the refined precipitation forecast map where precipitation intensity exceeds 50 mm / h as heavy precipitation risk points; for thunderstorm and gale warnings, extracting grid points with wind speeds exceeding a preset gale threshold from the VAD wind field inversion map, and verifying them based on the velocity fuzzy features in the radial velocity data, and then using these as thunderstorm and gale risk points; using a geospatial connectivity algorithm, performing spatial correlation analysis between the lightning risk points, the heavy precipitation risk points, and the thunderstorm and gale risk points and the township administrative division boundary vector data respectively, determining the township area to which each risk point falls, and classifying and statistically analyzing them according to risk type to generate the township-level risk warning list.

[0060] In some embodiments, the method further includes: configuring the main control program to cyclically call the independent executable scripts corresponding to the above steps at preset time intervals to achieve fully automated operation; the preset time interval is 10 minutes; automatically storing the generated refined precipitation forecast map, the VAD wind field inversion map, the severe convection forecast results, and the township-level risk warning list to a specified directory according to the generation date, and generating a dynamically refreshed webpage index file to achieve unattended operational operation.

[0061] The following will describe an exemplary application of the embodiments of the present invention in a practical application scenario.

[0062] This embodiment utilizes four core technologies—multi-data fusion support, topographic precipitation forecasting technology, radar wind field VAD inversion technology, and AI thunderstorm and gale forecasting technology—to achieve a fully automated nowcasting method. The specific steps are as follows: (I) Multi-data fusion support system Acquire FY-4B meteorological satellite TBB data (i.e., meteorological satellite observation data in the above embodiments), radar combined reflectivity map and radar base data, GRIB2 format numerical model wind field data (including two component files uwin.grb2 and vwin.grb2) and specific humidity data, topographic data (file name: terrain_data.txt), and administrative division boundary vector data of townships under the jurisdiction of the prefecture-level city (xz_hss.shp).

[0063] Using methods such as CUBR interpolation, the diverse multi-source data with varying origins and formats are uniformly transformed and resampled to a high-precision regular grid with a spatial resolution of 0.01° × 0.01°. This processing step ensures that data from different sources, such as satellites, radar, numerical models, and geographic information systems, have a completely consistent spatial reference and coordinate framework in subsequent calculations and analyses.

[0064] (II) Orographic Precipitation Forecasting Technology The calculation process of terrain vertical velocity: For short-term heavy precipitation forecasting, the amplification effect of terrain needs to be considered in mountainous areas. Among these factors, the calculation of terrain vertical velocity is the primary focus. This is based on wind field data provided by the model (including surface zonal wind speed u and radial wind speed v) and the spatial gradient of the mountainous terrain (i.e., the rate of change of terrain height in the x and y directions). In the formula, h is the terrain height; This represents the topographic gradient along the longitude direction; This represents the terrain gradient along the latitudinal direction. (Using the formula...) Quantitatively assess the vertical motion velocity caused by topographic forced uplift, where, Vertical velocity relative to the terrain; Ground density; It is the acceleration due to gravity; This is the horizontal wind speed vector; The result represents the spatial gradient of terrain height, expressed in hPa / s, reflecting the dynamic influence of terrain on airflow.

[0065] The method for estimating the topographic precipitation rate is as follows: Based on the obtained topographic vertical velocity, combined with the atmospheric specific humidity q (i.e., the specific humidity data in the above example), the precipitation condensation rate is used. The expression, Further calculations can be made to estimate the precipitation intensity caused by orographic uplift, i.e., the increase in rainfall. , In the formula, For a specific time period; the unit of the calculation result is mm / h, used to quantify the direct contribution of topography to precipitation.

[0066] Criteria for determining convergence zones: To ensure the reasonableness of orographic precipitation estimation, convergence zones are defined only when the horizontal divergence... The calculation result is less than 0, meaning that the contribution of orographic precipitation is only retained when the corresponding atmospheric horizontal convergence area is reached. This condition effectively eliminates false precipitation signals that may be generated in non-convergence areas, improving the accuracy of the forecast. Finally, the filtered orographic precipitation data will be fused with the precipitation field obtained by converting radar forecast reflectivity, and a refined precipitation forecast map for 0-1 hours will be generated using the dedicated program rainfall_01h_yb.py.

[0067] (III) Application of Radar Wind Field VAD Inversion Technology Radial velocity decoding steps: First, extract the two key parameters Scale and Offset from the radar base data (v9559). Then, accurately restore the radial velocity of each range library according to the formula "velocity (m / s) = (encoded value - Offset) / Scale", and convert the encoded data into the actual physical velocity value.

[0068] Gridded radial velocity processing: Next, the radial velocity data in the polar coordinate system is transformed and interpolated to regular latitude and longitude grid points through the distance-azimuth projection method. This process is implemented by the script v_cref.py, thereby obtaining a spatially continuously distributed gridded radial velocity field.

[0069] VAD wind field inversion calculation: Within each 3×3 local grid window, based on the physical relationship between radial velocity and horizontal wind field, the core is: in the Cartesian coordinate system, the horizontal wind vector is... Represents components, i.e. In polar coordinates, then... , Represents components, i.e. In polar coordinates ( , radial velocity in ) and tangential velocity They are respectively , In the formula This is the azimuth angle, due north. clockwise Increase. right Find the derivative, assuming the area element being calculated is ( , The internal wind field is uniform, with , ,in express along Increase in the direction, that is... Clockwise is positive, counterclockwise is negative. A value greater than 0 represents the radial velocity away from the radar, and a value opposite represents the radial velocity towards the radar. This decomposition results in... When using portions, combine Azimuth has 4 quadrants and a total of 16 modes, for example... , , The least squares method is used to solve this system of equations, and the two components of the horizontal wind field are derived: , The wind direction is defined by u (east-west wind) and v (north-south wind). The inversion results are then Gaussian smoothed to generate a more continuous and reasonable high-resolution two-dimensional wind field. This core algorithm is implemented by the v_fy.py script.

[0070] Wind plume plotting and visualization: The high-resolution two-dimensional wind field obtained from the above inversion was downsampled to a spatial resolution of 0.1°, and wind plume plots were drawn based on this. The township administrative boundaries of Huangshan City were superimposed on the plot as a geographical reference, thereby intuitively and clearly displaying the mid-level wind field structure and its distribution characteristics within the study area.

[0071] (iv) AI-based thunderstorm, gale, and precipitation forecasting technology This embodiment innovatively applies a multimodal large language model to the nowcasting of two types of severe convective weather: thunderstorms and strong winds, and short-duration heavy precipitation, achieving full automation from image feature recognition to township-level risk area output. This technology constructs targeted image input combinations and analysis strategies based on the different forecast targets.

[0072] 1. AI-powered thunderstorm and gale forecasting sub-technology (1) Multimodal feature recognition and processing stage: The system uses multimodal image sequences from radar combined reflectivity maps, VAD wind field inversion maps, radial velocity data, and meteorological satellite observation data as core input data. By calling the Tongyi Qianwen multimodal large language model API interface, the system can deeply integrate and comprehensively analyze these image contents, thereby extracting the rich meteorological information contained therein, laying a data foundation for subsequent weather feature recognition.

[0073] (2) Automatic model analysis and identification stage: The multimodal large language model autonomously and deeply analyzes the input images, accurately identifying the presence of various typical severe convective features such as bow echoes, gust fronts, mesocyclones, velocity pairs, and three-body scattering. The model not only identifies the features themselves but also combines geographic and topographical information to comprehensively assess the intensity, location, and evolution trend of various features, thereby evaluating the probability and risk level of future severe convective weather such as thunderstorms and strong winds.

[0074] (3) Structured result output stage: After analysis, the model outputs structured forecast results in a standardized JSON format, namely, severe convection forecast results. These results mainly include two core fields: first, the wind gap area, which lists the names of all towns and villages where thunderstorms with wind speeds reaching or exceeding 17.2 m / s (i.e., wind force 8 or above) are likely to occur. This list strictly matches the "name_x1" field name in the user-provided SHP file to ensure accurate spatial correspondence; second, the gust intensity, which provides the expected maximum wind speed range for each affected area. The values ​​are presented in a clear and concise format, such as "17.2-25 m / s," providing specific intensity references for forecasting decisions.

[0075] (4) Generation stage of pictures, reasons and explanatory text: The model simultaneously generates a detailed text description, systematically explaining the current weather situation, detailing the key features identified from radar and satellite images and their meteorological significance, and clearly providing the basis and reasoning logic for the final forecast. Furthermore, the AI-predicted locations of thunderstorms and strong winds in townships are presented visually: on the corresponding SHP map, the system highlights townships likely to experience severe weather in bold red, clearly labeling the township names at their corresponding locations, thus achieving an intuitive and accurate spatial representation of the forecast conclusions.

[0076] 2. AI Precipitation Forecasting Sub-technology (1) Modal image sequence construction stage: The system automatically retrieves three key image types from the latest data catalog: earlier combined reflectance (CREF) images, more recent combined reflectance (CREF) images, and infrared cloud images (TBB) from the FY-4B satellite. These three images constitute a sequence of images reflecting the dynamic evolution of the precipitation system from past to present, providing the AI ​​model with temporally continuous evolution information.

[0077] (2) Automatic model analysis and identification stage: The system calls the Tongyi Qianwen Multimodal Large Language Model API, taking the above three images as input, and supplementing them with precise time difference information (the time interval between two CREF images, and the time difference between the contour image and the latest CREF image) to guide the model to perform in-depth analysis: Motion vector inference: By comparing two CREF maps, the model automatically identifies the direction of movement, speed of movement (quantitatively calculated using a given time difference), and intensity change trend (increasing or decreasing) of heavy precipitation echo clusters.

[0078] Cloud image feature fusion: By combining cloud top brightness temperature (TBB) data from infrared cloud images, the vertical development intensity, cloud top height, and cold cloud area of ​​convective clouds can be determined, which helps to confirm the maturity and potential precipitation efficiency of precipitation systems.

[0079] Integrated forecast inference: The model combines echo extrapolation trends and cloud map characteristics to form a comprehensive judgment on the precipitation situation in the study area for the next hour.

[0080] (3) Structured result output stage: After the analysis is completed, the model outputs structured precipitation forecast results in a standardized JSON format.

[0081] (4) Precipitation field generation and visualization stage: Based on the structured forecasts output by AI, the system generates a high-resolution precipitation field using the "township boundary constraint" method: for township areas in the heavy_area and light_area, the median of their forecast intensity range is assigned; for unmentioned townships, a precipitation value of 0 is assigned. This method ensures the spatial physical consistency and clear boundaries of the precipitation forecast. Finally, this precipitation field, along with satellite cloud images and radar reflectivity maps, forms a 2×2 mosaic map, visually demonstrating the forecast basis and conclusions. Simultaneously, the system automatically generates a text file containing the rationale for the precipitation forecast and a JSON-formatted forecast data file for operational personnel to reference and archive.

[0082] Through the two parallel AI forecasting sub-techniques of thunderstorms and strong winds and precipitation, automated, intelligent, and township-level nowcasting of two major types of severe convective weather (thunderstorms and strong winds and short-term heavy precipitation) has been achieved, significantly improving the timeliness and precision of the early warning.

[0083] (v) Technology for displaying the list and images of high-risk townships 1. Spatial correlation analysis: The lightning warning process involves screening strong echo points with a combined reflectivity of not less than 65 dBZ, then using spatial connectivity (via the gpd.sjoin method) to match the echo points with township polygons to determine the townships affected by lightning, and automatically generating the corresponding list. This function is implemented by the cr_ldyb.py script.

[0084] The heavy rainfall early warning mechanism is based on the grid data of fine precipitation forecast for the next 0 to 1 hour. It counts the number of grid points within the administrative area of ​​each township that have reached or exceeded the preset threshold (≥50 mm / hour). When the number of such high precipitation intensity grid points exceeds the preset minimum number (e.g., 4), the system determines that there is a risk of short-term heavy rainfall in the township and triggers the early warning process. The relevant data processing and judgment are completed by the script rainfall_01h_yb.py.

[0085] Thunderstorm and gale warnings rely directly on the gust_area list output by the artificial intelligence model, which allows the system to quickly identify and delineate risk areas.

[0086] 2. Image captions for meteorological products: For different types of images such as reflectance maps, precipitation maps, and wind field maps, the system calculates the centroid of the polygon geometry of each township (specifically obtained using the representative_point() method), clearly marks the corresponding township name in black or blue font at the corresponding coordinate position on the map, and adds a white semi-transparent background frame to each mark to ensure that the text remains clearly legible against various complex background map backgrounds.

[0087] 3. Text report output: The system automatically generates several standardized text files, including *_lightning_towns.txt, *_rainfall_01h_towns.txt, and *_gust_towns.txt. These files provide a detailed list of towns at risk of lightning, short-term heavy rainfall, and thunderstorms, offering direct and clear data references and action guidelines for relevant decision-making departments.

[0088] (vi) Integration of full-process automation and visualization All the above steps are completed by the main control program sequentially calling each independent executable file. The entire process is automatically executed every 10 minutes, ensuring continuous and stable task operation. All generated product images (including the township information marked on them) and corresponding text reports are automatically categorized and stored in designated subdirectories according to their generation date (in YYYYMMDD format) for easy subsequent management and retrieval. In addition, the system dynamically generates an index.html page, which allows users to filter and view based on product type, time, and other conditions, and automatically refreshes every 5 minutes, thus achieving a truly unattended automated business operation mode.

[0089] In this embodiment, regarding the accuracy of topographic precipitation forecasts, by introducing convergence discrimination technology and considering the amplification effect of mesoscale topography on precipitation, the false alarm rate of nowcast precipitation in mountainous areas is significantly reduced by about 30%, while the TS score of 0-1 hour nowcast precipitation is improved by 0.12, effectively enhancing the forecast accuracy and reliability.

[0090] In terms of timeliness and accuracy of thunderstorm and gale warnings and short-term heavy precipitation warnings, the AI ​​model demonstrated excellent performance: the accuracy rate for identifying severe convective features such as bow echoes, gust fronts, velocity pairs, three-body scattering, and mesocyclones remained consistently above 85%; in terms of precipitation forecasting, the model could accurately infer the evolution trend of echo movement speed and intensity, and the township-level precipitation area forecasts generated by combining satellite cloud image features showed a high degree of agreement with the actual situation. From acquiring radar data to outputting township-level thunderstorm and gale area text and precipitation forecast products, the entire process takes only 1-3 minutes, which is 10-15 minutes earlier than the previous manual identification and subjective extrapolation methods, significantly enhancing the timeliness of severe convective weather warnings.

[0091] The application value of wind field inversion products has become even more apparent. The high-resolution wind plume images generated by the inversion clearly demonstrate the fine structure of low-level wind field convergence lines and their close correspondence with echo-forming regions. This provides forecasters with an important reference for intuitively and accurately judging the occurrence, development, and evolution trends of convective systems, effectively improving the accuracy of short-term forecast services and user satisfaction.

[0092] The early warning products have reached the township level in terms of precision. Each forecast can automatically generate a list of townships at risk, covering three types of disasters: lightning, heavy rainfall, and thunderstorms with strong winds. The township names are clearly marked on the generated early warning images, which can directly serve the government and grassroots disaster prevention personnel in townships, effectively achieving the goal of "early warnings reaching the township level."

[0093] The fully automated process significantly saves human resources. The system can run automatically 144 times a day without human intervention, completing a full forecast process every 10 minutes. Forecasters no longer need to manually process data and create charts, allowing them to focus more on forecast consultations and decision-making support, resulting in a substantial improvement in overall work efficiency.

[0094] The above description is merely an embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and scope of the present invention are included within the scope of protection of the present invention.

[0095] It should be understood that the phrase "one embodiment" or "an embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of the invention. Therefore, "in one embodiment" or "in an embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. It should be understood that in the various embodiments of the invention, the sequence numbers of the above-described processes do not imply a sequential order of execution; the execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the invention. The sequence numbers of the above-described embodiments of the invention are merely descriptive and do not represent the superiority or inferiority of the embodiments.

[0096] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, or apparatus. Without further limitations, an element defined by the phrase "comprising a..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. In the several embodiments provided by this invention, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods, such as: multiple units or components may be combined, or integrated into another system, or some features may be ignored or not performed.

[0097] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for applying AI technology to near-term weather forecasting in mountainous areas, characterized in that, The method includes: Multi-source data for predicting mountainous areas is acquired, and the multi-source data is preprocessed to obtain spatiotemporally matched multi-source gridded data. The multi-source data includes meteorological satellite observation data, radar base data and radar combined reflectivity map, numerical model wind field data, specific humidity data, topographic data and administrative division boundary vector data. Based on the numerical model wind field data, the specific humidity data, and the topographic data, the topographic vertical velocity is calculated; based on the topographic vertical velocity and the specific humidity data, combined with the horizontal divergence discrimination condition, the topographic precipitation contribution is calculated, and the topographic precipitation contribution is fused with the precipitation field obtained based on the radar combined reflectivity image conversion to generate a refined precipitation forecast map. Radial velocity data is decoded from the radar base data, projected onto a latitude and longitude grid, and then inverted and plotted based on the VAD algorithm to generate a VAD wind field inversion map. A multimodal image sequence comprising the radar combined reflectivity map, the VAD wind field inversion map, the radial velocity data, and the meteorological satellite observation data is constructed and input into a pre-trained multimodal large language model; the multimodal large language model identifies strong convection features and outputs structured strong convection forecast results; Based on the refined precipitation forecast map and the severe convection forecast results, areas with precipitation intensity exceeding the standard and severe convection feature points are extracted as meteorological risk points. The meteorological risk points are matched with the administrative boundaries of townships using a spatial connectivity algorithm to generate a township-level risk warning list, which is then overlaid onto the geographic information base map to generate a visualized warning image.

2. The method according to claim 1, characterized in that, Calculating the vertical velocity of the terrain specifically includes: Based on the horizontal wind speed component in the numerical model wind field data and the spatial gradient of terrain height determined based on the terrain data, the terrain vertical velocity is calculated; the formula for calculating the terrain vertical velocity is: ； In the formula, Vertical velocity relative to the terrain; Ground density; It is the acceleration due to gravity; This is the horizontal wind speed vector; This represents the spatial gradient of terrain elevation. The specific criteria for determining horizontal divergence include: retaining only horizontal divergence. Regional orographic precipitation contribution, eliminating spurious precipitation signals outside convergence zones; where, horizontal divergence u and v represent the zonal and radial wind speeds at ground level, respectively; x and y represent the spatial distances in the zonal and longitudinal directions, respectively.

3. The method according to claim 1, characterized in that, Generating a VAD wind field inversion map specifically includes: Preset key parameters are extracted from the radar base data based on the radar base data encoding format, and the radial velocity data is restored according to the corresponding encoding restoration formula; Within a local grid window, based on the physical relationship between the radial velocity data and the horizontal wind field in polar coordinates, the least squares method is used to solve the equations and invert the east-west wind speed components and north-south wind speed components. The east-west wind speed component and the north-south wind speed component are Gaussian smoothed, and a high-resolution two-dimensional wind field is synthesized based on the smoothed wind speed components. The VAD wind field inversion map is obtained by overlaying the administrative division boundary vector data onto the high-resolution two-dimensional wind field.

4. The method according to claim 1, characterized in that, The multimodal large language model identifies strong convection features and outputs structured strong convection forecast results, including: The multimodal large language model automatically identifies bow echoes, gust fronts, mesocyclones, velocity pairs, and three-body scattering features in images to determine the potential location of severe convective weather. The output is JSON-formatted structured data containing a wind gap area field, which includes a list of township names where the wind speed reaches a preset threshold, and data on the maximum wind speed range containing a gust intensity field.

5. The method according to claim 1, characterized in that, Generate a township-level risk warning list, including: For lightning warnings, strong echo points with a radar composite reflectivity of not less than 65dBZ are selected as lightning risk points. For heavy rainfall warnings, grid points in the refined precipitation forecast map with precipitation intensity exceeding 50 mm / h are identified as heavy rainfall risk points. For thunderstorm and gale warnings, grid points with wind speeds exceeding a preset gale threshold are extracted from the VAD wind field inversion map and verified based on the velocity fuzzy features in the radial velocity data, and then used as thunderstorm and gale risk points. Using a geospatial connectivity algorithm, the lightning risk points, heavy precipitation risk points, and thunderstorm and gale risk points are spatially correlated with the township administrative division boundary vector data to determine the township area to which each risk point falls. The risk points are then categorized and statistically analyzed according to their risk type to generate the township-level risk warning list.

6. The method according to claim 1, characterized in that, The method further includes: The main control program is configured to cyclically call the independent executable scripts corresponding to the above steps at preset time intervals to achieve fully automated operation of the entire process; the preset time interval is 10 minutes. The generated refined precipitation forecast map, VAD wind field inversion map, severe convection forecast results, and township-level risk warning list are automatically stored in a designated directory according to the generation date, and a dynamically refreshed webpage index file is generated to achieve unattended operational operation.

Images