High plateau airport severe convective weather prediction method based on numerical simulation and artificial intelligence
By combining the WRF-LES tool with artificial intelligence models, the problem of low accuracy in forecasting severe convective weather in high-altitude plateau regions has been solved, achieving high spatiotemporal resolution and real-time severe convective weather forecasting, thus ensuring aviation safety.
Patent Information
- Application Number
- CN202511731490.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-24
- Publication Date
- 2026-02-10
AI Technical Summary
Existing meteorological forecasting methods struggle to capture the fine details of local weather systems in high-altitude plateau regions due to sparse observation stations and insufficient data coverage. Traditional numerical simulation and observation methods are limited by resolution and physical parameterization schemes, making it difficult to accurately simulate the impact of complex terrain on airflow and convective activity. There is a lack of models that can comprehensively describe the severe convective environment field, and there is a lack of physical mechanism support, resulting in low accuracy in severe convective weather forecasts, especially in the prediction of extreme weather such as thunderstorms, strong winds, hail, and short-duration heavy precipitation.
By employing the WRF-LES tool for nested grid settings, combined with high-resolution terrain data and data assimilation methods, a multidimensional meteorological dataset is constructed. An artificial intelligence prediction model is used for feature extraction and training, and a loss function with physical constraints is incorporated to achieve high spatiotemporal resolution and real-time prediction of severe convective weather.
It has enabled refined flow field simulation and accurate short-term forecasting of severe convective weather at high-altitude airports, improved the accuracy of short-term forecasting, provided real-time warnings of severe convective weather, and ensured aviation safety.
Smart Images
Figure CN121503281A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of meteorological forecasting technology, and more specifically, relates to a method for predicting severe convective weather at high-altitude airports based on numerical simulation and artificial intelligence. Background Technology
[0002] Severe convective weather can trigger severe turbulence, icing, lightning strikes, and wind shear, all of which endanger flight safety. This is especially true at airports in high-altitude regions (above 3000 meters), where complex terrain and extreme weather conditions further exacerbate the unpredictability and hazards of severe convective weather. Airflow in high-altitude regions is significantly influenced by topographic forcing, and the dramatic changes in local circulation, turbulence, and thermal and dynamic conditions make the occurrence and development of severe convective weather more complex and variable.
[0003] Existing weather forecasting methods mainly rely on traditional meteorological observation methods (such as ground stations, radiosondes, radar, and satellites) and numerical weather prediction (NWP) models, which have the following drawbacks: (1) In high-altitude plateau regions, traditional methods are difficult to capture the fine features of local weather systems due to the sparse observation stations and insufficient data coverage; (2) The application of traditional numerical simulation and observation methods in high-altitude plateau areas is limited by resolution and physical parameterization schemes, making it difficult to accurately simulate the impact of complex terrain on airflow and convection, especially in the short-term (1-6 hours) forecast period; (3) Existing technologies have not fully utilized the potential of machine learning algorithms in severe convective weather forecasting, lack models that can comprehensively describe the severe convective environment field, and are difficult to effectively solve the problem of forecasting the potential of severe convective weather such as thunderstorms and strong winds, and are difficult to meet the requirements of aviation operations for high spatiotemporal resolution and real-time performance. (4) Existing technologies often lack physical mechanism support for the application of machine learning models in severe convective weather forecasting, and do not effectively integrate weather parameters into machine learning models, resulting in poor model interpretability. (5) The triggering mechanism of severe convective weather in high-altitude plateau areas is complex, involving multi-scale physical processes such as local thermal instability, convection triggering and topographic forcing. Traditional methods are difficult to fully describe the change process, resulting in low forecast accuracy of severe convective weather, especially in the prediction of extreme weather such as thunderstorms, hail and short-term heavy precipitation. Summary of the Invention
[0004] Severe convective weather can trigger severe turbulence, icing, lightning strikes, and wind shear, all of which endanger flight safety. This is especially true at airports in high-altitude regions (above 3000 meters), where complex terrain and extreme weather conditions further exacerbate the unpredictability and hazards of severe convective weather. Airflow in high-altitude regions is significantly influenced by topographic forcing, and the dramatic changes in local circulation, turbulence, and thermal and dynamic conditions make the occurrence and development of severe convective weather more complex and variable.
[0005] Existing weather forecasting methods mainly rely on traditional meteorological observation methods (such as ground stations, radiosondes, radar, and satellites) and numerical weather prediction (NWP) models, which have the following drawbacks: (1) In high-altitude plateau regions, traditional methods are difficult to capture the fine features of local weather systems due to the sparse observation stations and insufficient data coverage; (2) The application of traditional numerical simulation and observation methods in high-altitude plateau areas is limited by resolution and physical parameterization schemes, making it difficult to accurately simulate the impact of complex terrain on airflow and convection, especially in the short-term (1-6 hours) forecast period; (3) Existing technologies have not fully utilized the potential of machine learning algorithms in severe convective weather forecasting, lack models that can comprehensively describe the severe convective environment field, and are difficult to effectively solve the problem of forecasting the potential of severe convective weather such as thunderstorms and strong winds, and are difficult to meet the requirements of aviation operations for high spatiotemporal resolution and real-time performance. (4) Existing technologies often lack physical mechanism support for the application of machine learning models in severe convective weather forecasting, and do not effectively integrate weather parameters into machine learning models, resulting in poor model interpretability. (5) The triggering mechanism of severe convective weather in high-altitude plateau areas is complex, involving multi-scale physical processes such as local thermal instability, convection triggering and topographic forcing. Traditional methods are difficult to fully describe the change process, resulting in low forecast accuracy of severe convective weather, especially in the prediction of extreme weather such as thunderstorms, hail and short-term heavy precipitation. Attached Figure Description
[0006] Figure 1 This is a schematic diagram illustrating the principle of the high-altitude airport severe convective weather prediction method based on numerical simulation and artificial intelligence provided by the present invention. Detailed Implementation
[0007] 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, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.
[0008] As an example, see the attached document. Figure 1 As shown, to solve the above-mentioned technical problems, this embodiment provides a method for predicting severe convective weather at high-altitude airports based on numerical simulation and artificial intelligence, including: Load the WRF-LES (Weather Research and Forecasting Large-Eddy Simulation) tool, set nested grids, horizontal resolution, and airport area, enable LES mode, downscale the terrain data, select a physical parameterization scheme, process the terrain data and adapt it to the high-altitude environment, use data assimilation methods to adjust the initial field and boundary conditions, simulate the local flow field in the high-altitude region, output meteorological data for the high-altitude region, and use independent observation data to verify the simulation results and terrain effects of the WRF-LES tool; Key meteorological features were extracted from the simulation results, combined with field observation data, to construct a multidimensional meteorological dataset, and the data was preprocessed. After cleaning, normalizing and spatiotemporally aligning the multidimensional meteorological dataset, feature extraction is performed, including physical features, observational features, topographic features and spatiotemporal features, and feature data is obtained by filtering. An artificial intelligence prediction model is constructed, taking feature data as input and outputting the occurrence probability of each severe convective weather type, the intensity of severe convective weather, as well as the occurrence location, movement path, and duration of severe convective weather. The model is trained using historical meteorological data and severe convective event data to learn the relationship between the evolution pattern of severe convective weather and numerical simulation results. The model is trained through supervised learning and data augmentation, and the accuracy of the artificial intelligence prediction model is verified using a loss function, thus obtaining a severe convective weather prediction model. The simulation results obtained from the real-time WRF-LES model are input into the severe convective weather prediction model, which outputs the probability of occurrence, intensity, and spatiotemporal distribution of severe convective weather within a set time period. The probability of occurrence threshold and intensity threshold are set, and the output results of the severe convective weather prediction model are used to issue early warnings.
[0009] This invention is a method that effectively combines numerical simulation and artificial intelligence prediction models to improve the accuracy of short-term forecasts of severe convective weather at high-altitude airports. It is applicable to aviation meteorological early warning and flight safety assurance at high-altitude airports.
[0010] Optionally, you can set nested grids, horizontal resolution, and airport area, including: The horizontal resolution and extent of the outer, middle, inner, and LES regions are set, along with the mesh nesting ratio, vertical resolution, and time step. Vertical resolution includes the number of vertical layers and the model top height. The time step includes setting the time step for the outer mesh and the time step for the LES region. For example: the outer region has a horizontal resolution of 9 km, covering a large area (5000 km × 5000 km), used to simulate large-scale weather background fields (such as upper-level jet streams and frontal systems); the middle region has a horizontal resolution of 3 km, covering the high-altitude airport and surrounding area (1000 km × 1000 km), used to simulate mesoscale convection systems and topographic forcing effects; the inner region has a horizontal resolution of 1 km, covering the target airport and surrounding complex terrain area (300 km × 300 km), used to simulate local convection and topographically induced turbulence; the LES region has a horizontal resolution of 100 meters, covering the core area of the airport (50 km × 50 km), used to capture small-scale turbulence, local convergence, and convection-triggered processes. A 3:1 mesh nesting ratio is used to ensure numerical stability. The horizontal resolution of 100 meters in the LES region was achieved through downscaling. Vertical resolution settings: The number of vertical layers was set to 50-70, with the near-surface layer (0-1km from the ground) refined to approximately 20-50 meters per layer to capture low-level wind shear and surface thermal processes; the model top height was set to 50 hPa (approximately 20km), covering the troposphere and lower stratosphere to ensure simulation of thunderstorm cloud tops and upper-level wind fields; Time step settings: The time step for the outer grid (9 km) was 30 seconds, decreasing layer by layer; the time step for the LES region was 0.1-0.3 seconds to meet the numerical stability requirements of high-resolution simulation.
[0011] The terrain data is downscaled, specifically to a horizontal resolution of 100 meters, using the following steps: (1) WRF nested downscaling: Based on the inner region, the LES region mesh is further set by the nesting function of the WRF tool. The LES region uses one-way or two-way nesting and obtains the boundary conditions from D03 to ensure that the boundary conditions of the LES region are smoothly transitioned and to avoid the boundary effect from interfering with the simulation results.
[0012] (2) LES mode activation: The LES mode is activated in the LES region, the traditional boundary layer parameterization scheme is turned off, and the three-dimensional turbulence closure scheme (Deardorff TKE scheme) is adopted to directly analyze small-scale turbulent motion; the LES mode simulates key processes such as local turbulence, updraft, and local convergence through high-resolution grid (100 m) and fine time step. (3) Downscaling data processing: The WRF tool is used to interpolate the initial field and boundary conditions of the LES region from the 1 km output of the inner region. Combined with high-resolution terrain data (such as ASTER 30m terrain dataset), interpolation is performed to generate information such as terrain height, slope and land use type at a resolution of 100 meters, ensuring that the LES region accurately reflects the terrain details.
[0013] The physical parameterization schemes include setting up microphysics schemes, boundary layer schemes, radiation schemes, surface process simulation schemes, and convection parameterization schemes.
[0014] Optionally, the microphysics scheme employs the Morrison two-parameter scheme; the radiation scheme employs the RRTMG long-wave and short-wave schemes; the surface processes scheme uses the Noah-MP land surface model to simulate soil moisture, surface temperature, and heat flux in high-altitude plateau regions, and combines high-resolution land use data to optimize surface roughness and albedo parameters; the convection parameterization scheme in the outer and middle layers simulates mesoscale convection systems; and the convection parameterization is disabled in the inner layer and LES regions.
[0015] Specifically, considering the complex terrain and meteorological conditions in high-altitude plateau regions, the physical parameterization scheme of WRF is optimized to ensure that the model can accurately simulate the triggering and development of severe convective weather. For example: Microphysics Scheme: The Morrison two-parameter microphysics scheme is selected, which can accurately simulate microphysical processes such as cloud droplets, ice crystals, and raindrops, and supports the simulation of thunderstorms, hail, and heavy precipitation. For high-altitude, low-temperature environments, ice phase process parameters (such as ice crystal settling velocity and ice nucleus concentration) are optimized to improve the accuracy of icing and hail simulations. Boundary layer scheme: The YSU (Yonsei University) boundary layer scheme is used in the outer, middle, and inner regions of the outer mesh to simulate large-scale boundary layer turbulence. In the LES region (D04), traditional boundary layer parameterization is disabled, and the LES turbulence closure scheme (1.5th order turbulent kinetic closure) is adopted to directly analyze small-scale turbulent motion. Radiation scheme: The longwave and shortwave radiation schemes of RRTMG (Rapid Radiative Transfer Model for GCMs) are adopted, taking into account the effects of strong solar radiation and cloud-radiation interaction in high-altitude plateau areas; Surface processes: The Noah-MP (Multi-Physics) land surface model was used to simulate soil moisture, surface temperature and heat flux in high-altitude plateau regions, accurately reflecting the impact of surface thermal processes on convection triggering. Combined with high-resolution land use data (MODIS 30 m dataset), surface roughness and albedo parameters were optimized.
[0016] Convection parameterization: Enable the Kane-Fridge or Greer-Freitas convection parameterization scheme in the outer and middle regions to simulate mesoscale convective systems; disable convection parameterization in the inner grid (D03-D04, 1 km and 100 m) to directly analyze the convection process and ensure a refined simulation of severe convective weather.
[0017] Optionally, the terrain data can be processed, including: using the ASTER GDEM terrain dataset, generating terrain height, slope, and aspect data with a resolution greater than a set threshold through bilinear interpolation; preserving terrain features in the LES region to simulate terrain uplift and local convergence effects; adjusting the terrain smoothing factor; and simulating surface sensible and latent heat fluxes.
[0018] The complex terrain of high-altitude plateau regions is a significant triggering factor for severe convective weather. Therefore, the WRF-LES tool is specifically optimized for terrain processing, for example: High-resolution terrain data: Using the Global Digital Elevation Model (30-meter resolution) terrain dataset, 100-meter resolution terrain height, slope, and aspect data are generated through bilinear interpolation; in the LES region, terrain details (such as ridges, valleys, and steep slopes) are preserved to accurately simulate terrain uplift and local convergence effects; Terrain forcing optimization: Adjust the terrain smoothing factor in WRF to reduce numerical instability and ensure the accuracy of terrain forcing effects. Optimize the surface drag coefficient and terrain height correction parameters for terrain features around high-altitude airports (such as valley winds and slope flow).
[0019] Local thermal processes: Simulating surface sensible heat flux and latent heat flux, considering the influence of strong solar radiation and low air pressure on thermal instability in high-altitude plateau regions, and capturing local thermal instability and moisture convection-triggered processes caused by topography through high-resolution simulation.
[0020] Optionally, when simulating the local flow field in high-altitude plateau regions, a three-dimensional variational assimilation method is adopted, and the assimilation frequency is set. For the outer grid, global model data is assimilated with ground observation data every 6 hours; for the outer grid, high-frequency observation data is assimilated every 1 hour.
[0021] Specifically, to improve the accuracy of WRF-LES simulations, data assimilation techniques are used to optimize the initial field and boundary conditions: Data assimilation method: Using the WRFDA (WRF Data Assimilation) module, a three-dimensional variational assimilation (3DVAR) method is employed to fuse ground observation data (including airport self-observation data and encrypted automatic station observation data). Assimilation frequency: For the outer and middle grids, global model data (such as GFS 0.25° resolution) is assimilated every 6 hours; for the inner grid and LES region, high-frequency observation data (such as radar and satellite data) is assimilated every 1 hour to improve the accuracy of the initial field for short-term predictions.
[0022] Optionally, the output meteorological data for high-altitude plateau regions includes output variables, diagnostic variables, and resolution data; output variables include basic meteorological elements and convection parameters; diagnostic variables include echo top height, vertical cumulative liquid water content, surface sensible heat flux, and local convergence; resolution data includes temporal resolution and spatial resolution for the LES region.
[0023] Optionally, key meteorological feature data can be extracted from the simulation results, combined with field observation data, to construct a multidimensional meteorological dataset, and data preprocessing can be performed, including: Key meteorological characteristic data include meteorological element data and diagnostic variables; meteorological element data include three-dimensional wind field, temperature, humidity, air pressure, precipitation, available convective potential energy, convective suppression energy, vertical wind shear, K index and wet potential vortex; diagnostic variables include echo top height, vertical cumulative liquid water content, surface sensible heat flux and local convergence. Field observation data includes Doppler radar data, satellite data, ground observation data, radiosonde data, auxiliary data, historical severe convective event data, and land use type data; Doppler radar data includes reflectivity factor, radial velocity, and spectral width, used to identify severe convective features; severe convective features include thunderstorm cells, bow echoes, and wind shear; satellite data includes cloud top brightness temperature, cloud system movement speed, and convective initiation indicators extracted from high-resolution infrared, water vapor, and visible light channel data acquired using geostationary satellites; ground observation data: real-time observation data collected from high-altitude airports and surrounding meteorological stations, including wind speed, wind direction, and temperature. The data includes information on local meteorological conditions such as humidity and air pressure; radiosonde data, including vertical profile data obtained using upper-air radiosondes; vertical profile data, such as temperature, humidity, and wind field profiles, used to verify and correct the atmospheric vertical structure simulated by WRF-LES; auxiliary data, including high-resolution topographic data, including topographic height, slope, and aspect; historical severe convective event data, including the spatiotemporal distribution and characteristic parameters of severe convective weather events in high-altitude plateau regions; severe convective weather events, including thunderstorms, hail, and short-duration heavy precipitation; and land use type data, used to characterize surface thermal properties and the impact of convection triggering. By using key meteorological feature data and field observation data, a multidimensional meteorological dataset is constructed. Key meteorological feature data that contributes the most to the prediction of severe convective weather are selected, and dimensionality reduction is performed using linear discriminant analysis.
[0024] Specifically, correlation analysis (calculating Pearson correlation coefficient) or mutual information methods are used to screen features that contribute most to severe convective weather prediction, while redundant features or features with correlation below a set threshold are removed. Linear discriminant analysis is used for dimensionality reduction, and the variance explained rate of each LDA component is calculated. When the cumulative variance reaches 90%, the optimal dimensionality reduction dimension is determined to reduce computational complexity. A multidimensional feature dataset is constructed, including physical features, observational features, topographic features, and spatiotemporal features, constituting the input data for high-altitude severe convective weather prediction.
[0025] Optionally, a convolutional neural network model can be used to extract spatial features of meteorological data, including the spatial distribution of thunderstorm cells, cloud morphology, and wind field structure; a long short-term memory network can be used to capture the temporal evolution of meteorological elements, including the development of convective clouds, dynamic changes in wind fields, and changes in precipitation intensity; a spatiotemporal joint model can be used to analyze the spatiotemporal characteristics of meteorological element data; and model parameters can be adjusted to dynamically adjust the contribution of physical features and topographic features.
[0026] This invention trains a severe convective weather prediction model using historical meteorological data and severe convective event data, learning the evolution patterns of severe convective weather and the relationship between numerical simulation results. It incorporates topographic information as a feature input to enhance the model's adaptability to topographic influences and improve prediction accuracy. The model's output includes: classification output, regression output, and spatiotemporal distribution output. Classification outputs include the probability of occurrence of severe convective weather types (such as thunderstorms, hail, short-duration heavy rainfall, and gusts); regression outputs include the intensity of severe convective weather (such as maximum wind speed, precipitation, and thunderstorm intensity); and spatiotemporal distributions include the location, path, and duration of severe convective weather events.
[0027] Deep learning models, such as convolutional neural network models, are used to extract spatial features from meteorological data, such as the spatial distribution of thunderstorm cells, cloud morphology, and wind field structure. Multi-layer convolutional structures (such as 3×3 convolutional kernels), combined with pooling layers and batch normalization, are employed to improve the robustness of feature extraction.
[0028] Long Short-Term Memory (LSTM) networks can be used to capture the temporal evolution patterns of meteorological elements, such as the development of convective clouds, dynamic changes in wind fields, and the evolution of precipitation intensity. Employing a bidirectional LTM network structure can enhance the ability to model time series data.
[0029] A spatiotemporal joint model, combining convolutional neural network (CNN) and long short-term memory (LSTM) networks, is used to construct an encoder-decoder structure for comprehensive analysis of the spatiotemporal characteristics of meteorological elements. For example, after extracting spatial features using a CNN model, the results are input into an LSTM for time series forecasting.
[0030] Machine learning models, such as random forest models, can be used to process high-dimensional feature data. By integrating multiple decision trees, they can improve the model's ability to fit complex nonlinear relationships and its robustness.
[0031] Gradient boosting tree models can enhance the model's sensitivity to key features of severe convective weather through iterative optimization, making them particularly suitable for small sample scenarios.
[0032] Physical constraint embedding: Adding physical constraint terms (such as convection trigger thresholds) to the model's loss function can ensure that the prediction results conform to meteorological laws.
[0033] By designing feature weighting factors, the contributions of physical features (convective available potential energy and wind shear) and topographic features are dynamically adjusted, thereby improving the interpretability of the model.
[0034] When constructing the dataset, historical severe convective event data (including cases of thunderstorms, hail, and short-duration heavy precipitation) were collected in high-altitude plateau regions. Combined with simulation data from the WRF-LES tool, training, validation, and test sets (in a 7:2:1 ratio) were constructed. The dataset contains at least 5 years of historical data, covering various weather scenarios (such as spring thunderstorms and summer heavy precipitation) to ensure the model's generalization ability.
[0035] Optionally, a physical constraint term can be added to the loss function of the artificial intelligence prediction model; the physical constraint term includes the convection trigger threshold.
[0036] Optionally, the accuracy of the AI prediction model can be obtained by calculating the cross-entropy loss function; the hyperparameters of the AI prediction model can be optimized using cross-validation.
[0037] Training strategies for AI prediction models include: employing supervised learning methods, training the model based on labeled severe convective event data (such as radar-identified thunderstorm cells), or using cross-validation (such as 5-fold cross-validation) to optimize hyperparameters (such as learning rate, number of hidden layers, and tree depth) to prevent overfitting; or introducing data augmentation parameters (such as random perturbations, time series flipping, and spatial shifting) to enhance the model's adaptability to sparse data. Optimizing the prediction accuracy of severe convective weather types by calculating cross-entropy loss, or optimizing the stability of intensity prediction by using mean squared error (MSE) or Huber loss functions, or adding regularization terms (such as L2 regularization) to control model complexity and prevent overfitting.
[0038] Based on a trained artificial intelligence prediction model and combined with the simulation results of the latest WRF-LES tool, short-term (hourly) predictions of severe convective weather at high-altitude airports can be achieved. When the model predicts the occurrence of severe convective weather, it automatically triggers the early warning system to notify the airport control center and pilots in advance so that timely flight adjustments can be made.
[0039] The present invention has the following advantages: This invention enables refined flow field simulation: using the WRF-LES method, it can accurately simulate the flow field below 100 meters in local areas, capture micro-meteorological processes in the complex meteorological environment of high-altitude plateau regions, and provide more refined basic data for the prediction of severe convective weather; This invention combines numerical simulation with artificial intelligence algorithms. By integrating WRF-LES simulation with deep learning algorithms, it can more accurately simulate the complex meteorological environment of high-altitude airports and capture the details and dynamic changes of severe convective weather. This invention enables high-precision short-term forecasting: traditional numerical weather prediction is difficult to achieve short-term (hourly) severe convective weather forecasting, while this invention effectively improves the accuracy of short-term forecasting through artificial intelligence algorithms. This invention combines topographic information to improve upon the shortcomings of traditional numerical simulation methods, making the prediction of severe convective weather more adaptable to the special environment of high-altitude plateau regions. This invention enables real-time forecasting of severe convective weather. By combining the latest results of numerical simulations and artificial intelligence predictions, it provides rapid and accurate short-term warnings of severe convective weather, ensuring aviation safety.
[0040] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for predicting severe convective weather at high-altitude airports based on numerical simulation and artificial intelligence, characterized in that: include: Load the WRF-LES tool, set nested grids, horizontal resolution and airport area, enable LES mode, downscale the terrain data, select physical parameterization scheme, process the terrain data and adapt it to the high-altitude environment, use data assimilation method to adjust initial field and boundary conditions, simulate local flow field in high-altitude region, output meteorological data of high-altitude region, and use independent observation data to verify the simulation results of WRF-LES tool and terrain effect; Key meteorological features were extracted from the simulation results, combined with field observation data, to construct a multidimensional meteorological dataset, and the data was preprocessed. After cleaning, normalizing and spatiotemporally aligning the multidimensional meteorological dataset, feature extraction is performed, including physical features, observational features, topographic features and spatiotemporal features, and feature data is obtained by filtering. An artificial intelligence prediction model is constructed, taking feature data as input and outputting the occurrence probability of each severe convective weather type, the intensity of severe convective weather, as well as the occurrence location, movement path, and duration of severe convective weather. The model is trained using historical meteorological data and severe convective event data to learn the relationship between the evolution pattern of severe convective weather and numerical simulation results. The model is trained through supervised learning and data augmentation, and the accuracy of the artificial intelligence prediction model is verified using a loss function, thus obtaining a severe convective weather prediction model. The simulation results obtained from the real-time WRF-LES model are input into the severe convective weather prediction model, which outputs the probability of occurrence, intensity, and spatiotemporal distribution of severe convective weather within a set time period. The probability of occurrence threshold and intensity threshold are set, and the output results of the severe convective weather prediction model are used to issue early warnings.
2. The method for predicting severe convective weather at high-altitude airports based on numerical simulation and artificial intelligence according to claim 1, characterized in that, Configure nested grids, horizontal resolution, and airport area settings, including: Set the horizontal resolution and region range of the outer, middle, inner, and LES regions; set the mesh nesting ratio, vertical resolution, and time step; vertical resolution includes the number of vertical layers and the top height of the model; setting the time step includes setting the time step of the outer mesh and the time step of the LES region. The physical parameterization schemes include setting up microphysics schemes, boundary layer schemes, radiation schemes, surface process simulation schemes, and convection parameterization schemes.
3. The method for predicting severe convective weather at high-altitude airports based on numerical simulation and artificial intelligence according to claim 2, characterized in that, The microphysics scheme employs the Morrison two-parameter scheme; the radiation scheme uses the RRTMG long-wave and short-wave schemes; the surface processes scheme uses the Noah-MP land surface model to simulate soil moisture, surface temperature, and heat flux in high-altitude plateau regions, and combines high-resolution land use data to optimize surface roughness and albedo parameters; the convection parameterization scheme in the outer and middle layers simulates mesoscale convection systems; and the convection parameterization is turned off in the inner layer and LES region.
4. The method for predicting severe convective weather at high-altitude airports based on numerical simulation and artificial intelligence according to claim 1, characterized in that, The terrain data was processed, including: using the ASTER GDEM terrain dataset, generating terrain height, slope, and aspect data with a resolution greater than a set threshold through bilinear interpolation; preserving terrain features in the LES region to simulate terrain uplift and local convergence effects; adjusting the terrain smoothing factor; and simulating surface sensible and latent heat fluxes.
5. The method for predicting severe convective weather at high-altitude airports based on numerical simulation and artificial intelligence according to claim 1, characterized in that, When simulating the local flow field in high-altitude plateau regions, a three-dimensional variational assimilation method is adopted, and the assimilation frequency is set. For the outer grid, global model data is assimilated with ground observation data every 6 hours; for the outer grid, high-frequency observation data is assimilated every 1 hour.
6. The method for predicting severe convective weather at high-altitude airports based on numerical simulation and artificial intelligence according to claim 1, characterized in that, The output meteorological data for the high-altitude plateau region includes output variables, diagnostic variables, and resolution data. Output variables include basic meteorological elements and convective parameters. Diagnostic variables include echo top height, vertical cumulative liquid water content, surface sensible heat flux, and local convergence. Resolution data includes temporal resolution and spatial resolution for the LES region.
7. The method for predicting severe convective weather at high-altitude airports based on numerical simulation and artificial intelligence according to claim 1, characterized in that, Key meteorological features were extracted from the simulation results and combined with field observation data to construct a multidimensional meteorological dataset. Data preprocessing was then performed, including: Key meteorological characteristic data include meteorological element data and diagnostic variables; meteorological element data include three-dimensional wind field, temperature, humidity, air pressure, precipitation, available convective potential energy, convective suppression energy, vertical wind shear, K index and wet potential vortex; diagnostic variables include echo top height, vertical cumulative liquid water content, surface sensible heat flux and local convergence. Field observation data includes Doppler radar data, satellite data, ground observation data, radiosonde data, auxiliary data, historical severe convective event data, and land use type data; Doppler radar data includes reflectivity factor, radial velocity, and spectral width, used to identify severe convective features; severe convective features include thunderstorm cells, bow echoes, and wind shear; satellite data includes cloud top brightness temperature, cloud system movement speed, and convective initiation indicators extracted from high-resolution infrared, water vapor, and visible light channel data acquired using geostationary satellites; ground observation data: real-time observation data collected from high-altitude airports and surrounding meteorological stations, including wind speed, wind direction, and temperature. The data includes information on local meteorological conditions such as humidity and air pressure; radiosonde data, including vertical profile data obtained using upper-air radiosondes; vertical profile data, such as temperature, humidity, and wind field profiles, used to verify and correct the atmospheric vertical structure simulated by WRF-LES; auxiliary data, including high-resolution topographic data, including topographic height, slope, and aspect; historical severe convective event data, including the spatiotemporal distribution and characteristic parameters of severe convective weather events in high-altitude plateau regions; severe convective weather events, including thunderstorms, hail, and short-duration heavy precipitation; and land use type data, used to characterize surface thermal properties and the impact of convection triggering. By using key meteorological feature data and field observation data, a multidimensional meteorological dataset is constructed. Key meteorological feature data that contributes the most to the prediction of severe convective weather are selected, and dimensionality reduction is performed using linear discriminant analysis.
8. The method for predicting severe convective weather at high-altitude airports based on numerical simulation and artificial intelligence according to claim 1, characterized in that, Spatial features of meteorological data are extracted using convolutional neural network models, including the spatial distribution of thunderstorm cells, cloud morphology, and wind field structure; temporal evolution patterns of meteorological elements are captured using long short-term memory networks, including the development of convective clouds, dynamic changes in wind fields, and changes in precipitation intensity; spatiotemporal joint models are used to analyze the spatiotemporal characteristics of meteorological element data; and model parameters are adjusted to dynamically adjust the contribution of physical features and topographic features.
9. The method for predicting severe convective weather at high-altitude airports based on numerical simulation and artificial intelligence according to claim 8, characterized in that, Add a physical constraint term to the loss function of the artificial intelligence prediction model; the physical constraint term includes the convection trigger threshold.
10. The method for predicting severe convective weather at high-altitude airports based on numerical simulation and artificial intelligence according to claim 8, characterized in that, The accuracy of the AI prediction model is obtained by calculating the cross-entropy loss function; the hyperparameters of the AI prediction model are optimized using cross-validation.