A method for short-term gust weather forecast

Abstract

The present invention discloses a method for forecasting short-term gusts, including the following steps: determining a forecast evaluation standard time period; obtaining four downscaled short-term gust forecast results based on different factor data and forecast model methods for different gust technical capture angles according to the determined forecast evaluation standard time period; then, performing real-time evaluation on the four short-term gust forecast results based on real-time monitoring data of the forecast evaluation standard time period, determining the weight of each short-term gust forecast result for the next forecast evaluation standard time period in an evaluation weight expression; then, outputting the short-term gust forecast result for the next forecast evaluation standard time period according to the evaluation weight expression and the weight of each short-term gust forecast result; and repeating the forecasting and evaluation steps to obtain the short-term gust forecast results for subsequent forecast evaluation standard time periods. The present invention can achieve more accurate and refined downscaled short-term gust forecasts.

Description

Technical Field

[0001] The present invention belongs to the field of weather forecasting using general artificial intelligence technology, and in particular relates to a method for short-term gust weather forecasting. Background Art

[0002] Gust forecasts are crucial for specific activities and industries, including aviation, navigation, construction, and outdoor activities. Specifically, they include: Aviation safety: For pilots and air traffic controllers, timely and accurate information about gust conditions is crucial, helping them make informed flight decisions and avoid dangerous situations involving strong gusts. Maritime safety: For ships and crew members, advance knowledge of gusts at sea helps them adopt appropriate navigation strategies to prevent serious impacts or damage to their vessels. Construction safety: At construction sites, gusts can pose a threat to aerial work and lifting operations, so early warning of gusts helps ensure safety measures are taken to protect workers and equipment. Outdoor activity safety: For outdoor sports events, camping, mountaineering, and other activities, early warning of gusts can help organizers adjust their plans and ensure the safety of participants. Agricultural production: For agricultural production, timely information about gust conditions can help farmers take appropriate protective measures to protect crops and facilities from damage.

[0003] However, due to the chaotic characteristics of the atmospheric wind field, the wind field weather system is diverse and complex, and will exhibit weather phenomena with nonlinear, non-uniform and dynamically changing characteristics. Therefore, for these specific activities or industries that require timely response, the current gust short-term weather forecast has a generally low level of sophistication and accuracy, and cannot provide more effective, more precise and more accurate gust short-term weather forecast services. Summary of the Invention

[0004] In order to address the shortcomings of the existing technology, the present invention uses improved deep learning extrapolation, statistical method extrapolation, wind field downscaling diagnostic model, optical flow method extrapolation, data assimilation and other methods, combined with site wind speed observation, radar reflectivity observation, fusion of real-time grid points, and numerical model products to obtain four methods that capture wind field changes from different technical perspectives and forecast wind field values ​​through different technical means, and effectively integrates the forecast results of the four methods to achieve more effective, more precise and more accurate gust short-term weather forecast services.

[0005] To achieve the above object, a method for short-term gust weather forecasting according to one embodiment of the present invention comprises the following steps:

[0006] S1. Determine the forecast evaluation standard time period and set k=1;

[0007] S2. Based on the determined kth forecast assessment standard time period, CALMET and WRF are fused to obtain the first gust short-term forecast results based on terrain data correction and downscaling based on terrain data, based on terrain data and numerical model forecast data;

[0008] S3. Based on the determined kth forecast assessment standard time period, and based on the radar reflectivity motion vector, station wind speed observation data, and numerical model forecast data, multi-scale variational analysis is used to obtain the second gust short-term forecast result based on the fusion and extrapolation of multiple observation data;

[0009] S4. According to the determined kth forecast evaluation standard time period, based on radar reflectivity data, station wind speed observation data, and numerical model forecast data, the PhyDnet spatiotemporal convolutional grid model is used to obtain the third gust short-term forecast result based on capturing severe convective meteorology and extrapolation;

[0010] S5. According to the determined kth forecast assessment standard time period, based on HRCLDAS multi-source fusion data, station wind speed observation data, and numerical model forecast data, the extended orthogonal empirical decomposition is used to obtain the fourth gust short-term forecast result based on vector wind field propagation and extrapolation;

[0011] S6. Based on the real-time monitoring data of the kth forecast evaluation standard time period, respectively, perform real-time evaluation on the first gust of wind short-term forecast result, the second gust of wind short-term forecast result, the third gust of wind short-term forecast result, and the fourth gust of wind short-term forecast result, and determine the weight of each gust of wind short-term forecast result in the k+1th forecast evaluation standard time period in the evaluation weight expression;

[0012] S7. Outputting the gust short-term meteorological forecast result for the k+1th forecast assessment standard time period according to the evaluation weight expression and the weight of each gust short-term forecast result;

[0013] S8, let k=k+1;

[0014] S9. Repeat steps S2-S8 to obtain the gust short-term meteorological forecast results for subsequent forecast evaluation standard time periods.

[0015] Furthermore, the forecast evaluation standard time period in step S1 is 2 hours.

[0016] Furthermore, in step S2, based on the determined kth forecast evaluation standard time period, CALMET and WRF are fused to obtain the first gust short-term forecast result based on terrain data correction and downscaling, based on terrain data, specifically:

[0017] S21. Downloading numerical model forecast data of ground and atmospheric meteorological elements according to the determined kth forecast evaluation standard time period;

[0018] S22, converting the downloaded processed numerical model forecast data of ground and atmospheric meteorological elements into a WPS intermediate file, and generating an intermediate file recognizable by WRF by combining it with the underlying surface data;

[0019] S23, using the WRF mode to read the intermediate file and generate a data file that can be recognized by the CALWRF program;

[0020] S24. Continue to use CALWRF to generate data files that can be recognized by CALMET;

[0021] S25. Input the data file that CALMET can identify into the CALMET model to perform wind field downscaling diagnosis;

[0022] S26. Output the CALMET model results to obtain the first gust of wind short-term forecast results based on the correction and downscaling of terrain data.

[0023] Furthermore, in step S3, according to the determined kth forecast evaluation standard time period, based on the radar reflectivity motion vector, the site wind speed observation data, and the numerical model forecast data, a multi-scale variational analysis is used to obtain a second gust short-term forecast result based on the fusion and extrapolation of multiple observation data, specifically:

[0024] S31. Downloading numerical forecast products, station wind speed observation data, and radar reflectivity data for the approaching time period according to the determined kth forecast evaluation standard time period;

[0025] S32. Using interpolation, unify different data times into target data time;

[0026] S33. Calculating a radar reflectivity motion vector according to the interpolated radar reflectivity data based on an optical flow method;

[0027] S34. Based on multi-scale variational analysis, the station wind speed observation, radar reflectivity motion vector and numerical forecast data are fused to obtain fused data;

[0028] S35. Based on the CALMET model and the WRF model, the fused wind field data is downscaled and diagnosed to obtain the near-real-world near-surface wind field v that conforms to the law of conservation of mass. obs ;

[0029] S36, adjusting the wind field predicted by the current numerical model using the obtained near-real-time and near-surface wind fields;

[0030] S37. Output the adjusted numerical model forecast wind field as the final wind field extrapolation forecast, and obtain the second gust of wind short-term forecast result based on the fusion and extrapolation of multiple observation data.

[0031] Furthermore, the expression adjusted in step S36 is:

[0032] v adj =w t v obs +(1-w t )v fcst

[0033]

[0034] Among them, v adj is the wind field forecasted by the adjusted numerical model; w t represents weight; v obs represents the near-realistic near-surface wind field that complies with the law of conservation of mass; v fcst represents the wind field at the current time predicted by the current model; τ represents the length of the time window; t i Represents the time of the current forecast hour.

[0035] Furthermore, in step S4, according to the determined kth forecast evaluation standard time period, based on radar reflectivity data, station wind speed observation data, and numerical model forecast data, the PhyDnet spatiotemporal convolutional grid model is used to obtain the third gust short-term forecast result based on capturing severe convective weather and extrapolation, specifically:

[0036] S41. Constructing a deep learning instantaneous average wind forecast model based on the PhyDnet spatiotemporal convolutional grid model;

[0037] S42. Establish a model for the relationship between instantaneous average wind and maximum wind using the least squares fitting statistical method for each site;

[0038] S43. Download the numerical forecast products, station wind speed observation data and radar reflectivity data for the near future;

[0039] S44, using an interpolation method to unify different data times into a target data time;

[0040] S45, using the site wind speed observation and radar reflectivity as input, generating a site wind speed forecast for the target data time of the kth forecast evaluation standard time period;

[0041] S46. Using multi-scale variational analysis, the generated site wind speed forecast is integrated with the model forecast;

[0042] S47. Based on the established model of the relationship between instantaneous average wind and maximum wind, the instantaneous maximum wind data is predicted through the obtained fusion data, and the short-term forecast result of the third gust of wind based on capturing severe convective weather and extrapolation is output.

[0043] Furthermore, in step S5, based on the determined k-th forecast evaluation standard time period, the extended orthogonal empirical decomposition is used to obtain the fourth gust short-term forecast result based on vector wind field propagation and extrapolation based on the HRCLDAS multi-source fusion data, the site wind speed observation data, and the numerical model forecast data, specifically:

[0044] S51. Construct an extended orthogonal empirical decomposition wind speed forecast model;

[0045] S52, downloading the HRCLDAS multi-source fusion near-real-time wind field data and numerical forecast data and site wind speed observation data at the approaching time;

[0046] S53, using time interpolation, interpolating the HRCLDAS and numerical forecast data to the time frequency of the observation data;

[0047] S54. Based on multi-scale variational analysis, the interpolation results of HRCLDAS and numerical forecast data are corrected using station wind speed observations to obtain near-real-time grid wind speed data;

[0048] S55. Input the obtained near-real-life grid wind speed data into the extended orthogonal empirical decomposition wind speed forecast model to obtain the fourth gust short-term forecast result based on vector wind field propagation and extrapolation.

[0049] Furthermore, in the extended orthogonal empirical decomposition wind speed forecast model constructed in step S51, the typical gust excitation and propagation process is solved by matrix eigenvalues ​​and eigenvectors as follows:

[0050]

[0051] Among them, [U1...U T , V1...V T ] represents the abbreviation of the wind speed field covariance matrix of the sliding window, The time coefficient of the window wind field transmission process, EOF i Represents the empirical orthogonal decomposition component, each component represents a wind field transmission process with a sliding window duration, i represents the grid point, T represents the matrix transpose, and ε represents the residual error.

[0052] Furthermore, the evaluation weight expression in step S6 is:

[0053] Gust fin =w1Gust1+w2Gust2+w3Gust3+w4Gust4

[0054] Among them, Gust finis the gust short-term meteorological forecast result for the k+1th forecast evaluation standard time period; Gust1 is the first gust short-term forecast result, w1 is the weight coefficient of the first gust short-term forecast result; Gust2 is the second gust short-term forecast result, w2 is the weight coefficient of the second gust short-term forecast result; Gust3 is the third gust short-term forecast result, w3 is the weight coefficient of the third gust short-term forecast result; Gust4 is the fourth gust short-term forecast result, w4 is the weight coefficient of the fourth gust short-term forecast result.

[0055] Furthermore, the values ​​of the weight coefficients w1, w2, w3 and w4 are 0-100%.

[0056] The beneficial effects of the present invention are:

[0057] 1. The present invention uses a variety of wind gust prediction methods based on different factors and technologies to obtain multiple prediction results for wind fields at different capture angles. It then selects the method that is closest to the real-time monitoring results for forecast output. This solution fully utilizes the advantages of various methods, selects the more accurate method under different meteorological conditions and at different times, and finds the prediction result that is closest to the actual situation through real-time comparison and selection, thereby reducing forecast uncertainty, improving the credibility and timeliness of the forecast, and being more adaptable to forecasts of complex and chaotic wind field weather systems.

[0058] 2. The first short-term wind gust forecast result of the present invention combines the CALMET model with the WRF model, and uses the 3km resolution numerical forecast wind field output by the WRF model as the input of the CALMET model to obtain a short-term wind gust downscale forecast. This solves the wind field-terrain incompatibility problem caused by simply using the interpolation method to downscale the numerical forecast product. It can correct the wind field numerical forecast results based on more detailed and realistic terrain data, thereby improving the accuracy and resolution of wind field simulation;

[0059] 3. The second gust short-term forecast result of the present invention is based on radar reflectivity data, station wind speed observation data, and numerical model forecast data. It uses multi-scale variational data assimilation, optical flow method, and CALMET model to fuse multiple observation data with forecast data to obtain a more accurate and refined ground-level near-real-time wind field. The fused data provides a more precise and comprehensive estimate of the atmospheric state, effectively improving forecast accuracy by understanding the complex wind field dynamics.

[0060] 4. The third short-term wind gust forecast result of the present invention is based on radar reflectivity data, station wind speed observation data, and numerical model forecast data. By identifying strong convective gusts, the PhyDnet spatiotemporal convolutional network splits the motion of the echo into advection motion terms and nonlinear disturbance terms, and introduces an attention mechanism to establish an optimal deep learning model for instantaneous average wind forecast, thereby achieving short-term wind speed forecast;

[0061] 5. The fourth short-term gust forecast result of the present invention is based on HRCLDAS multi-source fusion data, numerical forecast products, and site wind speed observation data. By constructing a wind speed covariance matrix and simulating the excitation and propagation process of gusts based on the extended empirical orthogonal decomposition method, an innovative extended orthogonal empirical decomposition forecast model is established to achieve short-term grid wind speed forecasting. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 This is a flow chart of a method for short-term gust weather forecasting according to the present invention;

[0063] Figure 2 A schematic diagram of a code file call for obtaining a first gust of wind short-term forecast result based on terrain data correction and downscaling according to an embodiment of the present invention;

[0064] Figure 3 This is a schematic diagram of a code file call for obtaining a second gust short-term forecast result based on fusion and extrapolation of multiple observation data according to an embodiment of the present invention;

[0065] Figure 4 This is a design diagram of a PhyDnet spatiotemporal convolutional network residual network according to an embodiment of the present invention;

[0066] Figure 5 A schematic diagram of a code file call for obtaining a short-term forecast result of the third gust of wind based on capturing severe convective weather and extrapolating it according to an embodiment of the present invention;

[0067] Figure 6 This is a schematic diagram of a code file call for obtaining a fourth wind gust short-term forecast result based on vector wind field propagation and extrapolation according to an embodiment of the present invention;

[0068] Figure 7 This is a comparison chart of the effects of live site observation data, gust short-term weather forecast data, and existing wind field HCLDAS data according to an embodiment of the present invention. DETAILED DESCRIPTION

[0069] In order to make the purpose, technical solutions and advantages of the present invention more clear, further description is given below with reference to the accompanying drawings and embodiments.

[0070] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0071] As an example of an implementation of the present invention, as shown in the attached Figure 1 As shown, the present invention provides a method for short-term gust weather forecasting, which includes the following steps:

[0072] S1. Determine the forecast evaluation standard time period and set k=1;

[0073] Taking the need to obtain a 1km resolution gust forecast product with a frequency of 10 minutes for the next 2 hours in western Guangdong as an example, the forecast evaluation standard time period is 2 hours, and the cycle parameter k is set to 1.

[0074] S2. Based on the determined kth forecast assessment standard time period, the CALMET model is fused with the WRF model based on the terrain data and the numerical model forecast data to obtain the first gust short-term forecast result based on the terrain data correction and downscaling;

[0075] The CALMET model is a wind field diagnostic model based on the constraint of the law of conservation of mass, and is provided for public use as commercial software. CALMET takes into account the impact of terrain characteristics on the wind field, including the impact of mountains, valleys, plains and other terrains on wind speed and direction. By simulating these terrain characteristics, CALMET can better reflect the impact of actual terrain on the wind field and generate relatively detailed wind field data. In addition, the CALMET model itself also takes into account the principle of conservation of mass. During the wind field downscaling process, it ensures the consistency of wind field information in space and time to meet the requirements of mass conservation. Therefore, using the CALMET model for wind field downscaling diagnosis can obtain wind field data that is more consistent with the actual terrain characteristics and the principle of conservation of mass, and obtain a downscaled wind field that is relatively consistent with more detailed underlying surface characteristics.

[0076] The Weather Research and Forecasting Model (WRF) is an advanced weather research and forecasting model. It is open source and freely available. It is suitable for meteorological applications at scales ranging from tens of meters to thousands of kilometers. The WRF model is widely used to simulate short-term weather forecasts, atmospheric processes, and long-term climate. Its accuracy exceeds that of traditional numerical weather prediction models, and its spatial and temporal resolution is higher, making it a valuable tool for meteorological and atmospheric research.

[0077] The first-gust short-term wind forecast results of this invention combine the CALMET model with the WRF model to perform wind downscaling analysis, effectively improving the refinement of wind forecasts. This method primarily addresses the wind-terrain inconsistency problem caused by simply downscaling numerical forecast products using interpolation methods. It also allows for corrections to wind numerical forecast results based on more refined and realistic terrain data.

[0078] The specific steps are as follows:

[0079] S21. Downloading numerical model forecast data of ground and atmospheric meteorological elements according to the determined kth forecast evaluation standard time period;

[0080] When k=1, the kth forecast evaluation standard time period is the 2-hour period from now, and the existing numerical model forecast data of ground and atmospheric meteorological elements in western Guangdong for the next 2 hours are downloaded, where:

[0081] Ground element data include: 2-meter air temperature, 2-meter relative humidity, 10-meter wind direction and speed, precipitation, sea level pressure, ground pressure and surface temperature.

[0082] Atmospheric element data include: geopotential heights of 925, 850, 700, 500, and 200 hPa layers, temperature, wind direction and speed, and specific humidity.

[0083] These data are obtained by downloading existing numerical model forecast products. The original data are 3km resolution data at an hourly frequency. After obtaining 10-minute frequency data through time interpolation, quality control is carried out through climate extreme value test, temporal continuity test and spatial continuity test.

[0084] S22. Convert the downloaded processed numerical model forecast data of ground and atmospheric meteorological elements into a WPS intermediate file, and generate an intermediate file that can be recognized by WRF by combining it with the underlying surface data.

[0085] The underlying surface data used mainly include terrain height, land type, etc. The data are directly obtained from the official website of the WRF model. For details, please refer to https: / / www2.mmm.ucar.edu / wrf / users / download / get_sources_wps_geog.html.

[0086] The WPS intermediate file has a fixed format. After obtaining the ground and atmospheric elements predicted by the numerical model, the obtained data is called and passed to the corresponding function in the winter.py program to directly generate it. After the WPS intermediate file is generated, the WPS geogrid.exe program is called to generate the geo_em.d01.nc file based on the underlying surface data. This file stores the underlying surface data interpolated to the target grid. After completing the above two steps, metgrid.exe is called to automatically generate the intermediate files required by the WRF model based on the underlying surface data and the WPS intermediate data, namely met_em.d01.*.nc (* represents the file timestamp).

[0087] S23. Using the WRF mode, read the intermediate file and generate a data file that can be recognized by the CALWRF program.

[0088] Modify the WRF model simulation parameter file namelist.input according to the forecast area and time information. Execute real.exe to generate the wrfinput.d01 file based on the met_em.d01.*.nc file. The wrfinput_d01 file can be directly recognized by CALWRF.

[0089] S24. Use CALWRF to generate a data file that can be recognized by CALMET.

[0090] Set up the CALWRF configuration file, specify the WRF output file path, output format, time range, grid resolution and other parameters, run the CALWRF program, read the WRF output file, and convert it to a format that CALMET can recognize. Including:

[0091] Prepare WRF output files: Make sure the WRF model runs to completion and generates the following files: wrfout*: The main output files of the WRF model.

[0092] Set up CALWRF: Edit the CALWRF configuration file and specify the following parameters: the path of the WRF output file, the format and location of the output file, the time step and simulation time range, and grid information.

[0093] Run CALWRF: Run the CALWRF program in the command line to ensure that CALWRF executes successfully and generates output files. The files contain the following information: timestamp, wind speed and direction at the grid point, temperature, humidity, and pressure data. Ensure that the temporal and spatial resolutions and formats of the data are compatible with the CALMET model.

[0094] S25. Use the CALMET model to perform wind field downscaling diagnosis.

[0095] The resulting recognizable data file is fed into the CALMET model as input. The model then runs, downscaling the input data to generate wind field information at a higher resolution than 1 km, including horizontal and vertical wind speed and direction. Using the CALMET model for wind field downscaling diagnosis can produce a 1 km wind field that closely matches the fine-grained terrain characteristics and satisfies mass conservation.

[0096] S26. Output the final wind field downscaling forecast result.

[0097] The output is the short-term forecast result of the first gust of wind in western Guangdong with a downscaled resolution of 1 km and a frequency of 2 hours and 10 minutes in the next 2 hours.

[0098] The method code is mainly divided into two functional modules:

[0099] Module 1: This section of the program consists primarily of four Python code files: run_calmet_downscaling.py, DataInterface.py, Algorithm.py, and winter.py. run_calmet_downscaling.py is the method execution program (main program) responsible for calling other programs. DataInterface.py is responsible for downloading the required numerical model products and writing data. Algorithm.py primarily includes functions such as temporal interpolation of model product predictions, spatial interpolation routines, and calling Module 2. winter.py stores data in a format recognizable by WPS and is called by DataInterface.py.

[0100] Module 2: This part is mainly composed of the compiled WRF model pre-processor, CALWRF conversion tool, and CALMET model body. Among them, / calmet / WPS is the WRF pre-processor, which is responsible for reading the model forecast data and underlying surface data, and generating intermediate data files that can be recognized by WRF. / calmet / WRF stores the WRF model program, which is responsible for converting the data generated by the WPS program into the WRF model initial field; / calmet / bin / calwrf is the data format conversion program, which is responsible for reading the WRF model initial field data and converting it into 3D.dat and 2D.dat format files that can be recognized by the CALMET model; / calmet / bin / calmet is the CALMET model program, which is responsible for reading the data output by calwrf and performing wind field downscaling diagnosis to obtain a 1km wind field that conforms to the fine terrain characteristics and satisfies the mass conservation. An example of an implementation of the code file call is attached. Figure 2 shown.

[0101] S3. Based on the determined kth forecast assessment standard time period, and based on the radar reflectivity motion vector, station wind speed observation data, and numerical model forecast data, multi-scale variational analysis is used to obtain the second gust short-term forecast result based on multiple observation fusion and extrapolation;

[0102] The second gust short-term forecast results of the present invention are based on radar reflectivity data, station wind speed observation data, and numerical model forecast data. Multi-scale variational data assimilation, optical flow, and the CALMET model are used to fuse these observational data with the forecast data to obtain a relatively accurate and refined near-real-time ground wind field. This near-real-time wind field is then used to adjust the forecast data, quickly generating a 1km resolution gust forecast product for the next 2 hours and 10 minutes in western Guangdong. Throughout this process, the fused data provides a more accurate and comprehensive estimate of the atmospheric state, which is of great significance for improving forecast accuracy and understanding complex wind field dynamics. It can provide a more reliable foundation for meteorological forecasts and services.

[0103] Based on multi-scale variational analysis, the process of fusing station wind speed observations, radar reflectivity motion vectors, and numerical forecasts can be summarized into the following steps:

[0104] S31. Data Collection:

[0105] According to the determined kth forecast evaluation standard time period, when k=1, that is, 2 hours from now, the numerical forecast products, site wind speed observation data and radar reflectivity data of the nearby time are downloaded.

[0106] Nearby times refer to the most recent model product, that is, the algorithm uses the most recent model data for each forecast; the downloaded hourly 3km resolution numerical forecast products, site wind speed observation data and 1km resolution radar reflectivity data.

[0107] S32. Use the interpolation method to uniformly interpolate numerical forecast products, station wind speed observation data, and radar reflectivity data to a 10-minute resolution;

[0108] S33. Calculating a radar reflectivity motion vector according to the interpolated radar reflectivity data based on an optical flow method;

[0109] The optical flow method used in the present invention is the Variational Echo Tracking method, which takes two adjacent radar reflectivity data as input, calls the corresponding function in vet.py, and generates a reflectivity motion vector. Since the motion vector is based on a Cartesian coordinate system, it can be further converted into m / s based on the grid resolution and the time resolution information of the radar data. Taking a 1km resolution grid as an example, assuming that the time interval between two adjacent radar data is 6 minutes, the motion vector is 1 grid in the x direction and 0 grid in the y direction, then the moving speed can be converted to 16.6m / s, and the moving direction is from west to east. The present invention uses variational technology to calculate the motion vector field of the radar echo. Compared with the traditional cross-correlation method, a more accurate motion vector field can be obtained.

[0110] S34. Based on multi-scale variational analysis, the site wind speed observation, radar reflectivity motion vector and numerical forecast are fused to obtain the fused data.

[0111] Multiscale variational analysis is a type of optimization method that considers characteristics of different scales during the data assimilation process. The present invention uses the variational data assimilation algorithm STMAS (Static Scale Transformation and Adaptive Mesh Refinement) based on the multi-grid concept to perform multiscale variational analysis. By introducing static scale transformation and adaptive mesh refinement technology, STMAS can effectively adjust and optimize the analysis field at different scales, improving the accuracy and efficiency of assimilation analysis. By transferring information between coarse and fine grids, STMAS can better process information of different scales in the observation data, thereby achieving better performance in the assimilation process.

[0112] The equations are shown in equations (1)-(4).

[0113]

[0114] Among them, x b represents the grid background field; y kRepresents the station observation increment used in multi-grid solution, y o represents the site observation value, x k Indicates that each layer needs to calculate the grid background field correction increment when solving multiple grids. x represents the final corrected background field. h k represents the observation operator that interpolates the grid background field to the site coordinates. The superscript k represents the number of different multigrid layers. kmax represents the number of multigrid layers to be set, which is determined according to the actual grid dimension. o represents the empirical ratio of the observation error covariance to the background field error covariance, which is usually between 0 and 1. T represents the transpose. J k is the name of the cost function of the kth layer.

[0115] When solving the STMAS equation, the present invention first interpolates the grid background field whose error needs to be corrected to a 3×3 dimensional grid, then uses equation (2) to calculate the variables in equation (1), and uses the quasi-Newton method to solve equation (3) to obtain x 1 Then interpolate the background field to a 6×6 grid, use equation (2) to calculate the variables in equation (1), and solve equation (3) to get x 2 In this way, the grid dimension of the background field interpolation is increased step by step by 2 times, and the background field increment x of the corresponding level is obtained by solving the problem. k , until the dimension is the same as the original dimension of the background field. Finally, the corrected background field is calculated using formula (4).

[0116] During fusion, the data fusion principle is the same as above, except that the "grid background field" is replaced with the numerical model forecast, and the "station observation" is replaced with the "station wind speed observation" or "radar reflectivity motion vector." Specifically, the station wind speed observation and the radar reflectivity motion vector converted to m / s are used as observation data and fused with the model forecast data using the STMAS method. The station wind speed observation is fused with the numerical model forecast 10m wind field using the same data fusion principle as described above for the STMAS method, except that the "grid background field" is replaced with the numerical model forecast, and the "station observation" is replaced with the "station wind speed observation." The radar reflectivity motion vector converted to m / s is used as the 700hPa wind field and is fused with the numerical model forecast high-altitude 3D wind field using the same data fusion principle as described above for the STMAS method, except that the "grid background field" is replaced with the numerical model forecast high-altitude 3D wind field, and the "station observation" is replaced with the "radar reflectivity motion vector converted to m / s." That is, a total of two data fusions are performed. The site wind speed observations are fused with the 10m wind field forecast by the numerical model to obtain the ground wind field; the radar reflectivity motion vector converted into m / s units is fused with the 700hPa wind field forecast by the numerical model to obtain the high-altitude wind field.

[0117] The impact of multi-scale analysis on wind field forecast results is mainly reflected in the following aspects:

[0118] Improved forecast accuracy: Multiscale analysis can capture the complexity and dynamics of wind fields, particularly at small and medium scales, such as wind structures within convective storms and boundary layers. By integrating observational and model data at different scales, more accurate predictions of wind field changes can be made.

[0119] Enhanced understanding of wind dynamics: Multiscale analysis provides a deeper understanding of how winds form, develop, and evolve, including their interactions with large-scale weather systems and small-scale convective activity. This helps to more accurately predict wind intensity and direction.

[0120] Optimizing resource allocation for wind forecasting: By performing calculations at appropriate scales, multi-scale analysis can more efficiently utilize computing resources and reduce unnecessary computational burdens. This means more efficient management of computing and storage resources for forecast centers and service providers.

[0121] Improved model performance for wind forecasts: Multiscale analysis can help improve the performance of numerical weather prediction (NWP) models for wind forecasts, particularly in handling convective weather and boundary layer dynamics. This can enhance the model's predictive power, especially in regions with active convection.

[0122] Enhanced decision support for wind forecasts: For decision makers, multi-scale analysis provides more detailed and accurate information on wind status, helping to more effectively formulate and implement strategies to deal with extreme wind events.

[0123] Promoting scientific research on wind field forecasting: Multi-scale analysis provides new perspectives and tools for meteorological science research, helping to deepen our understanding of the formation mechanism and changing laws of wind fields, thereby promoting the advancement of wind field forecasting technology.

[0124] S35. Based on the CALMET model and the WRF model, the fused wind field data is downscaled and diagnosed to obtain the near-real-world near-surface wind field v that conforms to the law of conservation of mass. obs .

[0125] Using steps similar to those in S2 above, based on the CALMET model and the WRF model, the fused data is downscaled and diagnosed to obtain the near-real-world, near-surface wind field v that conforms to the law of conservation of mass. obs .

[0126] S36. Use the obtained near-real-time and near-surface wind fields to adjust the wind field forecasted by the current numerical model. The adjustment formula is as follows:

[0127] v adj =w t v obs +(1-w t )vfcst

[0128]

[0129] Among them, v adj is the wind field forecasted by the adjusted numerical model; w t represents weight; v obs represents the near-realistic near-surface wind field that complies with the law of conservation of mass; v fcst represents the wind field at the current time predicted by the current numerical model; τ represents the time window length, which is less than or equal to 2 hours; t i Represents the current forecast time, which is 10 minutes.

[0130] S37. The output of the adjusted numerical model forecast wind field is the final wind field extrapolation forecast, and the short-term forecast result of the second gust of wind in the western Guangdong region with a frequency of 1 km resolution and extrapolation in the next 2 hours and 10 minutes is obtained.

[0131] The method code is mainly divided into four functional modules:

[0132] Module 1: This section of the program is primarily responsible for calling each module, downloading data, and reading data. The module consists of four Python files: run_obs_extrapolation_main.py, AlgorithmLib.py, DataInterface.py, and winter.py. Among them, run_calmet_downscaling.py is the method execution program (main program), responsible for calling other programs. DataInterface.py is responsible for downloading the required numerical model products and writing data. Algorithm.py mainly includes functions such as temporal interpolation of model product predictions, spatial interpolation programs, and calling other modules. winter.py is responsible for storing data in a format recognizable by WPS and is called by DataInterface.py.

[0133] Module 2: This section of the program is responsible for extrapolating radar reflectivity using optical flow to obtain the reflectivity motion vector. This module contains two program files, vet.py and _vet.c, which together implement the VariationalEcho Tracking method and are ultimately called by Algorithm.py in Module 1.

[0134] Module 3: This section of the program is responsible for performing multiscale variational analysis and fusing different data. This module contains a program file, stmas_lib.f90, which implements the multiscale variational analysis method and is ultimately called by Algorithm.py in Module 1.

[0135] Module 4: This part is mainly composed of the compiled WRF model pre-processor, CALWRF conversion tool, and CALMET model body. Among them, / calmet / WPS is the WRF pre-processor, which is responsible for reading the model forecast data and underlying surface data, and generating intermediate data files that can be recognized by WRF. / calmet / WRF stores the WRF model program, which is responsible for converting the data generated by the WPS program into the WRF model initial field; / calmet / bin / calwrf is the data format conversion program, which is responsible for reading the WRF model initial field data and converting it into 3D.dat and 2D.dat format files that can be recognized by the CALMET model; / calmet / bin / calmet is the CALMET model program, which is responsible for reading the data output by calwrf and performing wind field downscaling diagnosis to obtain a 1km wind field that conforms to the fine terrain characteristics and satisfies the mass conservation. An example of an implementation of the code file call is attached. Figure 3 shown.

[0136] S4. According to the determined kth forecast evaluation standard time period, based on radar reflectivity data, station wind speed observation data, and numerical model forecast data, the PhyDnet spatiotemporal convolutional grid model is used to obtain the third gust short-term forecast result based on capturing severe convective meteorology and extrapolation;

[0137] The third gust short-term forecast result method of the present invention is based on radar reflectivity data, station wind speed observation data, and numerical model forecast data. It constructs an instantaneous average wind forecast model based on the PhyDnet spatiotemporal convolution grid model, obtains instantaneous average wind forecast data, and then uses the least squares fitting statistical method to establish a relationship model between instantaneous average wind and maximum wind, effectively capturing strong wind weather, achieving more accurate instantaneous maximum wind short-term forecast, and quickly obtaining a gust forecast product with a frequency of 1km resolution for the next 2 hours and 10 minutes in western Guangdong.

[0138] The detailed calculation process of the method is as follows:

[0139] S41. Based on historical radar reflectivity and site wind speed observations, the model parameters are trained using the PhyDnet method and stored locally. The specific implementation process of this method is as follows:

[0140] S41. Constructing a deep learning instantaneous average wind forecast model based on the PhyDnet spatiotemporal convolutional grid model;

[0141] Identification of severe convective gusts: Collect historical multi-year regional reflectivity factor network data and minute-level instantaneous average wind observation data:

[0142] Weather radar observation data in Guangdong Province is used for quality control and networking to form a weather radar sample dataset with an interval of 10 minutes and a sequence length of 2 hours. It includes:

[0143] Collect weather radar observation data from Guangdong Province. Ensure the data has a temporal resolution of 10 minutes; decode the raw radar data and convert it into a workable format (such as NetCDF, HDF5, or a simple binary format); and remove missing values, outliers, and non-meteorological echoes. Apply quality control algorithms such as: noise filtering to remove background noise; radar self-test to check whether the radar hardware is working properly; raindrop attenuation correction to correct for radar wave attenuation caused by raindrops; ground clutter removal to eliminate echoes from fixed ground objects (such as mountains and buildings); network the data from different radar stations in Guangdong Province to form a radar dataset covering the entire province. Georeference the coverage areas of different radars to ensure data compatibility; sample the data at 10-minute intervals. Select a continuous time series of 2 hours in length; and store the processed sample dataset for further analysis and use.

[0144] Utilizing ground-based instantaneous mean wind observations with a 5-minute resolution, resampled to a 10-minute temporal resolution, and then gridded, these observations form a sample instantaneous mean wind dataset that matches weather radar observations. This process involves: collecting ground-based instantaneous mean wind observations with a 5-minute resolution, typically from meteorological stations or other ground-based observation facilities; confirming that the data's timestamps are formatted correctly and contain no missing or outliers; using data preprocessing techniques to resample the 5-minute resolution instantaneous mean wind data to a 10-minute temporal resolution; and storing the gridded dataset for matching with weather radar data and further analysis.

[0145] This paper constructs an instantaneous average wind forecast dataset by matching radar network data from the previous hour and ground instantaneous average wind data as input factors, and using instantaneous average wind observations for the next two hours as the ground truth field. This includes ensuring that both radar network data and ground instantaneous average wind data have been processed and converted to the same spatial and temporal resolution. It also ensures that instantaneous average wind observations for the next two hours are available for verification. The radar data and ground wind data are aligned in time, ensuring that each record has corresponding input and ground truth fields. The dataset is constructed by creating a dataset in which each sample contains radar and wind data from the previous hour as features and instantaneous average wind observations for the next two hours as labels. The instantaneous average wind forecast dataset is split into training, validation, and test sets in an 8:1:1 ratio.

[0146] Establishment of severe convective gust nowcasting model: Based on the PhyDnet spatiotemporal convolutional grid model, an instantaneous average wind forecast model is constructed to perform instantaneous average wind forecast, including:

[0147] The PhyDnet spatiotemporal convolutional network improves the long-term forecast effect by splitting the motion of the echo into horizontal motion terms and nonlinear disturbance terms, and introducing an attention mechanism. The PhyDnet spatiotemporal convolutional network model is trained using a constructed sequence training set, and the model effect is tested using a validation set. Based on the test results, the model structure and loss function are continuously adjusted to continuously optimize the instantaneous average wind forecast model.

[0148] PhyDnet spatiotemporal convolutional network splits the motion of the echo into advection motion and nonlinear perturbation terms. PhyDNet adopts a dual-branch architecture of physical unit PhyCell and basic spatiotemporal convolution unit ConvLSTM. The physical dynamic terms and non-physical factor terms (such as nonlinear perturbation terms such as image texture and new details) that can be described by partial differential equations can be separated by residual network design. The specific design method of the residual network can be seen in Figure 4 Specifically, given that at time t there are several radar / precipitation images X = X(t) as input, the core problem can be simplified to how to obtain the output (prediction) sequence from the input sequence X Use h(t) to represent the sequence characteristics (latent space characteristics) of each channel at the input end, then in the framework of PhyDNet, h(t) can be represented by the following partial differential equation:

[0149]

[0150] Where M p (h p ,X) and M r (h r ,X) are the physical motion term and nonlinear disturbance term of h respectively. The former is learned through PhyCell, that is, it is assumed that the precipitation nowcasting problem conforms to a certain complex partial differential equation expression, and the latter is directly learned using ConvLSTM, that is, the model's characterization of the disturbance term does not depend on any preset physical laws, but only depends on the local variation characteristics of precipitation.

[0151] The instantaneous mean wind forecast results of the deep learning model are evaluated on the test set using scoring methods (such as TS and HSS), taking into account both the accuracy and speed of the model, and the optimal deep learning model for instantaneous mean wind forecast is obtained.

[0152] S42. On this basis, continuously collect historical observation data of instantaneous maximum wind and average wind in the study area over many years, and establish a relationship model between instantaneous average wind and maximum wind based on the least squares fitting statistical method between sites; including:

[0153] Data Collection: Gather several years of historical maximum and mean wind observation data. This data is typically available from weather stations or meteorological databases. Ensure that the data covers multiple weather stations and contains sufficient time series data for statistical analysis.

[0154] Data preprocessing: Clean the data to remove errors and outliers. Standardize or normalize the data if necessary. Ensure that the maximum and mean wind data are aligned in time.

[0155] Site-to-site analysis: For each meteorological site, the maximum wind data are paired with the corresponding mean wind data.

[0156] Least Squares Fitting: The maximum and mean wind data for each station are fitted using the least squares method to model the relationship between them. The goal of the least squares method is to minimize the squared difference between the predicted values ​​and the actual observed values.

[0157] The following are the specific statistical modeling steps:

[0158] Establish a hypothetical model: Assume that there is a linear relationship model, which can be expressed as:

[0159] V_max=a*V_avg+b

[0160] Where V_max is the maximum wind speed, V_avg is the average wind speed, and a and b are model parameters.

[0161] Apply the least squares method: For each site, calculate the parameters a and b so that the following loss function is minimized:

[0162]

[0163] Where n is the number of observations.

[0164] Parameter estimation: By taking the derivative and setting it to zero, we can get the estimated values ​​of parameters a and b:

[0165]

[0166] Model validation: Use cross-validation or other statistical methods to verify the accuracy of the model. Further, calculate the coefficient of determination (R 2 ) to assess the goodness of fit of the model.

[0167] Model application: Use the established model to predict the maximum wind data of the data point through the average wind observation data.

[0168] Through the above steps, a relationship model between instantaneous average wind and maximum wind based on historical observation data is established for future wind field prediction and analysis.

[0169] S43. Download the numerical forecast products, station wind speed observation data and radar reflectivity data for the upcoming time.

[0170] The numerical forecast product has a 3km resolution; the site wind speed observation data is site data, interpolated to a 1km resolution grid using the inverse distance weighted method; the radar composite reflectivity data has a 1km resolution

[0171] S44. Use interpolation to unify different data times to a 10-minute resolution.

[0172] S45. Read the pre-trained optimal deep learning model parameter file for instantaneous average wind forecast, use the site wind speed observation and radar reflectivity as input, and generate a site wind speed forecast for the next 2 hours at a frequency of 10 minutes.

[0173] S46. Use multi-scale variational analysis to fuse the generated site wind speed forecast with the model forecast.

[0174] The station wind speed forecast generated in step S44 is fused with the model forecast data using the multiscale variational analysis method in step S3. The station wind speed observations are fused with the numerical model forecast wind field. The data fusion principle is the same as described in the STMAs method above, except that the "grid background field" is replaced by the numerical model forecast, and the "station observations" are replaced by "station wind speed observations."

[0175] S47. Based on the established model of the relationship between instantaneous average wind and maximum wind, the instantaneous maximum wind data is predicted through the obtained fusion data, and the instantaneous maximum wind forecast data for the next 2 hours is output, and the short-term forecast results of the third gust of wind in the western Guangdong region with a frequency of 1km resolution and extrapolation for the next 2 hours and 10 minutes are obtained.

[0176] The method code is mainly divided into three functional modules:

[0177] Module 1: This part of the program is mainly responsible for calling each module. This module contains a Python file, run_GDW_wind_main.py, which is the main program of the entire method.

[0178] Module 2: This module is the core program of the PhyDnet model and contains a code file called phydnet_att.py. This module is responsible for reading the model pre-trained parameters stored in GDWRInterface / weight, encoding and decoding the data, loading the relevant deep learning library, and generating wind speed forecast results.

[0179] Module 3: This part of the program is mainly responsible for downloading forecast data and various types of observation data, performing spatiotemporal interpolation, and performing multi-scale variational analysis. The module contains two program files, __init__.py and stmas_lib.f90. Among them, __init__.py implements the downloading and spatiotemporal interpolation of forecast data and various types of observation data, and stmas_lib.f90 is the main program for multi-scale variational analysis. An example of the code file call is shown in the attached figure. Figure 5 shown.

[0180] S5. According to the determined kth forecast assessment standard time period, based on HRCLDAS multi-source fusion data, station wind speed observation data, and numerical model forecast data, the extended orthogonal empirical decomposition is used to obtain the fourth gust short-term forecast result based on vector wind field propagation and extrapolation;

[0181] This method is based on HRCLDAS multi-source fusion data, numerical forecast products, and station wind speed observations. It uses the extended orthogonal empirical decomposition method to establish a forecast model, solves the typical gust excitation and propagation process through matrix eigenvalues ​​and eigenvectors, and realizes short-term grid wind speed forecast based on vector wind field, obtaining a gust forecast product with a frequency of 1km resolution for the next 2 hours and 10 minutes in western Guangdong.

[0182] The detailed calculation process of the method is as follows:

[0183] S51. Construct an extended orthogonal empirical decomposition wind speed forecast model;

[0184] Download historical HRCLDAS data and numerical forecast data, use the extended orthogonal empirical decomposition method to train the model parameters, and store them locally.

[0185] The construction of the extended orthogonal empirical decomposition wind speed forecast model is mainly divided into the following three steps:

[0186] ①Construction of wind speed covariance matrix

[0187] Unlike scalar fields, wind fields are expressed as binary vectors, usually expressed as a meridional wind component and a zonal wind component. The covariance of a vector is not only related to the vector's magnitude, but also to the angle, so it is necessary to use a vector. This paper uses the cosine of the two components of the wind field as the definition of the wind field covariance. The specific formula is as follows:

[0188] f=v a 2 +v b 2 -2v a v b cosγ

[0189] Among them, v a Indicates the wind speed at the previous moment, v brepresents the wind speed at the next moment, and γ represents the vector angle between the wind speeds at the two moments. In actual calculations, the longitudinal and latitudinal components of the wind speed at different moments need to be listed in two columns of the wind speed matrix. That is, for a set of n grid points and t moments of data, an n-row, 2t-column matrix can be constructed. The transpose of this matrix multiplied by the resulting n×n matrix is ​​the wind field covariance matrix.

[0190] This solution takes into account the modulus of the two vectors and their angle, so that the subsequent typical feature extraction of the wind field propagation process based on this takes into account both wind speed and wind direction, with higher discrimination.

[0191] ②Simulate the excitation and propagation process of gusts based on the extended empirical orthogonal decomposition method

[0192] The wind speed field covariance matrix is ​​generated by the time window sliding method:

[0193]

[0194] Among them, U i,j Indicates the zonal wind speed component in the jth sliding window at the i-th grid point within a time sliding window, V i,j It represents the longitudinal wind speed component in the jth sliding window at the i-th grid point within a time sliding window. The difference from the wind speed covariance matrix in ① is that j represents a sliding window, that is, τ columns, and τ represents the width of the sliding window.

[0195] The excitation and propagation processes of typical gusts are solved through matrix eigenvalues ​​and eigenvectors.

[0196]

[0197] Among them, the left side of the formula is the abbreviation of the wind speed field covariance matrix of the sliding window, and the right side of the formula is The time coefficient of the window wind field transmission process, EOF i represents the empirical orthogonal decomposition component, each component represents a wind field transmission process with a sliding window duration (10 minutes in the present invention), T represents the matrix transpose, and ε represents the residual error.

[0198] This method can ensure that different typical processes are orthogonal, and the first several modes can explain most of the variance, that is, the typical process is representative.

[0199] ③Construct an extended orthogonal empirical decomposition wind speed forecast model

[0200] Based on the extended empirical orthogonal decomposition of the gridded wind speed field, the similarity between the wind field evolution process and various typical processes within the current window period (1 hour) can be analyzed. Based on the similarity time coefficients of different typical processes and combined with historical data, the path of subsequent gust evolution can be obtained. The prediction of the time coefficient is achieved by the Long Short-Term Memory (LSTM) network.

[0201] S52. Download the HRCLDAS multi-source fusion near-real-time wind field data, numerical forecast data, and site wind speed observation data at the approaching time.

[0202] Download the most recent model product. The original numerical forecast data has an hourly resolution of 3 km, the original site wind speed observation data has an hourly resolution, and the HCLDAS multi-source fusion near-real-time wind field data has a spatial resolution of 1 km and an hourly temporal resolution. Forecasts are valid for two hours after the algorithm is started.

[0203] S53. Use time interpolation to interpolate the HRCLDAS and numerical forecast data to the 10-minute frequency of the observation data.

[0204] S54. Based on multi-scale variational analysis, the interpolation results of HRCLDAS and numerical forecast data are corrected using station wind speed observations to obtain near-real-world grid wind speed data.

[0205] Based on the use of multi-scale variational analysis similar to the above S3 step, that is, a total of two data fusions are performed, and the interpolation results of HRCLDAS and numerical forecast data are corrected using station wind speed observations: the station wind speed observations are fused with the 10m wind field forecast data of HRCLDAS (the data of the High Resolution Land Data Assimilation System HRCLDAS mainly focuses on the ground level, usually including near-surface meteorological variables. In its meteorological observations, the standard reference height of variables such as wind speed is 10 meters) to obtain the ground wind field grid wind speed data; the station wind speed observations are fused with the 700hPa wind field forecast of the numerical model to obtain the high-altitude wind field grid wind speed data, and the combination obtains the near-actual grid wind speed data.

[0206] S55. Load the pre-trained model parameters, input the extended orthogonal empirical decomposition wind speed forecast model constructed by the obtained near-real-time grid wind speed data, and output the short-term forecast result of the fourth gust of wind in western Guangdong in the next 2 hours and 10 minutes with a resolution of 1 km based on the vector wind field and extrapolation.

[0207] The method code is mainly divided into two functional modules:

[0208] Module 1: This section of the program is responsible for data downloading, reading and writing, spatiotemporal interpolation, and calling Module 2. This module contains three Python code files: run_hrcldas_extrapolation_main.py, AlgorithmLib.py, and DataInterface.py. run_hrcldas_extrapolation_main.py is the main program, DataInterface.py is responsible for data downloading and reading and writing, and Algorithm.py is responsible for calling Module 2, reading pretrained model parameters, and performing spatiotemporal interpolation of data.

[0209] Module 2: This part of the program is mainly responsible for multi-scale variational analysis and fusing different data. The module contains a program file, stmas_lib.f90, which implements the multi-scale variational analysis method and is ultimately called by Algorithm.py in Module 1. An example of a code file call is shown in the attached file. Figure 6 shown.

[0210] S6. Based on the real-time monitoring data of the kth forecast evaluation standard time period, respectively, perform real-time evaluation on the first gust of wind short-term forecast result, the second gust of wind short-term forecast result, the third gust of wind short-term forecast result, and the fourth gust of wind short-term forecast result, and determine the weight of each gust of wind short-term forecast result in the k+1th standard time period in the evaluation weight expression;

[0211] Output file description of the above steps S2-S5:

[0212] The output results of the four methods are all in the netcdf format. The variables related to the short-term wind speed forecast in the file are described as follows:

[0213] lon: longitude information; lat: latitude information; time: forecast time; u10_fcst: zonal wind forecast, three-dimensional array (time, latitude, longitude); v10_fcst: radial wind forecast, three-dimensional array (time, latitude, longitude).

[0214] According to the kth forecast evaluation standard time period, when k=1, that is, the real-time monitoring data for the next 2 hours, the first gust of wind short-term forecast result, the second gust of wind short-term forecast result, the third gust of wind short-term forecast result and the fourth gust of wind short-term forecast result are respectively evaluated in real time to determine the weight of each gust of wind short-term forecast result in the evaluation weight expression for the k=2th standard time period, that is, the forecast evaluation time period for the 3rd to 4th hours in the future.

[0215] S7. Outputting the k+1th (the second when k=1) gust short-term meteorological forecast result for the forecast evaluation standard time period according to the evaluation weight expression;

[0216] Evaluate the weight expression:

[0217] Gust fin =w1Gust1+w2Gust2+w3Gust3+w4Gust4

[0218] Among them, Gust fin Gust1 is the first gust forecast, w1 is its weighting factor; Gust2 is the second gust forecast, w2 is its weighting factor; Gust3 is the third gust forecast, w3 is its weighting factor; Gust4 is the fourth gust forecast, w4 is its weighting factor. Weighting factors w1, w2, w3, and w4 range from 0 to 100%.

[0219] The evaluation weight expression may adopt a weighted sum expression, in which the forecast result with the larger weight accounts for the largest proportion in the final forecast result.

[0220] Depending on the value of the weight coefficient, the evaluation weight expression can also be expressed as the optimal method weight expression. For example, after comparing and evaluating the four gust short-term forecast results in the first standard time period with the real-time monitoring data, the weight of the gust short-term forecast result closest to the real-time monitoring data is 100%, and the weights of other gust short-term forecast results are 0%. That is, the forecast result of the gust short-term forecast method that performs best at the first moment in the second standard time period is taken as the final gust short-term meteorological forecast result for the second standard time period.

[0221] According to the evaluation weight expression, the output is the second forecast evaluation standard time period, that is, the gust short-term meteorological forecast result for the 3rd to 4th hour in the future.

[0222] S8, let k=k+1;

[0223] Let k=k+1. When k=1, let k=2.

[0224] S9. Repeat steps S2-S8 to obtain the gust short-term meteorological forecast results for subsequent forecast evaluation standard time periods.

[0225] Steps S2 to S8 are repeatedly executed for the first time to obtain the gust short-term meteorological forecast result for the subsequent third forecast evaluation standard time period.

[0226] By analogy, steps S2-S8 are repeatedly executed in a loop, and the results of the four forecast methods for each standard time period are evaluated in real time to obtain the gust short-term weather forecast results for each future 2-hour standard time period.

[0227] Taking the western Guangdong Province as an example, its wind field weather system exhibits weather phenomena with nonlinear, non-uniform and dynamic change characteristics. These weather phenomena usually involve interactions at multiple scales, including large-scale weather systems (such as fronts, high-pressure and low-pressure systems) and small and medium-scale weather systems (such as convective storms, thunderstorms and tornadoes). These systems usually have rapid change processes, including the movement, intensity changes and structural adjustments of weather systems, which makes it difficult to effectively predict wind field weather systems.

[0228] As attached Figure 7 As shown, after comparing the actual site observation data, the gust short-term weather forecast data of the present invention and the existing wind field HCLDAS data issued by the National Meteorological Administration, it can be found that the method of the present invention has made a relatively effective correction to the existing wind field data.

[0229] This invention utilizes a variety of gust forecasting methods based on different weather factors, tools, and technologies. This method studies and predicts wind farm weather systems from different technical perspectives, yielding multiple forecast results. The method then evaluates weighted expressions to select the method that most closely matches real-time monitoring for high-quality forecast output, thereby providing more refined and accurate wind farm forecast services. By selecting the forecast method that most closely matches real-time monitoring for output, the invention improves the real-time and accuracy of forecasts, reduces the impact of uncertainties, adapts to complex weather systems, enhances the timeliness of forecasts, strengthens decision support, and promotes scientific research.

[0230] Because a fixed prediction method has been proven to have limitations for different wind speed fields in different regions or in the same region, this method implements the latest gust short-term forecast method. In view of the chaotic characteristics of the atmospheric wind field, it predicts future changes from various different determining factors of wind field changes and applies them in real-time and effective comprehensive applications. By innovatively designing a gust statistical forecast method based on empirical orthogonal decomposition, and four gust observation extrapolation methods based on optical flow method and wind field diagnostic downscaling model, it realizes gust diagnostic downscaling forecast for traditional numerical forecast products. Taking into account the characteristics of different methods (such as basic data download time, model calculation time, forecast accuracy, data capacity, etc.), the specific implementation process of the method is optimized, and finally an implementation method that integrates four gust short-term forecast methods is obtained. Then, by real-time evaluation and comprehensive application of the four results, more accurate and detailed wind field forecast results are obtained.

[0231] Furthermore, the method of the present invention can be continuously expanded in the subsequent process, and other new research and development analysis angles and technical forecasting methods can be added, incorporated into the evaluation weight expression, and multiple results can be evaluated and comprehensively applied in real time to obtain more accurate and detailed wind field forecast results.

[0232] In summary, the method of the present invention can capture different technical aspects of meteorological phenomena through different factors and different methods. In particular, in complex or changing environments, the use of multiple prediction methods can provide multiple perspectives, help reduce uncertainty, and improve the credibility of forecasts. In addition, wind field meteorological conditions are changeable, and a single prediction method may not be able to adapt to all situations. Through multiple methods, it is possible to better adapt to changing environments, improve the flexibility and adaptability of forecasts, better predict and respond to extreme weather events, and reduce the risks they bring. Furthermore, by storing and analyzing the performance of different prediction methods in different locations, different environments, different seasons and time periods to form a data set, it can promote more in-depth scientific research and technological progress, help improve prediction technologies and methods, and enhance the understanding and prediction capabilities of future weather systems.

[0233] The present invention and its embodiments are described above. This description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if a person skilled in the art is inspired by this and, without departing from the purpose of the present invention, designs structures and embodiments similar to this technical solution without inventiveness, they shall fall within the scope of protection of the present invention.

Claims

1. A method for short-term gust weather forecasting, characterized in that: It includes the following steps: S1. Determine the forecast evaluation standard time period and set k=1; S2. Based on the determined kth forecast assessment standard time period, CALMET and WRF are fused to obtain the first gust short-term forecast results based on terrain data correction and downscaling based on terrain data, based on terrain data and numerical model forecast data; S3. According to the determined kth forecast assessment standard time period, based on the radar reflectivity motion vector, the station wind speed observation data, and the numerical model forecast data, a multi-scale variational analysis is used to obtain the second gust short-term forecast result based on the fusion and extrapolation of multiple observation data; S4. According to the determined kth forecast evaluation standard time period, based on radar reflectivity data, station wind speed observation data, and numerical model forecast data, the PhyDnet spatiotemporal convolutional grid model is used to obtain the third gust short-term forecast result based on capturing severe convective meteorology and extrapolation; S5. According to the determined kth forecast evaluation standard time period, based on HRCLDAS multi-source fusion data, station wind speed observation data, and numerical model forecast data, the extended orthogonal empirical decomposition is used to obtain the fourth gust short-term forecast result based on vector wind field propagation and extrapolation; S6. Based on the real-time monitoring data of the kth forecast evaluation standard time period, respectively, perform real-time evaluation on the first gust of wind short-term forecast result, the second gust of wind short-term forecast result, the third gust of wind short-term forecast result, and the fourth gust of wind short-term forecast result, and determine the weight of each gust of wind short-term forecast result in the k+1th forecast evaluation standard time period in the evaluation weight expression; S7. Outputting the gust short-term meteorological forecast result for the k+1th forecast assessment standard time period according to the evaluation weight expression and the weight of each gust short-term forecast result; S8, let k=k+1; S9. Repeat steps S2-S8 to obtain the gust short-term meteorological forecast results for subsequent forecast evaluation standard time periods.

2. The method for short-term gust weather forecasting according to claim 1, characterized in that: The forecast evaluation standard time period in step S1 is 2 hours.

3. The method for short-term gust weather forecasting according to claim 1, characterized in that: The step S2, according to the determined kth forecast assessment standard time period, based on the terrain data and the numerical model forecast data, uses CALMET and WRF fusion to obtain the first gust short-term forecast result based on terrain data correction and downscaling, specifically: S21. Downloading numerical model forecast data of ground and atmospheric meteorological elements according to the determined kth forecast evaluation standard time period; S22, converting the downloaded processed numerical model forecast data of ground and atmospheric meteorological elements into a WPS intermediate file, and generating an intermediate file recognizable by WRF by combining it with the underlying surface data; S23, using the WRF mode to read the intermediate file and generate a data file that can be recognized by the CALWRF program; S24. Continue to use CALWRF to generate data files that can be recognized by CALMET; S25. Input the data file that CALMET can identify into the CALMET model to perform wind field downscaling diagnosis; S26. Output the CALMET model results to obtain the first gust of wind short-term forecast results based on the correction and downscaling of terrain data.

4. The method for short-term gust weather forecasting according to claim 1, characterized in that: The step S3, according to the determined kth forecast evaluation standard time period, based on the radar reflectivity motion vector, the site wind speed observation data, and the numerical model forecast data, uses multi-scale variational analysis to obtain the second gust short-term forecast result based on the fusion and extrapolation of multiple observation data, specifically: S31. Downloading numerical forecast products, station wind speed observation data, and radar reflectivity data for the approaching time period according to the determined kth forecast evaluation standard time period; S32. Using interpolation, unify different data times into target data time; S33. Calculating a radar reflectivity motion vector according to the interpolated radar reflectivity data based on an optical flow method; S34. Based on multi-scale variational analysis, the station wind speed observation, radar reflectivity motion vector and numerical forecast data are fused to obtain fused data; S35. Based on the CALMET model and the WRF model, the fused wind field data is downscaled and diagnosed to obtain the near-real-world near-surface wind field v that conforms to the law of conservation of mass. obs ; S36, adjusting the wind field predicted by the current numerical model using the obtained near-real-time and near-surface wind fields; S37. Output the adjusted numerical model forecast wind field as the final wind field extrapolation forecast, and obtain the second gust of wind short-term forecast result based on the fusion and extrapolation of multiple observation data.

5. The method for short-term gust weather forecasting according to claim 4, characterized in that: The expression adjusted in step S36 is: v adj =w t v obs +(1-w t )v fcst Among them, v adj is the wind field forecasted by the adjusted numerical model; w t represents weight; v obs represents the near-realistic near-surface wind field that complies with the law of conservation of mass; v fcst represents the wind field at the current time predicted by the current model; τ represents the length of the time window; t i Represents the time of the current forecast hour.

6. The method for short-term gust weather forecasting according to claim 1, characterized in that: In step S4, according to the determined kth forecast evaluation standard time period, based on radar reflectivity data, site wind speed observation data, and numerical model forecast data, the PhyDnet spatiotemporal convolutional grid model is used to obtain the third gust short-term forecast result based on capturing severe convective weather and extrapolation, specifically: S41. Constructing a deep learning instantaneous average wind forecast model based on the PhyDnet spatiotemporal convolutional grid model; S42. Establish a model for the relationship between instantaneous average wind and maximum wind using the least squares fitting statistical method for each site; S43. Download the numerical forecast products, station wind speed observation data and radar reflectivity data for the near future; S44, using an interpolation method to unify different data times into a target data time; S45, using the site wind speed observation and radar reflectivity as input, generating a site wind speed forecast for the target data time of the kth forecast evaluation standard time period; S46. Using multi-scale variational analysis, the generated site wind speed forecast is integrated with the model forecast; S47. Based on the established model of the relationship between instantaneous average wind and maximum wind, the instantaneous maximum wind data is predicted through the obtained fusion data, and the short-term forecast result of the third gust of wind based on capturing severe convective weather and extrapolation is output.

7. The method for short-term gust weather forecasting according to claim 1, characterized in that: In step S5, according to the determined kth forecast evaluation standard time period, based on HRCLDAS multi-source fusion data, site wind speed observation data, and numerical model forecast data, an extended orthogonal empirical decomposition is used to obtain the fourth gust short-term forecast result based on vector wind field propagation and extrapolation, specifically: S51. Construct an extended orthogonal empirical decomposition wind speed forecast model; S52, downloading the HRCLDAS multi-source fusion near-real-time wind field data and numerical forecast data and site wind speed observation data at the approaching time; S53, using time interpolation, interpolating the HRCLDAS and numerical forecast data to the time frequency of the observation data; S54. Based on multi-scale variational analysis, the interpolation results of HRCLDAS and numerical forecast data are corrected using station wind speed observations to obtain near-real-time grid wind speed data; S55. Input the obtained near-real-life grid wind speed data into the extended orthogonal empirical decomposition wind speed forecast model to obtain the fourth gust short-term forecast result based on vector wind field propagation and extrapolation.

8. The method for short-term gust weather forecasting according to claim 7, characterized in that: In the extended orthogonal empirical decomposition wind speed forecast model constructed in step S51, the typical gust excitation and propagation process is solved by matrix eigenvalues ​​and eigenvectors as follows: Among them, [U1...U T , V1...V T ] represents the abbreviation of the wind speed field covariance matrix of the sliding window, The time coefficient of the window wind field transmission process, EOF i represents the empirical orthogonal decomposition component, each component represents a wind field transmission process with a sliding window duration; i represents the grid point, T represents the matrix transpose, and ε represents the residual error.

9. The method for short-term gust weather forecasting according to claim 1, characterized in that: The evaluation weight expression in step S6 is: Gust fin =w1Gust1+w2Gust2+w3Gust3+w4Gust4 Among them, Gust fin is the gust short-term meteorological forecast result for the k+1th forecast evaluation standard time period; Gust1 is the first gust short-term forecast result, w1 is the weight coefficient of the first gust short-term forecast result; Gust2 is the second gust short-term forecast result, w2 is the weight coefficient of the second gust short-term forecast result; Gust3 is the third gust short-term forecast result, w3 is the weight coefficient of the third gust short-term forecast result; Gust4 is the fourth gust short-term forecast result, w4 is the weight coefficient of the fourth gust short-term forecast result.

10. The method for short-term gust weather forecasting according to claim 9, characterized in that: The values ​​of the weight coefficients w1, w2, w3 and w4 are 0-100%.

Images