Weather forecasting method and system, readable storage medium and program product

Abstract

The invention discloses a weather forecasting method and system, a readable storage medium and a program product. The weather forecasting method comprises the following steps: downloading a global analysis field and a global forecasting field; operating an artificial intelligence meteorological large model, and outputting global forecast data based on the global analysis field; operating a numerical weather forecast model, and generating a regional meteorological field based on the global forecast field; interpolating the global forecast data into the regional meteorological field to generate an interpolated regional meteorological field; operating the numerical weather forecast model, and generating an initial field and a boundary field of a corresponding region based on the interpolated regional meteorological field; and operating the numerical weather forecasting model, and performing ensemble forecasting based on the initial field and the boundary field. Therefore, the artificial intelligence large model can be applied in the numerical mode, the advantages of the two models are combined, and the accuracy of weather forecast is improved.

Description

Technical Field

[0001] The present application belongs to the field of meteorological numerical simulation, and specifically relates to a weather forecasting method, system, readable storage medium and program product. Background Art

[0002] Traditional weather forecasting relies on numerical weather prediction (NWP). These physics-driven models are based on an understanding of the nature of atmospheric processes. By inputting a three-dimensional initial meteorological field and numerically solving partial differential equations using high-performance computing clusters, they couple physical processes such as cloud microphysics, cumulus convection, and the boundary layer to predict weather conditions for the next few hours, days, or even months. However, these traditional numerical models consume significant computing power, and their accuracy decreases rapidly beyond seven days.

[0003] With the advancement of computing power and deep learning, researchers have begun exploring artificial intelligence (AI) weather forecasting methods. Based on AI model architectures such as visual transformer networks, graph neural networks, neural operator networks, and diffusion generative models, they have developed a series of large-scale AI meteorological models, including Pangu, GraphCast, FourCastNet, Fuxi, and Fengwu. These large-scale models have significantly increased their computational speed and are more capable of predicting both high-altitude and surface weather conditions, and even extreme weather such as typhoons.

[0004] However, large AI meteorological models are primarily trained based on reanalysis data such as ERA5, outputting global forecasts at a 0.25° angle. These models have low horizontal and vertical resolutions, lacking the ability to forecast regional small- and medium-scale weather, and are unable to accurately predict small- and medium-scale weather events such as localized rainstorms and thunderstorms. Furthermore, the output forecast variables only include limited physical quantities such as temperature, wind speed, and humidity, which cannot meet the needs of meteorological forecasting services. Furthermore, the training is based on L1 / L2 error constraints, which may lack physical consistency compared to traditional numerical models based on physical equations, resulting in non-physical characteristics such as energy spectrum deviations. Summary of the Invention

[0005] Technical problem to be solved by the invention

[0006] This application is formed to solve the above-mentioned technical problems, and its purpose is to provide a weather forecast method, system, readable storage medium and program product that can apply artificial intelligence large models in numerical models, combine the advantages of both, and improve the accuracy of weather forecasts.

[0007] Technical solutions used to solve technical problems

[0008] The present application provides a weather forecasting method, comprising: downloading a global analysis field and a global forecast field; running an artificial intelligence meteorological large model to output global forecast data based on the global analysis field; running a numerical weather forecast model to generate a regional meteorological field based on the global forecast field; interpolating the global forecast data into the regional meteorological field to generate an interpolated regional meteorological field; running the numerical weather forecast model to generate an initial field and a boundary field of a corresponding region based on the interpolated regional meteorological field; and running the numerical weather forecast model to perform an ensemble forecast based on the initial field and the boundary field.

[0009] Preferably, as the artificial intelligence meteorological big model, any one of the GraphCast, FourCastNet, Pangu, Fuxi, and Fengwu meteorological big models is selected.

[0010] Preferably, the artificial intelligence meteorological model is run on a GPU or an accelerator card.

[0011] Preferably, the high-resolution, medium- and small-scale WRF model is selected as the numerical weather forecast model.

[0012] Preferably, when running a numerical weather prediction model to generate a regional meteorological field based on the global forecast field, it includes: running a WPS to generate the regional meteorological field based on the global forecast field and static terrain data.

[0013] Preferably, when WPS is run and the regional meteorological field is generated based on the global forecast field and static terrain data, it includes: configuring WPS operation parameters to match the area, time, and resolution; reading the static terrain data, running geogrid.exe, and generating terrain data; reading the global forecast field, running ungrib.exe, and parsing the data of the global forecast field; based on the data output by ungrib.exe and the terrain data generated by geogrid.exe, running metgrid.exe to generate met_em*.nc files, that is, the regional meteorological field.

[0014] Preferably, when the global forecast data is interpolated into the regional meteorological field to generate an interpolated regional meteorological field, it includes: reading the regional meteorological field; reading the global forecast data output by the artificial intelligence meteorological model and extracting key meteorological variables; interpolating the key meteorological variables into the regional meteorological field and replacing the variables at the corresponding time to generate an interpolated regional meteorological field.

[0015] Preferably, the key meteorological variables include: latitudinal wind speed, longitudinal wind speed, temperature, humidity, geopotential height, surface temperature and surface pressure.

[0016] Preferably, NCL language is used for data input, output and interpolation, and the data is read and written in NetCDF format.

[0017] Preferably, when running the numerical weather prediction model to perform regional ensemble forecast, it includes: running an ensemble forecast based on random perturbations.

[0018] Preferably, the random perturbation includes adding SKEBS physical parameter perturbations to the geopotential, wind speed and air temperature respectively.

[0019] With this help, based on the physical equations of the NWP model, the output of the interpolated AI meteorological model can ensure high resolution and rich meteorological variables in small and medium-scale regions to the greatest extent (the amount of data of the first global forecast data is much larger than that of the second global forecast data), effectively solving the shortcomings of the AI meteorological model such as low resolution, few output variables, and lack of physical constraints. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 A flow chart of a weather forecasting method according to an embodiment of the present application is shown.

[0021] Figure 2 The simulation range set in this application and the path observation of the simulated case typhoon "Yutu" are shown.

[0022] Figure 3 The figure shows the forecast errors of global and regional forecasts for Typhoon Yutu's path, minimum sea level pressure and maximum wind speed, with the horizontal axis representing the forecast timeliness.

[0023] Figure 4 The observation data at 12:00 on October 21, 2018, the path forecast of the traditional regional forecast, and the path forecast of the regional forecast according to the present application are shown. DETAILED DESCRIPTION

[0024] The present application is further described below in conjunction with the following embodiments. It should be understood that the following embodiments are only used to illustrate the present application and are not intended to limit the present application. The same or corresponding reference numerals in the figures represent the same components, and repeated descriptions are omitted. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0025] Hereinafter, a weather forecast method according to an embodiment of the present application will be described with reference to the accompanying drawings.

[0026] The weather forecasting method proposed in this application aims to couple the numerically-based weather forecast model with an AI-based weather model, overcoming the shortcomings of both while leveraging their strengths to improve the accuracy of high-resolution ensemble forecasts. The specific steps are as follows:

[0027] First, in step S1, the global analysis field and the global forecast field are downloaded. Among them, the global analysis field (GlobalAnalysisField) is the optimal initial state of the atmosphere obtained by combining the numerical model with the observation data (such as satellites, radars, ground stations, etc.) at the start of the forecast. The global forecast field (GlobalForecastField) is the future weather forecast result starting from a certain analysis moment calculated by the numerical model based on the analysis field using physical equations. The global analysis field and the global forecast field can be obtained from the data server provided by the meteorological agency, for example. As the global analysis field, for example, it can include the reanalysis data (ERA5) generated by ECMWF, the reanalysis data (NCEP FNL) generated by NOAA, the initial field (GFS-FNL) generated by the GFS business system, etc. As the global forecast field, for example, the forecast data at the time corresponding to the global analysis field can be downloaded.

[0028] In addition, in this embodiment, the high-resolution mesoscale weather forecast model, WRF (Weather Research and Forecasting Model), is selected as the numerical weather prediction (NWP) weather forecast model. Specifically, WRF, as a mesoscale numerical weather model, is suitable for local and regional weather forecasts. It can analyze small and medium-scale weather phenomena such as rainstorms, thunderstorms, typhoons, hurricanes, and air pollution, and can achieve refined weather forecasts with high spatial resolution (e.g., 1 km level). However, numerical weather forecast models (NWP) are not limited to these and can be selected according to needs. For example, as global numerical weather forecast models, you can choose GFS (Global Forecast System), ECMWF (European Centre for Medium-Range Weather Forecasts), ICON (ICOsahedral Nonhydrostatic Model), etc. As regional numerical weather forecast models, in addition to WRF, you can also choose RAP (Rapid Refresh), HRRR (High-Resolution Rapid Refresh), AROME (Applications of Research to Operations at Mesoscale), COSMO (Consortium for Small-Scale Modeling), etc.

[0029] Next, in step S2, the AI meteorological model runs, taking the downloaded data file from the global analysis field (hereinafter referred to as global analysis data) as input, performing calculations, and outputting global forecast data. For clarity, the downloaded data file from the global forecast field will be referred to as the first global forecast data, and the global forecast data output by the AI meteorological model will be referred to as the second global forecast data.

[0030] Specifically, global analysis data is typically stored in formats such as GRIB2 or NetCDF. NetCDF data can be used directly. GRIB2 data can be parsed using tools such as wgrib2, Python, and pygrib, and converted to the input formats specified by large AI meteorological models, such as numpy and NetCDF.

[0031] In this embodiment, the artificial intelligence meteorological large model can be run on a GPU or an accelerator card, for example, and after inputting global analysis data and calculating, it outputs the second global forecast data. Preferably, the second global forecast data is converted into NetCDF data with a unified naming rule for subsequent direct use. As artificial intelligence meteorological large models, for example, GraphCast, Pangu-Weather (Pangu), FourCastNet, Fuxi, Fengwu, etc. can be selected. Since the artificial intelligence meteorological large model does not require complex physical solutions, compared with the traditional NWP model, the AI model can complete the prediction within a few seconds, while the NWP may take several hours to calculate, which not only greatly improves the computing power and efficiency, but also reduces the cost of the equipment.

[0032] Then, in step S3, the numerical weather prediction model is run to generate the required regional meteorological field based on the downloaded first global forecast data. In the present embodiment, WPS is run to generate the corresponding regional meteorological field based on the downloaded first global forecast data and static terrain data. Wherein, WPS (WRF Preprocessing System) is the preprocessing system of WRF (Weather Research and Forecasting Model), which is responsible for the preparation of terrain, land use, meteorological initial and boundary conditions. In other words, WPS is used to convert global data into a WRF usable format. Static terrain data can be provided by WPS itself or can be obtained from the official website. These static terrain data can include data such as terrain, altitude, land type, soil type, etc. In addition, in the WRF mode, regional meteorological field (Regional Meteorological Field) refers to the atmospheric state within a specific area, including the spatial distribution of meteorological elements such as temperature, humidity, air pressure, wind speed, and precipitation.

[0033] Specifically, in this embodiment, for example, WPS can be run as follows:

[0034] Configure WPS operating parameters, such as editing namelist.wps to match the corresponding region, time, resolution, etc. For example, you can set the WPS simulation time and region based on the weather system of interest to ensure that the life cycle of the weather system is covered;

[0035] Read static terrain data, run geogrid.exe, and generate terrain data, such as providing lower boundary conditions for the meteorological field according to the accuracy in the configuration table;

[0036] Read the first global forecast data, run ungrib.exe, parse the first global forecast data, such as decoding the binary file in GRIB2 format and reading key meteorological variables and reprojecting (output);

[0037] Based on the data output by ungrib.exe and the terrain data generated by geogrid.exe, run metgrid.exe to interpolate the data onto the WRF grid, generating met_em*.nc files, representing regional meteorological data (and therefore, regional meteorological fields). These met_em*.nc files are intermediate files used to convert global data into a regional model. They contain the initial meteorological field information required by the regional model (WRF) and serve as input data for the subsequent real.exe run to generate the initial and boundary fields.

[0038] Next, in step S4, key meteorological variables extracted from the second global forecast data are interpolated into the regional meteorological field at the corresponding time. The AI meteorological model outputs meteorological variables in the form of global grid data (e.g., 0.25°), while the WRF model requires high-resolution regional grid data (e.g., 1km). Therefore, interpolation of the second global forecast data is necessary to drive the WRF model.

[0039] Specifically, read the met_em*.nc files generated by WPS, which contain at least the grid and initial conditions required for WRF simulation;

[0040] The second global forecast data output by the artificial intelligence meteorological model is read. In step S2 above, the second global forecast data has been uniformly stored in NetCDF format, so the required key meteorological variables can be directly read and extracted. In this embodiment, the key meteorological variables that can be extracted include, for example, five three-dimensional variables: zonal wind speed, meridional wind speed, temperature, humidity, and geopotential height; and two two-dimensional variables: surface temperature and surface pressure.

[0041] Interpolate the extracted key meteorological variables into the WRF grid, for example, using Python xarray, scipy.interpolate.griddata, NCL (NCAR Command Language), etc.

[0042] Replace the variables in the above met_em*.nc file at the corresponding time with the interpolated data, generate a new met_em*.nc file and save it, that is, generate the interpolated regional meteorological field.

[0043] In this implementation, time interpolation is mainly used when interpolating key meteorological variables into met_em*.nc files. Specifically, the spatial structure of the regional meteorological field has been processed by WPS and has a high resolution, while the spatial resolution of the data of the large meteorological model is low (the amount of data is also relatively small). If spatial interpolation is performed, it may cause data distortion in complex terrain areas, making it difficult to maintain the consistency of the WRF grid, and may also cause the loss of local meteorological features. In addition, the temporal changes of meteorological variables are far more drastic than the spatial changes. Under the premise of ensuring temporal consistency, the spatial impact is small. Therefore, the use of time interpolation can ensure the correct propagation of important weather systems as much as possible, and can also reduce the time discretization error of the numerical model and improve the simulation accuracy. Therefore, interpolation is only performed spatially when necessary.

[0044] Next, in step S5, WPS runs to generate the initial and boundary fields for the regional model based on the new met_em*.nc file. Specifically, real.exe is run to convert the interpolated new met_em*.nc file into WRF format, generating the initial fields (wrfinput_d01) and boundary fields (wrfbdy_d01) for the WRF regional model. The initial fields (Initial Conditions) are the atmospheric variables required at the start of the WRF simulation (T=0). The boundary fields (BoundaryConditions) provide the atmospheric conditions outside the WRF simulation domain, constraining the atmospheric evolution within the model domain.

[0045] Next, in step S6, WRF is run to perform ensemble forecasting.

[0046] Specifically, run wrf.exe for numerical simulation, and use the initial field and boundary field as input data to drive the WRF regional model simulation. WRF ensemble forecasting (Ensemble Forecasting) mainly introduces uncertainty through three aspects: initial condition perturbation, boundary condition perturbation, and physical condition perturbation, in order to improve the reliability and robustness of the forecast. For example, by running multiple WRF instances (ensemble members), each ensemble member uses slightly different initial / boundary conditions or physical schemes, and runs a complete set of WRF processes (WPS+AI interpolation+real.exe+wrf.exe) for each ensemble member. After the run is completed, a result file is generated for each member and the ensemble mean / standard deviation is calculated. In this way, uncertainty information can be obtained and the reliability of the forecast can be improved.

[0047] More specifically, initial condition perturbations represent the uncertainty of the initial state of the atmosphere, boundary condition perturbations represent the uncertainty of the evolution of external weather systems, and physical condition perturbations represent the uncertainty of the model's physics scheme. Initial condition perturbations can be performed by providing different analysis data for each ensemble member, or by generating a different initial field (wrfinput_d01) for each ensemble member through error perturbations or random perturbations. This is acceptable as long as each ensemble member obtains different initial conditions. Boundary condition perturbations can be performed by providing different boundary data for each ensemble member, or by generating a different boundary field (wrfbdy_d01) through random perturbations. This is acceptable as long as each ensemble member obtains different boundary conditions. Physical condition perturbations are primarily performed when running wrf.exe (i.e., after initial and boundary condition perturbations). By modifying the physical scheme parameters in namelist.input, different physical process schemes (such as microphysics, convection, boundary layer, and radiation) are introduced. This allows each ensemble member to adopt a different combination of physical parameterization schemes, or by perturbing model parameters to increase the uncertainty of the numerical model. This generates a different wrfinput_d01 and / or wrfbdy_d01 file for each ensemble member. Using these files as input, we run multiple instances of wrf.exe, outputting result files for each ensemble member. Based on these result files, we calculate the ensemble mean and standard deviation, improving forecast confidence.

[0048] However, in this embodiment, step S6-1 can also be used instead of step S6. In step S6-1, a WRF ensemble forecast based on random perturbations is run. The AI meteorological large model is a deterministic forecast, and the boundary conditions after replacing key meteorological variables cannot provide perturbations. However, small and medium-scale physical processes have a significant impact on the simulation results. Therefore, the WRF integrated random perturbation method is used to introduce uncertainty and enhance the performance of the ensemble forecast.

[0049] Specifically, in this implementation, SKEBS (Stochastic Kinetic Energy Backscatter Scheme) is used as a stochastic perturbation scheme. SKEBS can simulate the backflow of subgrid-scale energy in the atmosphere, enhancing the diversity of ensemble forecasts. In other words, SKEBS is a type of physical perturbation.

[0050] The SKEBS perturbation is added to the core meteorological variables, such as geopotential, wind speed, and temperature, as follows:

[0051]

[0052] where ψ′(x, y, t) represents the two-dimensional stream function perturbation; k represents the meridional wave number, l represents the zonal wave number; e 2 πi(kx / X+ly / Y) Represents the standard waveform. Therefore, the disturbance at a certain point in space is equal to the superposition of waves of different frequencies in the longitudinal and latitudinal directions.

[0053]

[0054]

[0055] Among them, u′(x, y, t) represents the longitudinal wind speed; v′(x, y, t) represents the latitudinal wind speed.

[0056] The resulting wrfout_d01 and / or wrfbdy_d01 files for each ensemble member contain the SKEBS perturbation simulation results. Multiple instances of wrf.exe are then run based on these wrfout_d01 and / or wrfbdy_d01 files, outputting result files for each ensemble member. The ensemble mean and standard deviation are then calculated based on these result files to improve forecast confidence.

[0057] This application couples an artificial intelligence meteorological large model with a high-resolution numerical model, organically combining the advantages of both, thereby solving the shortcomings of the AI meteorological large model, such as low resolution, few output variables, and lack of physical constraints, while also solving the shortcomings of NWP, such as time-consuming calculations and lack of accuracy in long-term forecasts.

[0058] Specifically, traditional numerical weather forecasts (such as WRF, GFS, ECMWF) mainly rely on physical models (based on fluid mechanics, thermodynamics, and atmospheric dynamics equations), and supercomputers take several hours to run. The artificial intelligence meteorological large model (such as GraphCast, Pangu-Weather, FourCastNet, Fuxi) is a data-driven mathematical model that is completely based on machine learning. It extracts patterns from historical weather data through large-scale data training (deep learning), learns the spatiotemporal evolution of atmospheric variables, skips the calculation of traditional physical equations, and directly predicts future weather conditions. Therefore, the AI meteorological large model is much faster than NWP and can predict global weather within seconds. However, due to the lack of physical constraints in the AI meteorological large model, false weather phenomena may occur. Based on the above, this application uses the AI meteorological large model to generate a second global forecast data based on global analysis data, uses WPS to generate a regional meteorological field based on the first global forecast data, and inserts key meteorological variables extracted from the second global forecast data into the regional meteorological field, thereby generating an initial field and boundary field adjusted by mathematical interpolation.

[0059] With this, based on the physical equations of the NWP model, the output of the interpolated AI meteorological model can ensure high resolution and rich meteorological variables in small and medium-scale regions to the greatest extent (the amount of data of the first global forecast data is much larger than that of the second global forecast data), effectively solving the shortcomings of the AI meteorological model with low resolution and few output variables. In addition, the rapid fusion of multi-source data by the AI meteorological model is not only efficient but also highly accurate in analysis, so that the initial field and boundary field generated after interpolation by the AI meteorological model are of high precision, which can optimize regional model simulation. In addition, traditional numerical forecasts rely on NWP model driving, and errors will amplify over time. By effectively utilizing the deep learning capabilities of the AI meteorological model, such as extracting rules based on large amounts of data training, capturing long-term trends, and data training based on historical errors, extreme weather can be better captured, so that its output data can be interpolated to the regional grid and replaced to improve the long-term forecast accuracy and forecast reliability. In addition, since the interpolation processing of the AI meteorological large model is performed after generating the regional meteorological field and before generating the initial field and boundary field, the physical model can be adjusted again after the interpolation of the mathematical model on the basis of converting the global model into the regional model, thereby greatly reducing the non-physical impact brought by the AI meteorological large model.

[0060] Furthermore, as mentioned above, the boundary conditions after replacing the key meteorological variables cannot provide disturbances, resulting in the boundary field after mathematical interpolation having non-physical characteristics compared to before interpolation. Therefore, this application adopts SKEBS perturbation to increase the disturbance of physical tendencies to increase physical constraints, thereby reducing the impact of mathematical interpolation on physical equations, which can further improve the forecasting capability after coupling and provide decision support for forecasting services.

[0061] The following, combined Figures 2 to 4 , which is described in more detail through specific embodiments.

[0062] In this embodiment, the simulation is performed for the 2018 typhoon Yutu.

[0063] Figure 2 The simulation range set in this application and the observed path of the simulated Typhoon Yutu are shown. TD stands for tropical depression, TS for tropical storm, STS for severe tropical storm, TY for typhoon, STY for severe typhoon, and SUPERTY for super typhoon. These are the intensity classification standards for tropical cyclones (typhoons) in the Northwest Pacific and South China Sea regions.

[0064] Global forecasts use the Global Forecasting System (GFS) model, with an eight-day forecast launched every 12 hours between 12:00 UTC on October 20 and 12:00 UTC on October 22. Download global analysis data and the corresponding GFS global model forecast data (i.e., the first global forecast data). This data is available for real-time download and can be used for forecasting operations.

[0065] The Pangu Meteorological Model and the Fuxi Meteorological Model are deployed and run on accelerator cards (GPU / TPU). The horizontal resolution of both models is 0.25° (about 25km), and global forecast data (i.e., the second global forecast data) are generated based on global analysis data.

[0066] The regional forecast uses the WRF model, and the 8-day forecast is activated every 12 hours between 12:00 on October 20 and 12:00 on October 22. Figure 2 As shown, the WRF simulation domain is set. Specifically, the d01 domain is set to 9 km resolution, with a grid size of 586 × 391, encompassing China and the northwest Pacific to resolve large-scale weather patterns. The number of vertical layers is set to 57, and the model top is set to 10 hPa to fully cover the flight altitudes required for civil aviation meteorological support. These settings can of course be adjusted as needed.

[0067] In this embodiment, NCL language is used for data input, output and interpolation, and the NetCDF format is used for both reading and writing.

[0068] Set 20 ensemble members, and configure the ensemble perturbation amplitude and other related parameters as follows:

[0069] tot_backscat_psi=1.0E-05;

[0070] tot_backscat_t=1.0E-06;

[0071] ztau_psi=10800.0;

[0072] ztau_t=10800.0;

[0073] rexponent_psi=-1.83;

[0074] rexponent_t = -1.83;

[0075] zsigma2_eps=0.0833;

[0076] zsigma2_eta=0.0833.

[0077] The physical parameterization schemes used by WRF include:

[0078] Microphysics scheme: WSM6 (WRF single-moment 6-class);

[0079] Radiation scheme: RRTM-G (Rapid Radiative Transfer Mode);

[0080] Boundary Layer Program: YSU (Yonsei University Planetary Boundary Layer Program);

[0081] Land surface model: NOAH (unified Noah land surface model);

[0082] Cumulus parameterization scheme: Tiedtke scheme.

[0083] Figure 3 The following table shows the forecast errors for Typhoon Yutu's track, minimum sea level pressure, and maximum wind speed using three global forecasts (GFS, Pangu, and Fuxi) and three regional forecasts (WRF_GFS, WRF_Pangu, and WRF_Fuxi). The vertical axis represents the track error (Track RMSE), the minimum sea level pressure error (SLP RMSE), and the maximum wind speed error (MaximalWind RMSE), respectively. The horizontal axis represents the forecast time (Time). WRF_GFS represents the regional forecast using a traditional numerical model, while WRF_Pangu and WRF_Fuxi represent the regional forecasts based on the method proposed in this application.

[0084] Figure 4 The figure shows the path forecasts for the observation data (OBS), the traditional regional forecast (WRF_GFS), and the regional forecasts (WRF_PANGU and WRF_FUXI) according to this application at 12:00 on October 21, 2018. The ensemble path forecast is represented by a relatively light solid line, the deterministic path forecast is represented by a dotted line, and the ensemble mean path forecast is represented by a relatively dark solid line.

[0085] like Figure 3 、 4 As shown in the figure, based on global model forecast results, the forecast error of GFS exceeds 3,000 km in about 7 days, and the forecast error of WRF_GFS also reaches 2,000 km in about 7 days. The path forecasts of Pangu and Fuxi are better than those of GFS, but the global model resolution is low, so the typhoon intensity (minimum sea level pressure, maximum wind speed) forecast is poor. After coupling the AI meteorological large model with WRF, WRF_PANGU and WRF_FUXI retain the advantages of global forecasts and significantly improve the intensity forecast skills, indicating that the coupled model has strong capabilities in multi-scale forecasting. Furthermore, after adding the SKEBS perturbation, the ensemble mean improves compared to the deterministic forecast, and at the same time provides a probability distribution of the path, which can support forecasters in making decisions and improve forecast accuracy.

[0086] It can be seen that after coupling the AI meteorological large model with the mesoscale numerical weather model WRF, it can not only give play to the advantages of AI's fast calculation and WRF's physical consistency, but also provide refined and high-resolution weather forecasts, thereby improving the accuracy and reliability of weather forecasts.

[0087] The functions of the elements disclosed in this specification may be performed using circuits or processing circuits comprising general-purpose processors, special-purpose processors, integrated circuits, ASICs (Application Specific Integrated Circuits), existing circuits and / or combinations thereof that are constructed or programmed to perform the disclosed functions. A processor is considered a processing circuit or circuit because it contains transistors or other circuits. In this disclosure, a circuit, unit or means is hardware that performs the listed functions, or hardware that is programmed to perform the listed functions. The hardware may be the hardware disclosed in this specification, or other known hardware that is programmed or constructed to perform the listed functions. When the hardware is considered to be a processor that is a type of circuit, the circuit, means or unit is a combination of hardware and software, and the software is used in the construction of the hardware and / or processor.

[0088] As another aspect, the present invention further provides a computer-readable storage medium, which may be included in the computer device described in the above embodiments and their variations, or may exist independently and not incorporated into the computer device. The computer-readable storage medium carries one or more programs. When executed by the computer device, the computer device implements the method described in the above embodiments and their variations. For example, the computer device may implement the steps shown in the various figures.

[0089] According to one aspect of the present invention, a computer program product is provided, comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the methods provided in various optional implementations of the above-described embodiments.

[0090] The above specific implementation methods further describe the purpose, technical solutions and beneficial effects of this application in detail. It should be understood that the above is only a specific implementation method of this application and is not limited to the scope of protection of this application. Without departing from the purpose of the basic characteristics of this application, this application can be embodied in various forms. Therefore, the implementation forms in this application are used for illustration rather than limitation. Since the scope of this application is defined by the claims rather than the specification, and all changes that fall within the scope defined by the claims or the equivalent range of the scope defined by them should be understood to be included in the claims. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of this application should be included in the scope of protection of this application.

Claims

1. A weather forecast method, characterized in that: include: Download global analysis fields and global forecast fields; Running an artificial intelligence meteorological large model to output global forecast data based on the global analysis field; running a numerical weather prediction model to generate a regional meteorological field based on the global forecast field; interpolating the global forecast data to the regional meteorological field to generate an interpolated regional meteorological field; Running the numerical weather forecast model to generate an initial field and a boundary field of the corresponding region based on the interpolated regional meteorological field; The numerical weather prediction model is run to perform ensemble forecast based on the initial field and the boundary field.

2. The weather forecast method according to claim 1, characterized in that: As the artificial intelligence meteorological big model, any one of the GraphCast, FourCastNet, Pangu, Fuxi, and Fengwu meteorological big models is selected.

3. The weather forecast method according to claim 2, characterized in that: Run the artificial intelligence meteorological model on a GPU or accelerator card.

4. The weather forecast method according to claim 1, wherein: As the numerical weather forecast model, the high-resolution, medium- and small-scale WRF model was selected.

5. The weather forecast method according to claim 4, characterized in that: When the numerical weather prediction model is run to generate the regional meteorological field based on the global forecast field, it includes: running WPS to generate the regional meteorological field based on the global forecast field and static terrain data.

6. The weather forecast method according to claim 5, characterized in that: When the WPS is run to generate the regional meteorological field based on the global forecast field and static terrain data, it includes: Configure WPS operation parameters to match the region, time, and resolution; Read the static terrain data, run geogrid.exe, and generate terrain data; Read the global forecast field, run ungrib.exe, and parse the data of the global forecast field; Based on the data output by ungrib.exe and the terrain data generated by geogrid.exe, run metgrid.exe to generate met_em*.nc files, which are the meteorological fields of the region.

7. The weather forecast method according to claim 1, characterized in that: When interpolating the global forecast data to the regional meteorological field to generate the interpolated regional meteorological field, the method includes: Reading the regional meteorological field; Reading the global forecast data output by the artificial intelligence meteorological model and extracting key meteorological variables; The key meteorological variables are interpolated into the regional meteorological field and replace the variables at the corresponding time to generate an interpolated regional meteorological field.

8. The weather forecast method according to claim 7, characterized in that: The key meteorological variables include at least: latitudinal wind speed, longitudinal wind speed, temperature, humidity, geopotential height, surface temperature and surface pressure.

9. The weather forecast method according to claim 1, wherein: NCL language is used for data input, output and interpolation, and NetCDF format is used for data reading and writing.

10. The weather forecast method according to claim 1, characterized in that: When the numerical weather forecast model is run to perform regional ensemble forecast, it includes: running an ensemble forecast based on random perturbations.

11. The weather forecast method according to claim 10, characterized in that: The random perturbation includes a scheme of adding SKEBS physical parameter perturbations to the potential, wind speed and air temperature respectively.

12. A weather forecast system comprising a memory, a processor, and a computer program stored in the memory, wherein: The processor executes the computer program to implement the weather forecasting method according to any one of claims 1 to 11.

13. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the weather forecasting method according to any one of claims 1 to 11 is implemented.

14. A computer program product comprising a computer program, characterized in that When the computer program is executed by a processor, the weather forecasting method according to any one of claims 1 to 11 is implemented.

Images