A method, medium, and computer program for typhoon intensity prediction based on AI initial field-driven NWP regional models and fusion with vortex dynamic initialization.

Abstract

This invention discloses a typhoon intensity prediction method, medium, and computer program based on an AI initial field-driven NWP regional model and fused with vortex dynamic initialization. Belonging to the field of artificial intelligence and regional numerical simulation technology, it aims to improve the accuracy of intensity prediction and structure reconstruction capability for rapidly intensifying typhoons. The method first constructs an AI weather model based on deep learning, generating continuous initial and boundary fields covering multiple meteorological elements to drive the regional NWP model. Second, a vortex dynamic initialization process is introduced before forecasting to enhance the consistency and intensity accuracy of the typhoon's axisymmetric structure in the simulated initial field. Subsequently, integral simulation is conducted based on an atmosphere-ocean bidirectional coupling model, and local structure dynamic optimization is achieved by assimilating high-frequency multi-source observation data. Finally, an AI-NWP residual feedback mechanism is constructed to achieve closed-loop correction of prediction errors and model fine-tuning. This invention is suitable for high-precision typhoon intensity prediction and has good engineering applicability.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and regional numerical simulation technology, and relates to high-precision modeling and prediction of typhoon intensity evolution. Specifically, it is a typhoon intensity prediction method, medium and computer program based on AI initial field driven NWP regional model and fused with vortex dynamic initialization, which can improve the intensity prediction accuracy and structure restoration capability of strong typhoons in the rapid intensification stage. Background Technology

[0002] As a typical tropical cyclone system, typhoons exhibit particularly complex intensity changes, influenced by a combination of atmospheric, oceanic, and topographic factors. Especially during the rapid intensification (RI) phase, they display extremely high nonlinearity and uncertainty. Accurate typhoon intensity prediction is not only crucial for disaster prevention and mitigation but also represents one of the key technical challenges currently facing numerical weather prediction (NWP). Traditional global NWP models typically use static data assimilation methods to generate initial fields. Limited by resolution, dynamic framework, and parameterization schemes, these models struggle to accurately characterize key features such as the typhoon's eyewall structure, low-level wind intensity, and mid-level warm core structure. This results in a systematic underestimation of typhoon intensity in traditional NWP models, leading to particularly weak predictions.

[0003] In recent years, artificial intelligence (AI)-driven meteorological forecasting methods have developed rapidly, and some AI global weather forecasting models have demonstrated superior performance in short- and medium-term weather element prediction. However, because AI models are inherently regression-based, they suffer from intensity smoothing and structure diffusion issues, leading to significant biases in the intensity prediction of severe weather systems such as typhoons, particularly in terms of near-surface wind speed (10m), minimum central pressure, and vortex axisymmetric structure. In existing technologies, Chinese patent CN119202880A discloses a method for rapid identification of long-distance typhoon heavy rainfall based on an AI meteorological model. This method analyzes the long-distance precipitation effect by removing the typhoon core vortex and constructing sensitivity experiments, but its technical solution does not address the systematic underestimation problem of AI models in typhoon intensity prediction. CN108983320A trains a deep neural network using numerical weather prediction models and measured data, and combines this with mesoscale meteorological numerical models for downscaling calculations to predict extreme wind speeds. However, this method focuses on predicting extreme wind speeds, and the coupling method between the AI ​​model and the numerical model is relatively simple, making it difficult to fully leverage the advantages of AI models in large-scale circulation prediction. CN119106249A uses an AI model to generate simulated typhoon samples to assess the impact of climate change on extreme typhoon winds, focusing on long-term climate impact assessment rather than short-term or medium-term typhoon intensity prediction, thus failing to fully utilize the advantages of AI models in processing complex meteorological data.

[0004] To compensate for the limitations of global models in local weather simulation, regional high-resolution numerical weather prediction models, such as the Weather Research and Forecasting Model (WRF), are widely used for high-resolution reproduction and intensity evolution simulation of typhoon processes. By setting nested grids, local physical parameter schemes, and initialization processing of vortex structures, the WRF model can better capture and reproduce the internal dynamic structure of typhoons. However, the simulation results of regional numerical models are highly dependent on the quality of their initial and boundary conditions. Traditional initial fields provided by the global National Weather Service (NWP) are still insufficient in simulating typhoon intensity and struggle to capture the true three-dimensional atmospheric structure. While AI weather prediction has advantages in large-scale circulation prediction and the WRF model performs well in simulating fine-scale processes, there is currently a lack of mature technical solutions for effectively integrating the advantages of both to improve the accuracy of typhoon intensity prediction. In particular, how to use the prediction results of AI models as driving conditions for regional models, and how to optimize vortex initialization techniques for the characteristics of AI prediction results, remain technological gaps.

[0005] In summary, while existing AI meteorological models possess strong large-scale circulation prediction capabilities, they exhibit significant weaknesses in characterizing typhoon intensity. Traditional WRF models, despite their high resolution and structural simulation capabilities, still suffer from significant errors due to limitations in initial conditions and boundary field quality. Therefore, effectively integrating AI weather forecasting techniques with regional numerical weather models and combining them with vortex dynamic initialization techniques to construct a high-precision typhoon intensity prediction method is a pressing technical challenge. Summary of the Invention

[0006] (a) Purpose of the invention

[0007] To address the aforementioned deficiencies and shortcomings of existing technologies, this invention aims to provide a typhoon intensity prediction method, medium, and computer program based on an AI-driven NWP regional model and fused with vortex dynamic initialization. By employing an AI meteorological prediction model to generate high-quality initial fields and boundary conditions, this method drives a regional numerical weather prediction model and superimposes the vortex dynamic initialization process, enhancing the dynamic consistency and intensity fitting capability of the initial vortex structure. This achieves high-precision simulation and structural reproduction of the intensity evolution of strong typhoons, especially during their rapid intensification phase. This method fully integrates the advantages of AI models in capturing mesoscale circulation with the capabilities of high-resolution NWP models in local structure analysis, significantly improving the simulation accuracy of regional models for key parameters such as maximum wind speed, central pressure, and eyewall structure of typhoons. This effectively addresses the systematic underestimation problem in existing prediction systems for typhoon intensity prediction.

[0008] (II) Technical Solution

[0009] To achieve the objective of this invention and solve its technical problems, the present invention adopts the following technical solution:

[0010] The first objective of this invention is to provide a typhoon intensity prediction method based on an AI initial field-driven NWP regional model and fused with vortex dynamic initialization, to improve the intensity prediction accuracy and structure reconstruction capability of tropical cyclone systems, especially typhoons, during their rapid intensification phase. The method includes the following steps:

[0011] S100. Construct an AI weather model and generate the initial and boundary fields:

[0012] An AI weather model is built and trained using a deep learning architecture based on historical reanalysis data. The AI ​​weather model includes multiple deep learning sub-networks to support different forecast lead times. It uses an Earth spherical sensing structure for input encoding and outputs multi-leader continuous forecast results covering upper-air meteorological elements, which serve as the initial and boundary fields for subsequent NWP regional models.

[0013] S200. Construct the NWP region model and perform vortex dynamic initialization processing:

[0014] A multi-layered nested NWP regional model computational domain is constructed, in which the intermediate and inner domains are set as adaptive tracking moving grids based on the typhoon vortex center position; and a vortex dynamic initialization process is introduced in the preprocessing stage before the report is initiated. By extracting the axisymmetric vortex structure of the typhoon in the AI ​​initial field, spectral perturbation enhancement and position relocation are performed to construct a dynamic initial field with intensity fitting ability and dynamic consistency.

[0015] S300. Initiate AI-driven atmospheric-oceanic two-way coupled model integral simulation:

[0016] The NWP regional atmospheric model is started based on the initial dynamic field after vortex initialization, and the ocean model is started simultaneously for coupled integral simulation. In each integral time step, the ocean model updates key underlying surface variables, including at least sea surface temperature, latent heat flux, and sensible heat flux, in real time, and uses them as feedback input to the atmospheric model to dynamically correct the sea surface boundary conditions along the typhoon's path.

[0017] S400. Real-time assimilation and correction of multi-source observation data to optimize the simulation process:

[0018] During the integration of the atmospheric-oceanic regional coupled model, multi-source near-surface observation data are dynamically introduced based on a predetermined time step. A data assimilation algorithm is used to perform local high-frequency frequency value assimilation correction on the typhoon eyewall and its adjacent areas, and the local wind field and temperature field status are dynamically updated.

[0019] S500. Construct an AI-NWP residual closed-loop feedback correction mechanism:

[0020] During the regional coupled model integration process, the error residuals between the simulation results and the typhoon observation data are periodically acquired. The error information is then fed back into the AI ​​weather model as a feedback input by constructing a residual driving network. This serves as a fine-tuning signal to correct the initial field output of the AI ​​weather model for the next prediction cycle, thereby achieving dynamic two-way closed-loop coupling between the AI ​​weather model and the NWP regional model.

[0021] S600. Output and evaluate the typhoon intensity simulation results:

[0022] After completing the full-time integration, the key parameters of typhoon intensity are output, and the structural consistency evaluation is carried out in combination with the actual observation data of the typhoon to complete the quantitative prediction and analysis of typhoon intensity.

[0023] The second objective of this invention is to provide a computer program product, including computer instructions, which are used to execute the above-mentioned typhoon intensity prediction method based on AI initial field-driven NWP regional model and fused with vortex dynamic initialization.

[0024] The third objective of this invention is to provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-mentioned typhoon intensity prediction method based on AI initial field-driven NWP regional model and fused with vortex dynamic initialization.

[0025] (III) Technical Effects

[0026] Compared with the prior art, the present invention has the following significant technical effects: (1) The present invention generates a continuous and timely high-dimensional meteorological initial field and boundary field through AI weather model, and constructs an NWP regional model based on AI initial field, which significantly improves the refinement and timeliness of initial conditions; at the same time, combined with the vortex dynamic initialization strategy, the axisymmetric circulation structure of the typhoon is extracted and reconstructed, which enhances the intensity fitting and dynamic consistency of the typhoon initial field, effectively improves the problem of insufficient early structure characterization of rapidly intensifying typhoons in traditional models, and improves the accuracy and physical consistency of intensity simulation. (2) The present invention constructs an AI-NWP two-way closed-loop coupling system, introduces an AI-NWP residual feedback mechanism in the integration process of the regional atmospheric-ocean coupled model, realizes dynamic identification and real-time correction of simulation errors; in conjunction with the high-frequency multi-source observation data assimilation process, local high-resolution wind field and temperature field updates are implemented for key structural areas such as the typhoon eyewall, which strengthens the constraint capability of key physical processes in the typhoon evolution process and significantly improves the forecast accuracy and timeliness stability of typhoon path, intensity and structural evolution. Attached Figure Description

[0027] Figure 1 The diagram shows the implementation flowchart of the typhoon intensity prediction method based on AI initial field-driven NWP regional model and fused with vortex dynamic initialization of the present invention.

[0028] Figure 2 The figure shows a comparison of the distribution of the difference in the initial field validation column integral water vapor flux between the AI ​​weather model and the ECMWF model. (a) represents the difference between the AI ​​weather model and ERA5 (AI-ERA5); (b) represents the difference between the ECMWF model and ERA5 (EC-ERA5).

[0029] Figure 3 The figure shows a comparison of the effects before and after vortex dynamic initialization. In the figure: (a) is the vortex structure predicted by the AI ​​weather model; (b) is the enhanced vortex structure after vortex dynamic initialization.

[0030] Figure 4 The figure shows a comparison and analysis of the typhoon's path and intensity. In the figure: (a) is a comparison between the path simulation results and the optimal path; (b) is a comparison of the evolution of the maximum wind speed within 10 meters of the outer perimeter. Detailed Implementation

[0031] This invention aims to provide a typhoon intensity prediction method, medium, and computer program based on an AI initial field-driven NWP regional model and fused with vortex dynamic initialization, to improve the intensity prediction accuracy and structure reconstruction capability of tropical cyclone systems, especially typhoons, during their rapid intensification phase. To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions in the embodiments of this invention will be described in more detail below with reference to the accompanying drawings. The described embodiments are some, but not all, embodiments of this invention, and are exemplary, intended to explain the invention, and should not be construed as limiting the invention.

[0032] Example 1: Typhoon Intensity Prediction Method

[0033] The typhoon intensity prediction method provided in this invention is based on an AI initial field-driven NWP regional model and integrates vortex dynamic initialization. Figure 1 As shown, its implementation mainly includes the following steps:

[0034] S100. Construct an AI weather model and generate the initial and boundary fields:

[0035] An AI weather model is built and trained using a deep learning architecture based on historical reanalysis data. The AI ​​weather model includes multiple deep learning sub-networks to support different forecast lead times. The input is encoded using an Earth spherical sensing structure, and the output is a multi-leader continuous forecast result covering upper-air meteorological elements, which serves as the initial field and boundary field for subsequent NWP regional models.

[0036] In this embodiment of the invention, the AI ​​weather model adopts an encoder-decoder structure during its construction. The input end uses an Earth spherical sensing encoding structure to process globally gridded input data, and explicitly enhances the latitude and longitude information in the input data through a location embedding mechanism to improve the model's ability to express the changes in tropical cyclone location and geographical dependence characteristics. The AI ​​weather model enhances its cross-temporal dependence modeling capability through a deep residual connection structure, and generates multiple continuous upper-air meteorological variable fields with prediction time at the output end. Its input variables include at least upper-air variables such as atmospheric temperature, air pressure, specific humidity, zonal wind speed, and geopotential height. The missing surface variables are supplemented and reconstructed through external reanalysis data or ocean models to ensure that a complete initial field and boundary conditions are provided for subsequent NWP regional models.

[0037] It should be noted that the training dataset for the AI ​​weather model uses ERA5 reanalysis data, covering global atmospheric reanalysis data from 1979 to 2023, with a spatial resolution of 0.25°×0.25° and a temporal resolution of 1 hour, ensuring the model's ability to generalize to different climate regions and seasonal variations globally. During model training, the sample weights for typhoon-active areas in the Northwest Pacific were significantly enhanced. Data augmentation techniques were used to generate more typhoon-related training samples, improving the AI ​​weather model's ability to identify and forecast typhoon systems. Furthermore, considering the limitations of AI weather models regarding surface variables, this invention supplements the model by fusing external ocean model data and surface reanalysis data, ensuring a complete three-dimensional atmospheric state field is provided to the regional model. Simultaneously, to improve the AI ​​weather model's adaptability to different typhoon development stages, a multi-source historical typhoon sample set was introduced during model training, and its sample distribution range was expanded through data augmentation strategies, enhancing its generalization and prediction capabilities for rapidly intensifying, path-changing, and multi-vortex structure typhoons. A multi-objective loss function was also employed during model training to balance prediction errors among different variables and improve the dynamic consistency of the overall initial field.

[0038] S200. Construct the NWP region model and perform vortex dynamic initialization processing:

[0039] A multi-layered nested NWP regional model computational domain is constructed, in which the intermediate and inner domains are set as adaptive tracking moving grids based on the typhoon vortex center position; and a vortex dynamic initialization process is introduced in the preprocessing stage before the report is initiated. By extracting the axisymmetric vortex structure of the typhoon in the AI ​​initial field, spectral perturbation enhancement and position relocation are performed to construct a dynamic initial field with intensity fitting ability and dynamic consistency.

[0040] As a preferred approach, the NWP regional model selects the WRF model and adopts a three-layer nested structure, including an outer domain, an intermediate domain, and an inner domain. The outer domain has a grid spacing of 15-20 km to provide large-scale circulation background conditions; the intermediate domain has a grid spacing of 5-8 km to analyze the mesoscale circulation structure; and the inner domain has a grid spacing of 500 m-2 km to simulate the typhoon eyewall, rainband, and local severe convective structure with high accuracy. The intermediate and inner domains are set as typhoon vortex adaptive tracking moving grids, which automatically adjust the coordinates of the computational domain center every 1-10 minutes according to the typhoon center position to ensure that the high-resolution grid domain always covers the typhoon main vortex region. A two-way feedback mechanism is adopted between the three nested layers to ensure that the structural information of the fine grid domain is transmitted upward to the coarse resolution region, enhancing the coordination and dynamic consistency of the entire simulation.

[0041] In this embodiment of the invention, a vortex dynamics initialization process is introduced in the preprocessing stage before the report is initiated. This process preferably includes the following sub-steps:

[0042] S201. Extracting the axisymmetric vortex structure of the typhoon: Based on the initial field generated by the AI ​​weather model, extract the key structural variables of the main vortex region of the typhoon, including the geopotential height field, zonal wind field, temperature field and specific humidity field on the isobaric surface, extract the vortex circulation structure with axisymmetric characteristics, remove the asymmetric disturbance components, and construct an idealized initial structural framework of the typhoon.

[0043] S202. Perform spectral embedding enhancement simulation: embed the vortex structure into the initial background field of the NWP regional model, use spectral embedding technology to constrain the large-scale background field, and release the energy of small and medium-scale disturbances. Perform a 6-hour short-time pre-integration simulation to enhance the typhoon core intensity, eyewall structure and symmetry.

[0044] S203. Implement spatial repositioning of vortex structure: After the short-time integration is completed, based on the deviation between the actual observation of the typhoon center and the typhoon center position in the simulated field, the vortex extraction and displacement function interpolation algorithm is used to spatially reposition and re-embed the simulated vortex structure to ensure the spatial consistency of the initial dynamic field.

[0045] S204. Constructing the final initial dynamic field: The vortex simulation results after structural enhancement and position correction are weighted and fused with the background field of the regional model to form an initial dynamic field that takes into account both large-scale accuracy and local vortex intensity. This initial field serves as the starting input for the regional model's integral simulation, improving the reliability and structural reconstruction capability of subsequent typhoon intensity simulations.

[0046] It is worth noting that the vortex dynamic initialization process is one of the core technological innovations of this invention. This process effectively solves the problems of weak typhoon intensity and incomplete structure output by AI weather models. The spectral perturbation enhancement technology, through 6 hours of dynamic integration, not only enhances the axisymmetric circulation intensity of the typhoon, but more importantly, maintains the dynamic and thermodynamic equilibrium of the vortex structure, avoiding numerical instability that may be caused by traditional artificial enhancement methods. The vortex relocation algorithm adopts a dual positioning mechanism based on the vorticity center and the pressure center, ensuring that the enhanced typhoon vortex accurately corresponds to the actual observation location, providing a high-quality dynamically consistent initial field for subsequent coupled simulations.

[0047] S300. Initiate AI-driven atmospheric-oceanic two-way coupled model integral simulation:

[0048] The NWP regional atmospheric model is launched based on the initial dynamic field after vortex initialization, and the ocean model is launched simultaneously for coupled integral simulation. In each integral time step, the ocean model updates key underlying surface variables in real time, including at least sea surface temperature, latent heat flux, and sensible heat flux, and uses them as feedback inputs to the atmospheric model to dynamically correct the sea surface boundary conditions along the typhoon's path.

[0049] As a preferred method, the ocean model is either ROMS or HYCOM, with the grid resolution set to match the inner domain of the atmospheric model. Its computational depth is extended to 1000 meters to fully cover the upper mixing layer and thermocline structure. The ocean state parameters along the typhoon path are refreshed frequently. During the integration process, it is bidirectionally coupled with the NWP regional atmospheric model through a coupling interface. The coupling variables include at least sea surface temperature, latent heat flux, sensible heat flux, and 10-meter wind stress. The parameters are dynamically fed back to the NWP atmospheric model to update its boundary layer physical parameters and dynamically correct the sea surface thermal boundary conditions along the typhoon's path.

[0050] In this embodiment of the invention, the AI-driven atmospheric-oceanic bidirectional coupled model integral simulation is further subdivided into the following sub-steps during implementation:

[0051] S301. Coupled System Synchronous Initialization: Based on the initial atmospheric dynamic field after vortex initialization and the initial temperature and salinity field of the ocean model, the initial state matching of the atmosphere-ocean interface is performed through the coupler, and consistency constraints of sea surface temperature, sea surface height and 10-meter wind field are established to ensure that the physical fields of the atmospheric model and the ocean model are coordinated and unified at the interface boundary when the coupled system is started.

[0052] S302. Real-time calculation of air-sea interface fluxes: In each integral time step, the ocean model calculates the momentum flux, sensible heat flux, latent heat flux and net radiation flux of the air-sea interface in real time. The flux exchange coefficient is dynamically determined based on the difference between sea surface temperature and atmospheric temperature, 10-meter wind speed and relative humidity, and a nonlinear flux parameterization scheme is established.

[0053] S303. Calculation of ocean dynamic response: The ocean model calculates the Ekman transport, vertical mixing and upwelling processes of the ocean surface in real time based on the wind stress provided by the atmospheric model. It simulates the deepening of the ocean mixing layer, cooling of sea surface temperature and upwelling of subsurface water caused by strong typhoon winds. The turbulent mixing scheme is used to handle the enhanced vertical mixing of the ocean induced by the typhoon, and the cold water vortex characteristics in the typhoon wake are simulated by the ocean vortex parameterization scheme.

[0054] S304. Dynamic Update of Atmospheric Boundary Layer: The atmospheric model receives information on sea surface temperature, sea surface roughness, and air-sea flux from the ocean model, and updates the thermal structure and dynamic characteristics of the atmospheric boundary layer in real time. The spatial distribution of sea surface temperature affects the temperature profile and stability stratification of the atmospheric boundary layer, changes in sea surface roughness regulate near-surface wind speed shear and turbulence intensity, and air-sea flux input adjusts the water vapor content and thermal balance of the atmospheric boundary layer, thereby achieving a dynamic response to the surface conditions above and below the typhoon's path.

[0055] It is worth noting that atmospheric-ocean coupled simulation is a key technique to compensate for the lack of ocean physical processes in AI weather models. The introduction of ocean models not only provides dynamically changing sea surface temperature boundary conditions, but more importantly, it simulates the interaction between typhoons and the ocean, including ocean mixing, upwelling, and sea surface temperature cooling effects caused by strong typhoon winds. This two-way coupling mechanism enables the simulation system to capture the physical mechanisms of rapid typhoon intensification and deterioration, particularly the process by which typhoons obtain sufficient energy supply in deep, warm water regions, and the phenomenon of typhoon intensity suppression in shallow or cold water regions.

[0056] S400. Real-time assimilation and correction of multi-source observation data to optimize the simulation process:

[0057] During the integration of the atmospheric-oceanic regional coupled model, multi-source near-surface observation data are dynamically introduced based on a predetermined time step. A data assimilation algorithm is used to perform local high-frequency frequency value assimilation correction on the typhoon eyewall and its adjacent areas, and the local wind field and temperature field status are dynamically updated.

[0058] As a preferred approach, the multi-source near-surface observation data introduced include synthetic aperture radar (SAR) inverted wind field data, wind profiler radar data, microwave satellite inverted data, and / or automatic weather station ground observation data. The ensemble Kalman filter algorithm is used for numerical assimilation, and an adaptive local enhancement coefficient is set in the typhoon eyewall region to improve the ability to capture the high-resolution structure of wind speed and temperature fields in the core area of ​​strong convection. During the assimilation process, an observation data quality control mechanism is established to remove abnormal observations through background field verification and spatial consistency analysis to ensure the physical rationality and numerical stability of the assimilated data.

[0059] In this embodiment of the invention, the real-time assimilation correction of multi-source observation data includes the following sub-steps:

[0060] S401. Quality control and preprocessing of multi-source observation data: Dynamically introduce multi-source observation data, perform format unification and time series alignment based on spatiotemporal matching strategy, and perform preprocessing and quality control on all observation data to remove outliers, missing measurement points and low-confidence samples that do not conform to the assimilation window;

[0061] S402. Perform observation data preprocessing and error modeling: For different types of observation data, based on their temporal resolution, spatial distribution characteristics and systematic error characteristics, construct independent observation error covariance estimation models and optimize them in conjunction with the background error covariance matrix to meet the requirements of subsequent numerical assimilation algorithms for error synergy and weight allocation. For wind field observation data, use wind direction inversion correction algorithm and radial filtering technology to perform directional consistency processing to enhance its physical availability in the typhoon eyewall region.

[0062] S403. Constructing a numerical assimilation framework based on regional models: Integrating data assimilation algorithms into the NWP regional model, setting the assimilation time window length and time step, and setting assimilation interpolation strategies based on the available frequency of observation data to achieve multi-time, localized, and dynamic assimilation of different key variables.

[0063] S404. Implement high-frequency local assimilation in key areas: Set up assimilation sensitive areas in the typhoon eyewall area and its adjacent areas, and perform high-frequency local reanalysis and update of the observation data to ensure accurate fitting of the peak wind speed structure and asymmetric characteristics during the rapid intensification process, thereby improving the ability to characterize eyewall wind speed and radial wind distribution.

[0064] S405. Assimilation results are fed back to the integral simulation system in real time: After each assimilation cycle, the updated local wind field and temperature field states are fed back to the atmospheric-oceanic regional coupled model as constraints, replacing the original predicted state variables and restarting the NWP integration, continuously optimizing the structural evolution path and intensity development trend in typhoon simulation.

[0065] It should be noted that the multi-source observation data used in this embodiment is not simply superimposed, but fused through a data assimilation algorithm. The data assimilation algorithm can perform a weighted average based on the error characteristics and spatial distribution of different observation data, thereby obtaining a more accurate analysis field. In this embodiment, an adaptive local enhancement coefficient is set in the typhoon eyewall region to improve the high-resolution structure capture capability of wind speed and temperature fields in the strong convection core region, which is crucial for accurately predicting typhoon intensity.

[0066] S500. Construct an AI-NWP residual closed-loop feedback correction mechanism:

[0067] During the regional coupled model integration process, the error residuals between the simulation results and the typhoon observation data are periodically acquired. The error information is then fed back into the AI ​​weather model as a feedback input by constructing a residual driving network. This serves as a fine-tuning signal to correct the initial field output of the AI ​​weather model for the next prediction cycle, thereby achieving dynamic two-way closed-loop coupling between the AI ​​weather model and the NWP regional model.

[0068] As a preferred approach, the residual-driven network employs a deep neural network architecture. The input layer receives the deviation vector between the NWP simulation results and the observed data. The hidden layer extracts the spatial-temporal feature patterns of the deviation through a multilayer perceptron. The output layer generates a correction signal corresponding to the weight space of the AI ​​weather model. The network training adopts an online learning method with an adaptively adjustable learning rate that is dynamically optimized based on the correction effect. The fine-tuning process only adjusts the weights of the output layer and the last two hidden layers of the AI ​​weather model to maintain the stability of the main model architecture and avoid catastrophic forgetting. Furthermore, the AI-NWP residual feedback correction mechanism introduces a control threshold strategy. When the error residual between the simulation results of the NWP regional model and the observed data exceeds a set threshold, a full-cycle re-initialization process is initiated to simultaneously reconstruct the input of the AI ​​weather model and the initial field of the NWP regional model, thus preventing the expansion of prediction deviations caused by error accumulation.

[0069] In this embodiment of the invention, the AI-NWP residual feedback correction mechanism includes the following sub-steps during implementation:

[0070] S501. Multidimensional forecast error analysis and quantification: Compare the forecast results of the NWP coupled model with the typhoon observation data according to the preset period, calculate the prediction deviations of the typhoon center location, maximum sustained wind speed, minimum central pressure and typhoon movement speed, and use weighted root mean square error and / or mean absolute error to quantify the magnitude and distribution characteristics of the systematic deviations.

[0071] S502. Residual Feature Extraction and Pattern Recognition: Construct a deep learning-based residual analysis network. The input layer receives the deviation vector between the NWP simulation results and the observed data. The spatial distribution pattern of the error is extracted through a convolutional neural network. A nonlinear mapping relationship between the error and environmental factors is established, and the corresponding correction signal for the weight space of the AI ​​weather model is output.

[0072] S503. Online Fine-tuning and Optimization of AI Weather Model Weights: Based on the correction signal output by the residual analysis network, the model weight parameters are adjusted online. The fine-tuning process only incrementally updates the weights of the model output layer and the last two hidden layers to adapt to the NWP prediction error performance in the current simulation period.

[0073] S504. Constructing an AI-NWP bidirectional closed-loop coupling process: During the regional simulation process, after each predetermined integration period (such as 6 hours or 12 hours), the residual extraction and AI correction process is repeatedly executed to realize dynamic cyclical feedback between NWP and AI weather models, enabling AI weather models to have dynamic response and adaptive correction capabilities based on NWP simulation errors.

[0074] It is worth noting that the AI-NWP residual feedback correction mechanism is a key innovation of this invention. This mechanism identifies systematic bias patterns in the AI ​​model through a deep learning network and transforms this bias information into correction signals for the model weights, enabling online learning and continuous optimization of the AI ​​model. The fine-tuning process employs a conservative strategy, adjusting only the weights of the output layer and the last two hidden layers to avoid damaging the core architecture of the AI ​​model. The threshold control mechanism prevents the negative impact of erroneous feedback information on model performance. When an abnormally large forecast bias is detected, the system automatically triggers a re-initialization process to ensure the stability and reliability of the forecast system.

[0075] S600. Output and evaluate the typhoon intensity simulation results:

[0076] After completing the full-time integration, key parameters of typhoon intensity are output, and structural consistency evaluation is performed in conjunction with actual typhoon observation data to complete the quantitative prediction and analysis of typhoon intensity. As a preferred approach, the evaluation indicators for the typhoon intensity simulation results include the maximum 10-meter wind speed, minimum central pressure, eyewall radius, wind field asymmetry index, and path error. A time series analysis is used to generate the temporal evolution curve of typhoon intensity. The structural consistency evaluation is quantitatively assessed by comparing the simulated typhoon eye diameter, eyewall thickness, and spiral rainband distribution with the actual typhoon observation data.

[0077] Example 2: Application Case

[0078] Based on the above embodiment 1, in order to further verify the practical application effect of the technical solution of the present invention, this embodiment takes Typhoon Doksuri in 2023 as a case to demonstrate in detail the complete implementation process and forecast effect verification of the typhoon intensity prediction method based on AI initial field driven NWP regional model and fused with vortex dynamic initialization.

[0079] According to the technical solution of Example 1, the NWP regional model adopts a three-layer nested structure: outer domain D01 (18km resolution, 311×251 grid), middle domain D02 (6km resolution, 271×271 grid), and inner domain D03 (2km resolution, 211×211 grid). The middle and inner domains are configured with typhoon-tracking moving grids, with 50 vertical layers and a model top of 50 hPa. The model employs WSM6 microphysics, Dudhia shortwave and RRTM longwave radiation, modified MM5 boundary layer, and Kain-Fritsch cumulus parameterization schemes.

[0080] AI Initial Field Quality Validation. Following step S100, the initial field and boundary conditions were generated using the AI ​​weather model. Validation was performed by comparing the AI ​​model with ECMWF model and ERA5 reanalysis data. The AI ​​weather model showed high consistency with observations in predicting the column integral water vapor flux field, accurately capturing monsoon transport, subtropical high pressure, and the dual typhoon circulation structure. Quantitative analysis showed that the root mean square error of the geopotential height field in the AI ​​weather model was 2 m lower than that in the ECMWF model. 2 s -2 The wind field error was reduced by 0.5 ms. -1 This verifies the effectiveness of AI weather models as regional model-driven fields. However, as... Figure 2 As shown in Figures a and b, both the AI ​​weather model and the ECMWF model exhibit a significant underestimation of typhoon intensity in the core region, further confirming the necessity of introducing vortex dynamic initialization technology.

[0081] Then, follow step S200 to perform vortex dynamic initialization. For example... Figure 3 As shown in Figures a and b, the maximum wind speed predicted by the AI ​​weather model before initialization was only 23.9 ms. -1 The eyewall structure was blurred; after vortex separation, 6-hour dynamic enhancement, and repositioning, the typhoon's maximum wind speed increased to 49.0 ms. -1 It is close to the SAR observation value of 48.3 ms. -1 The eyewall structure is clear, and the radial wind field distribution is reasonable. The vortex enhancement process strictly follows dynamic equilibrium, ensuring the physical consistency between the enhanced vortex and the background field.

[0082] Then, the atmosphere-ocean two-way coupled integration was initiated according to step S300. The ROMS ocean model was used, with a 2km resolution matching the inner atmospheric domain and a calculation depth of 1000 meters. The simulation results show that the sea surface temperature cooling effect caused by the typhoon reaches 2-3°C, the characteristics of cold water eddies in the ocean wake are consistent with observations, the air-sea interaction process is accurately reproduced, and an important ocean thermal feedback mechanism is provided for typhoon intensity changes.

[0083] Next, multi-source observation data assimilation is performed according to step S400. An ensemble Kalman filter algorithm is used to assimilate SAR wind field, AMSU-A profile, and ground observation data. After assimilation, the typhoon center position error is reduced by 15%, the maximum wind speed error is reduced by 20%, the eyewall wind field structure is closer to the observed values, and the asymmetric characteristics of the typhoon are accurately described.

[0084] The results of the typhoon track and intensity forecast assessment are as follows: Figure 4As shown in Figures a and b, the typhoon track simulation results indicate that the AI-driven NWP model (AI_NWP) shows a high degree of agreement with the optimal track data in the first 24 hours. Compared to the pure AI model, the track deviation in the later stages is reduced, and the overall track forecast accuracy is better than the traditional ECMWF-driven NWP model (EC_NWP). In terms of typhoon intensity prediction, the AI-driven WRF model achieves significant improvements compared to the pure AI model. During the 72-hour forecast period for Typhoon Doksuri, the maximum root mean square error of the maximum wind speed of the pure AI model reached as high as 29.3 ms. -1 The system systematically underestimated the typhoon's intensity; the root mean square error of the AI-driven NWP model was reduced to 5.1 ms. -1 The forecast accuracy has been improved by 87%. The path forecast 24 hours in advance is highly consistent with the optimal path, and the overall accuracy is better than the traditional EC-driven NWP mode.

[0085] Furthermore, independent verification using high-resolution SAR data showed that at 10:00 AM on July 24th, the maximum wind speed of the typhoon observed by SAR was 48.3 m / s. -1 AI-driven NWP simulation 49.0 ms -1 The error was 1.5%; at 10:00 on July 25, the SAR observation was 57.2 ms. -1 Simulated time: 59.1ms -1 The error is 3.3%. Compared with the pure AI model's underestimation of more than 50%, this method significantly improves both typhoon intensity and eyewall structure forecasts.

[0086] A residual feedback mechanism was constructed according to step S500. By comparing the deviations between the NWP simulation results and observed data, a deep neural network was used to extract error features and fine-tune the AI ​​model weights online. A feedback loop was executed every 12 hours to allow the AI ​​model to gradually adapt to regional characteristics, further improving forecast accuracy and system stability. The results show that this method significantly outperforms traditional methods in terms of typhoon intensity forecast accuracy, structure reconstruction capability, and rapid change prediction, especially in the forecasting capability during the rapid intensification phase of typhoons. It provides advanced technical means for typhoon disaster prevention and mitigation and operational marine meteorology, possessing significant scientific value and broad application prospects.

[0087] The objectives of this invention have been fully and effectively achieved through the above embodiments. Those skilled in the art will understand that this invention includes, but is not limited to, the contents described in the accompanying drawings and the specific embodiments described above. Although the invention has been described with reference to what is currently considered the most practical and preferred embodiments, it should be understood that the invention is not limited to the disclosed embodiments, and any modifications that do not depart from the functional and structural principles of the invention will be included within the scope of the claims.

Claims

1. A typhoon intensity prediction method based on AI initial field-driven NWP regional models and fused with vortex dynamic initialization, characterized in that, It should include at least the following steps: S100. An AI weather model is built and trained using a deep learning architecture and based on historical reanalysis data. The model includes multiple deep learning sub-networks to support different forecast lead times. It uses an Earth spherical sensing structure for input encoding and outputs multi-leadership continuous forecast results covering upper-air meteorological elements. S200. Construct a multi-nested NWP regional model computational domain, set the intermediate domain and inner domain as adaptive tracking moving grids based on the typhoon vortex center position, and introduce vortex dynamic initialization in the preprocessing stage before reporting to construct a dynamic initial field with intensity fitting ability and dynamic consistency. S300. Based on the initial dynamic field after vortex initialization, the NWP regional atmospheric model and ocean model are simultaneously coupled for integral simulation. Within each integral time step, the ocean model updates key underlying surface variables, including sea surface temperature, latent heat flux, and sensible heat flux, in real time as feedback input to the atmospheric model to dynamically correct the sea surface boundary conditions along the typhoon's path. S400. During the integration of the regional coupled model, multi-source near-surface observation data are dynamically introduced based on a predetermined time step. A data assimilation algorithm is used to perform local high-frequency frequency value assimilation correction on the typhoon eyewall and its adjacent areas, and the local wind field and temperature field status are dynamically updated. S500. During the integration process of the regional coupled model, the error residuals between the simulation results and typhoon observation data are periodically acquired. A residual-driven network is constructed to input this error information as feedback into the AI ​​weather model, serving as a fine-tuning signal to online correct the initial field output of the AI ​​weather model for the next forecast period. The residual-driven network adopts a deep neural network architecture. The input layer receives the deviation vector between the NWP simulation results and the observed data. The hidden layer extracts the spatial-temporal feature patterns of the deviation through a multilayer perceptron. The output layer generates a correction signal corresponding to the weight space of the AI ​​weather model. The network training adopts an online learning method with an adaptively adjustable learning rate that is dynamically optimized based on the correction effect. The fine-tuning process only adjusts the weights of the output layer and the last two hidden layers of the AI ​​weather model. Furthermore, the AI-NWP residual feedback correction mechanism introduces a control threshold strategy. When the error residual between the NWP regional model simulation results and the observed data exceeds a set threshold, a full-cycle re-initialization process is initiated, simultaneously reconstructing the input of the AI ​​weather model and the initial field of the NWP regional model. S600. After completing the full-time integration, output the key parameters of typhoon intensity, and combine them with the actual typhoon observation data to conduct a structural consistency evaluation, thereby completing the quantitative prediction and analysis of typhoon intensity.

2. The typhoon intensity prediction method based on AI initial field-driven NWP regional model and fused with vortex dynamic initialization as described in claim 1, characterized in that, In step S100, the AI ​​weather model adopts an encoding-decoding structure. The input end uses an Earth spherical sensing encoding structure and explicitly enhances the latitude and longitude information in the input data through a location embedding mechanism. The AI ​​weather model enhances its cross-temporal dependency modeling capability through a deep residual connection structure. The output end generates multiple continuous upper-air meteorological variable fields with prediction time. The input variables include at least atmospheric temperature, air pressure, specific humidity, latitude and longitude wind speed, and geopotential height. Missing surface variables are supplemented and reconstructed through external reanalysis data or ocean models.

3. The typhoon intensity prediction method based on AI initial field-driven NWP regional model and fused with vortex dynamic initialization as described in claim 1, characterized in that, In step S200, the NWP regional model adopts a three-layer nested structure, including an outer domain, an intermediate domain, and an inner domain. The grid spacing of the outer domain is 15-20 kilometers, the grid spacing of the intermediate domain is 5-8 kilometers, and the grid spacing of the inner domain is 500 meters to 2 kilometers. The intermediate and inner domains are set as typhoon vortex adaptive tracking moving grids, which automatically adjust the coordinates of the computational domain center every 1-10 minutes according to the typhoon center position to ensure that the high-resolution grid domain always covers the typhoon main vortex region. A two-way feedback mechanism is adopted between the three nested layers to ensure that the structural information of the fine grid domain is transmitted upward to the coarse resolution region.

4. The typhoon intensity prediction method based on AI initial field-driven NWP regional model and fused with vortex dynamic initialization as described in claim 1 or 3, characterized in that, In step S200, the vortex dynamic initialization process is introduced in the preprocessing stage before the report is initiated, which includes at least the following sub-steps: S201. Extracting the axisymmetric vortex structure of typhoons: Based on the initial field generated by the AI ​​weather model, key structural variables of the main vortex region of the typhoon are extracted, the vortex circulation structure with axisymmetric characteristics is extracted, asymmetric disturbance components are eliminated, and an idealized initial structural framework of the typhoon is constructed. S202. Perform spectral embedding enhancement simulation: embed the vortex structure into the initial background field of the NWP regional model, use spectral embedding to constrain the large-scale background field, and release the energy of small and medium-scale disturbances. Perform a 6-hour short-time pre-integration simulation to enhance the typhoon core intensity, eyewall structure and symmetry. S203. Implement spatial repositioning of vortex structure: After the short-time integration is completed, based on the deviation between the actual observation of the typhoon center and the typhoon center position in the simulated field, the vortex extraction and displacement function interpolation algorithm is used to spatially reposition and re-embed the simulated vortex structure. S204. Constructing the final dynamic initial field: The vortex simulation results of structural enhancement and position correction are weighted and fused with the background field of the regional model to form a dynamic initial field that takes into account both large-scale accuracy and local vortex intensity, which serves as the initial input for the regional model integral simulation.

5. The typhoon intensity prediction method based on AI initial field-driven NWP regional model and fused with vortex dynamic initialization as described in claim 1, characterized in that, In step S300, the ocean model is selected as ROMS or HYCOM, the ocean grid resolution is set to match the inner domain of the atmospheric model, the ocean model calculation depth is extended to 1000 meters to fully cover the upper mixing layer and thermocline structure, and the ocean state parameters along the typhoon path are refreshed at high frequency. During the integration process, bidirectional coupling is performed with the atmospheric model through the coupling interface. The coupling variables include at least sea surface temperature, latent heat flux, sensible heat flux and 10-meter wind stress, and are dynamically fed back to the NWP atmospheric model to update its boundary layer physical parameters.

6. The typhoon intensity prediction method based on AI initial field-driven NWP regional model and fused with vortex dynamic initialization as described in claim 1 or 5, characterized in that, In step S300, the AI-driven atmospheric-oceanic two-way coupled model integral simulation includes the following sub-steps: S301. Coupled System Synchronous Initialization: Based on the initial atmospheric dynamic field after vortex initialization and the initial temperature and salinity field of the ocean model, the initial state matching of the atmosphere-ocean interface is performed through the coupler to establish consistency constraints for sea surface temperature, sea surface height and 10-meter wind field. S302. Real-time calculation of air-sea interface fluxes: Within each integration time step, the ocean model calculates the momentum flux, sensible heat flux, latent heat flux, and net radiation flux of the air-sea interface in real time, and dynamically determines the flux exchange coefficient based on the difference between sea surface temperature and atmospheric temperature, 10-meter wind speed, and relative humidity. S303. Calculation of ocean dynamic response: The ocean model calculates the Ekman transport, vertical mixing and upwelling processes of the ocean surface in real time based on the wind stress provided by the atmospheric model. The turbulent mixing scheme is used to handle the enhanced vertical mixing of the ocean induced by the typhoon, and the cold water vortex characteristics in the typhoon wake are simulated by the ocean vortex parameterization scheme. S304. Dynamic Update of Atmospheric Boundary Layer: The atmospheric model receives information on sea surface temperature, sea surface roughness, and air-sea flux from the ocean model, and updates the thermal structure and dynamic characteristics of the atmospheric boundary layer in real time, enabling dynamic response to the surface conditions above and below the typhoon's path.

7. The typhoon intensity prediction method based on AI initial field-driven NWP regional model and fused with vortex dynamic initialization as described in claim 1, characterized in that, In step S400, the multi-source near-surface observation data introduced are SAR inverted wind field data, wind profiler radar data, microwave satellite inverted data, and / or automatic weather station ground observation data. The EnKF algorithm is used for numerical assimilation, and an adaptive local enhancement coefficient is set in the typhoon eyewall region to improve the ability to capture the high-resolution structure of wind speed and temperature fields in the core area of ​​strong convection. During the assimilation process, an observation data quality control mechanism is established, and abnormal observation values ​​are eliminated through background field verification and spatial consistency analysis.

8. The typhoon intensity prediction method based on AI initial field-driven NWP regional model and fused with vortex dynamic initialization as described in claim 1 or 7, characterized in that, In step S400, the real-time assimilation and correction of multi-source observation data to optimize the simulation includes the following sub-steps during implementation: S401. Quality control and preprocessing of multi-source observation data: Multi-source observation data are dynamically introduced during the integration process. Based on the spatiotemporal matching strategy, the format is unified and the time series is aligned. Preprocessing and quality control are performed on all observation data to remove outliers, missing points, and low-confidence samples that do not conform to the assimilation window. S402. Perform observation data preprocessing and error modeling: For different types of observation data, based on their temporal resolution, spatial distribution characteristics and systematic error characteristics, construct independent observation error covariance estimation models and optimize them in combination with the background error covariance matrix. Use wind direction inversion correction algorithm and radial filtering technology to process the wind field observation data for directional consistency. S403. Constructing a numerical assimilation framework based on regional models: Integrating data assimilation algorithms into the NWP regional model, setting the assimilation time window length and time step, and setting assimilation interpolation strategies based on the available frequency of observation data to achieve multi-time, localized, and dynamic assimilation of different key variables. S404. Implement high-frequency local assimilation in key areas: Set up assimilation sensitive areas in the typhoon eyewall area and its adjacent areas, and perform high-frequency local reanalysis and update of the observation data to ensure accurate fitting of the peak wind speed structure and asymmetric characteristics during the rapid intensification process; S405. Assimilation results are fed back to the coupled model in real time: After each assimilation cycle, the updated local wind field and temperature field states are fed back to the atmospheric-oceanic regional coupled model as constraints, replacing the original predicted state variables and restarting the NWP integration.

9. The typhoon intensity prediction method based on AI initial field-driven NWP regional model and fused with vortex dynamic initialization as described in claim 1, characterized in that, In step S500, the AI-NWP residual feedback correction mechanism includes the following sub-steps during implementation: S501. Multidimensional forecast error analysis and quantification: Compare the forecast results of the NWP coupled model with the typhoon observation data according to the preset period, calculate the prediction deviations of the typhoon center location, maximum sustained wind speed, minimum central pressure and typhoon movement speed, and use weighted root mean square error and / or mean absolute error to quantify the magnitude and distribution characteristics of the systematic deviations. S502. Residual Feature Extraction and Pattern Recognition: Construct a deep learning-based residual analysis network. The input layer receives the deviation vector between the NWP simulation results and the observed data. The spatial distribution pattern of the error is extracted through a convolutional neural network. A nonlinear mapping relationship between the error and environmental factors is established, and the corresponding correction signal for the weight space of the AI ​​weather model is output. S503. Online Fine-tuning and Optimization of AI Weather Model Weights: Based on the correction signal output by the residual analysis network, the weight parameters of the AI ​​weather model are adjusted online. The fine-tuning process only incrementally updates the weights of the model output layer and the last two hidden layers to adapt to the NWP prediction error performance in the current simulation period. S504. Constructing an AI-NWP bidirectional closed-loop coupling process: During the regional simulation process, after each predetermined integration period, the residual extraction and AI correction process is repeatedly executed to achieve dynamic cyclic feedback between the NWP and the AI ​​weather model, enabling the AI ​​weather model to have the ability to dynamically respond and adaptively correct with the NWP simulation error.

10. The typhoon intensity prediction method based on AI initial field-driven NWP regional model and fused with vortex dynamic initialization as described in claim 1, characterized in that, In step S600, the evaluation indicators of the typhoon intensity simulation results include the maximum 10-meter wind speed, minimum central pressure, eyewall radius, wind field asymmetry index and path error. The time evolution curve of typhoon intensity is generated through time series analysis. The structural consistency evaluation is quantitatively assessed by comparing the degree of matching between the simulated typhoon eye diameter, eyewall thickness and spiral rainband distribution and the actual typhoon observation data.

11. A computer program product comprising computer instructions, characterized in that, The computer instructions are used to execute the typhoon intensity prediction method according to any one of claims 1 to 10, which is based on AI initial field-driven NWP regional model and fused with vortex dynamic initialization.

12. A computer-readable storage medium storing a computer program thereon, characterized in that, When the computer program is executed by the processor, it implements the typhoon intensity prediction method according to any one of claims 1 to 10, which is based on AI initial field-driven NWP regional model and fused with vortex dynamic initialization.

Images