Wind field inversion method and system based on multi-source fusion and physical collaborative modeling

Abstract

The invention provides a wind field inversion method and system based on multi-source fusion and physical collaborative modeling, and belongs to the technical field of atmospheric science and remote sensing, and the method comprises the steps: obtaining and carrying out the time-space alignment of a GNSS radio occultation refractive index profile, a GNSS reflection measurement sea surface wind vector, a foundation GNSS zenith total delay, a radar wind profile and mobile platform observation data; constructing a standardized observation vector; the observation vector is input into a physical information neural network, a loss function comprises a data fitting item and a physical constraint item composed of a heat-to-wind relation, a mass continuity equation and a vortex equation, and observation data and an atmospheric physical rule are fitted at the same time; dynamically adjusting the fusion weight of each observation source based on the weather type identification result; and finally, outputting a three-dimensional wind field with uncertainty quantization and derivative products. According to the invention, high-precision, high-resolution and physically credible three-dimensional wind field inversion is realized, and the inversion robustness and application value under complex weather conditions are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of atmospheric remote sensing and meteorological information inversion technology, and in particular to a wind field inversion method and system based on multi-source fusion and physical collaborative modeling. Background Technology

[0002] Three-dimensional wind fields are core dynamic variables for characterizing atmospheric motion, understanding the evolution of weather systems, and conducting numerical weather prediction. Currently, the main technologies for acquiring three-dimensional wind fields include radiosonde, wind profiler radar, satellite remote sensing, and related technologies of the Global Navigation Satellite System (GNSS). However, each technology has inherent limitations: radiosonde observations have low spatiotemporal resolution; wind profiler radar has limited coverage; traditional GNSS radio occultation (RO) technology relies on the geostrophic equilibrium assumption to retrieve wind fields, resulting in significant errors in non-geostrophic systems such as typhoons and strong convection, and its horizontal resolution is coarse (approximately 200-300 km); GNSS reflectometry (GNSS-R) technology can only acquire sea surface wind speeds and lacks vertical structure information.

[0003] In recent years, research on directly retrieving atmospheric parameters using deep learning has gradually increased. However, purely data-driven models are prone to producing "illusion" solutions that violate physical laws (such as mass non-conservation and vorticity anomalies), exhibit poor generalization ability in sparse observation regions, and are unable to meet the stringent requirements of meteorological operations for "reliable forecasts." Furthermore, existing multi-source fusion methods based on data assimilation typically have a fixed observation error covariance matrix, making it impossible to adaptively adjust the reliability of different observation sources according to specific weather conditions (such as typhoons versus clear skies), resulting in decreased fusion performance under extreme weather conditions.

[0004] Therefore, existing technologies face three major bottlenecks: 1) Single observation methods cannot simultaneously consider the three-dimensional structure of wind fields, high spatiotemporal resolution, and physical consistency; 2) Multi-source fusion lacks the ability to dynamically adapt to uncertainties; and 3) Artificial intelligence models are disconnected from atmospheric physical laws. There is an urgent need for a novel three-dimensional wind field inversion technology that can coordinate multi-source heterogeneous observations, deeply integrate physical knowledge and artificial intelligence, and intelligently adapt to different weather scenarios. Summary of the Invention

[0005] This invention aims to solve at least one of the technical problems existing in the prior art, and proposes a wind field inversion method and system based on multi-source fusion and physical co-modeling, which solves the problems of the limitations of single observation methods and the physical inconsistency of pure data-driven models.

[0006] In a first aspect, embodiments of the present invention provide a wind field inversion method based on multi-source fusion and physical co-modeling, including: S1. Acquire multi-source heterogeneous observation data; S2. Interpolation and mapping to a unified spatiotemporal grid to construct standardized observation vectors; S3. Input the standardized observation vector into the physical information neural network. The physical information neural network outputs a three-dimensional wind field. The loss function includes a data fitting term and an atmospheric physical constraint term. The inverted three-dimensional wind field is obtained by minimizing the loss function. The atmospheric physical constraint term includes at least constraints based on the thermal wind relationship, constraints based on the mass continuity equation, and constraints based on the vorticity equation. S4. Based on the weather system type or observation data quality, dynamically adjust the weights of the multi-source heterogeneous observation data in the fusion and inversion process; S5. Output three-dimensional wind field grid data with mean and standard deviation.

[0007] The technical effect of the wind field inversion method based on multi-source fusion and physical co-modeling disclosed in this invention is to achieve high-resolution, physically consistent, and extreme weather robust wind field inversion, which significantly improves the support capabilities for numerical forecasting, aviation safety, and wind power prediction.

[0008] Furthermore, the multi-source heterogeneous observation data includes: radio occultation refractive index profiles of space-based global navigation satellite systems, sea surface wind vectors measured by reflection from spaceborne or shore-based global navigation satellite systems, total zenith delay of ground-based global navigation satellite systems, wind profile data from weather radar or lidar, and dynamic observation data collected by mobile observation platforms.

[0009] Furthermore, in step S2, the construction of the standardized observation vector specifically includes: interpolating the radio occultation refractive index profile of the Global Navigation Satellite System to the vertical layer, mapping the sea surface wind vector measured by the Global Navigation Satellite System reflection to the bottom grid, converting the total zenith delay of the ground-based Global Navigation Satellite System into prior information on boundary layer humidity, interpolating the radar wind profile data to the grid, and mapping the mobile platform data according to the actual coordinates; performing gross error removal, consistency verification, and signal-to-noise ratio filtering on the mapped data to form a four-dimensional observation tensor containing valid observation values, data source identifiers, and preset uncertainties.

[0010] Furthermore, in step S3, the physical information neural network adopts an encoder-decoder structure, and the input channels include: the refractive index profile of the radio occultation of the Global Navigation Satellite System, the zenith total delay time series of the ground-based Global Navigation Satellite System, the radar radial velocity, the wind speed and direction measured by the Global Navigation Satellite System reflection, and the background wind field data of the reanalysis.

[0011] Furthermore, in step S3, the constraint term based on the thermal wind relationship is implemented through the following loss function components: ; Where u and v are horizontal wind components, T is the temperature profile obtained by inversion of the refractive index profile of radio occultation of the Global Navigation Satellite System, f is the Coriolis parameter, g is the gravitational acceleration, and x, y, and z are the coordinates in the east-west, north-south, and vertical directions, respectively.

[0012] Furthermore, step S4 specifically includes: using a convolutional neural network to classify the input observation field and identify the weather type as a typhoon, front, convective cell, or clear sky; and adjusting the channel weights corresponding to each observation data source through an attention module based on the classification results.

[0013] Furthermore, when the weather type is identified as a typhoon, the fusion weight of GNS reflectance measurement data and radar data is increased, while the fusion weight of GNS radio occultation data is decreased.

[0014] Furthermore, in step S1, the mobile observation platform is an unmanned aerial vehicle (UAV) equipped with a global navigation satellite system reflectance measurement receiver, used for vertical profile observation at a height of 0-2km in the boundary layer.

[0015] Furthermore, in step S5, the output products also include the vertical wind shear index, turbulence intensity estimate, and wind energy density map calculated based on the three-dimensional wind field grid product.

[0016] Secondly, embodiments of the present invention provide a wind field inversion system based on multi-source fusion and physical co-modeling, used to implement the method, including: The multi-source data access module is used to receive and preprocess multi-source heterogeneous observation data; The physics-AI fusion engine is equipped with the aforementioned physical information neural network, which is used to perform spatiotemporal alignment, physical constraint inversion, and dynamic weight fusion. The product generation and distribution module is used to generate and output three-dimensional wind field products and derivative products with uncertainty quantification.

[0017] The technical effects of the wind field inversion system based on multi-source fusion and physical collaborative modeling disclosed in this invention are as follows: This system constitutes a complete "observation-processing-application" operational chain, realizing full-process automation of "observation-modeling-product-application"; the multi-terminal distribution mode ensures the rapid deployment and application of the product, meteorological departments can optimize typhoon path forecasts based on the product, aviation departments can avoid turbulent areas in advance, and wind power companies can accurately predict power generation. Attached Figure Description

[0018] Figure 1 A flowchart illustrating a wind field inversion method based on multi-source fusion and physical co-modeling provided in an embodiment of the present invention; Figure 2A schematic diagram of the overall architecture of the three-dimensional wind field inversion system provided in an embodiment of the present invention. Detailed Implementation

[0019] To enable those skilled in the art to better understand the technical solutions of the present invention, exemplary embodiments of the present invention are described below in conjunction with the accompanying drawings, including various details of the embodiments of the present invention to aid understanding. These should be considered merely exemplary. Therefore, those skilled in the art should recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the present invention. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0020] A wind field inversion method based on multi-source fusion and physical co-modeling, referencing Figure 1 and Figure 2 As shown, its core lies in constructing a closed-loop intelligent inversion framework of "observation-physical laws-AI model". This method integrates multi-source observations from GNSS-RO, GNSS-R, ground-based GNSS, radar, and mobile platforms, and designs a Physical Information Neural Network (PINN) constrained by atmospheric physical equations as the inversion engine. Simultaneously, it introduces a dynamic weight fusion mechanism based on weather recognition, ultimately outputting a high-precision three-dimensional wind field product with uncertainty quantification. Specifically, it includes the following steps: S1. Acquire multi-source heterogeneous observation data: including radio occultation refractive index profiles of space-based global navigation satellite systems, sea surface wind vectors measured by reflection from spaceborne or shore-based global navigation satellite systems, total zenith delay of ground-based global navigation satellite systems, wind profile data from weather radar or lidar, and dynamic observation data collected by mobile observation platforms.

[0021] Space-based GNSS-RO data: Obtained from six low-Earth orbit satellites of the COSMIC-2 constellation, capturing refractive index profiles below 30 km altitude in the typhoon-affected area. Vertical sampling interval is 50 m, and horizontal coverage density is approximately 1 satellite per 100 km. Refractive index (N) profiles obtained from constellations such as COSMIC-2 are used to retrieve temperature, air pressure, and humidity.

[0022] Spaceborne / shore-based GNSS-R data: Sea surface wind vectors (wind speed and direction) provided by 8 NASA CYGNSS satellites, with an observation accuracy of ±1.5 m / s, covering the outer sea area of ​​the typhoon. Sea surface wind speed and direction acquired from CYGNSS satellites or coastal-based receivers.

[0023] Ground-based GNSS data: Twenty GNSS reference stations within the region were selected, and zenith total delay (ZTD) data with a sampling frequency of 1 Hz were output. After preprocessing, this data was converted into boundary layer humidity prior information. The ZTD time series was obtained from a dense GNSS receiver network.

[0024] Radar data: Radial wind profiles provided by two S-band Doppler weather radars, with a detection radius of 250 km, a vertical resolution of 100 m, and a time update interval of 6 minutes. Radial velocity (Vr) obtained from the S-band weather radar, and vertical wind profiles obtained from wind profiler radar / lidar.

[0025] Mobile platform data: Unmanned aerial vehicles (UAVs) equipped with miniature GNSS-R receivers and meteorological sensors were deployed to conduct vertical profile observations at altitudes of 0-2 km in nearshore areas potentially affected by the typhoon, transmitting encrypted wind and GNSS-R data back in real time. Specifically, six UAVs equipped with small GNSS-R receivers were deployed to conduct vertical profile observations at altitudes of 0–2 km above the typhoon boundary layer. Each UAV completed three profile surveys along a pre-set route, and the observation data was transmitted back to the system in real time via a 5G network. Simultaneously, ten ocean buoys were activated to provide measured wind field data from the sea surface as a verification reference.

[0026] By aggregating multi-source observations with varying spatiotemporal resolutions, accuracies, and sensitivities, an unprecedented data foundation has been provided for constructing high-resolution 3D wind fields. In particular, the addition of mobile platforms effectively fills the gaps in traditional observations regarding the boundary layer and specific regions. Specifically, the UAV mobile platform enabled intensive observations of key areas of the typhoon boundary layer, compensating for the sparsity of traditional observations in low-level, high-impact regions. The multi-source data covers all dimensions from high altitude to sea surface, and from fixed to dynamic observations, providing rich observational constraints for 3D wind field inversion. Compared to single RO or radar observations, data coverage completeness is improved by 60%, laying the foundation for subsequent high-precision inversion.

[0027] S2. Spatiotemporal alignment and quality control to construct standardized observation vectors: The multi-source heterogeneous observation data obtained in step S1 are interpolated and mapped to a unified spatiotemporal grid, and quality control is performed to construct standardized observation vectors.

[0028] This step aims to unify the aforementioned heterogeneous data under the same spatiotemporal framework.

[0029] 1. Define a unified grid: Set a 3D grid for the target area, with a horizontal resolution of 3 km and a vertical grid of 40 layers from the sea surface to a height of 15 km. The time window is a 10-minute slide.

[0030] 2. Data Mapping and Interpolation: The GNSS-RO refractive index profile is assigned to the vertical grid column corresponding to its horizontal position using spline interpolation. The GNSS-R sea surface wind vector (wind speed, wind direction) is assigned to the nearest sea surface grid point. Using known statistical models, the ZTD of the ground-based GNSS is converted into prior information on the near-surface (e.g., 0-1 km) humidity mixing ratio and assigned to the corresponding grid points. For radar radial velocity data, the horizontal wind component is first obtained by inversion using velocity azimuth display (VAD) or volumetric velocity processing (VVP) techniques, and then interpolated to the three-dimensional grid. Each set of data observed by the UAV is mapped to the corresponding four-dimensional grid cell based on its precise GPS position and timestamp.

[0031] 3. Quality Control and Vector Construction: Gross error removal (e.g., 3σ criterion), cross-source consistency checks (e.g., comparing wind speeds retrieved from GNSS-R and radar inversion in the same area), and signal-to-noise ratio (SNR) filtering are performed on the mapped grid data. Finally, a fixed-dimensional observation vector is constructed for each grid point. This vector contains valid observations from each data source (missing values ​​are filled with masks or ERA5 reanalysis background fields), a data source identifier, and pre-set observation errors based on instrument characteristics (e.g., GNSS-R wind speed error is set to ±1.5 m / s). The observation vectors from all grid points together constitute a four-dimensional (time, longitude, latitude, altitude) observation tensor.

[0032] This step solves the "language unification" problem of multi-source data fusion, transforming the chaotic and heterogeneous raw observations into tensors with regular structure, controllable quality, and direct processing by AI models, which is a prerequisite for subsequent high-precision fusion and inversion.

[0033] S3. Physics-AI Fusion Inversion: The standardized observation vector obtained in step S2 is input into a physical information neural network. The physical information neural network outputs a three-dimensional wind field, and its loss function includes a data fitting term and an atmospheric physics constraint term. The inverted three-dimensional wind field is obtained by minimizing the loss function. The atmospheric physics constraint term includes at least constraints based on the thermal wind relationship, constraints based on the mass continuity equation, and constraints based on the vorticity equation.

[0034] This step is the core of the method, which inputs the observation tensor obtained in step S2 into a carefully designed physical information neural network.

[0035] 1. Network Structure: An improved U-Net encoder-decoder structure is adopted as the backbone network. The encoder input channels include: GNSS-RO refractive index field, ground-based GNSS humidity prior field, radar radial velocity field / wind profile field, GNSS-R sea surface wind field, and background wind and temperature fields provided by ERA5 reanalysis data. The final output of the decoder is a three-dimensional wind vector (U, V, W) for each grid point.

[0036] 2. Loss Function Design: Total Loss Function The key innovation point is defined as: ; Among them, the data fitting term This involves calculating the mean square error between the wind field predicted by the network and the wind field indirectly derived from observation data or provided by the background field. For example, comparing the predicted U and V with the wind retrieved by radar VAD, and comparing the predicted sea surface wind with GNSS-R observations.

[0037] Physical constraint terms: Thermal wind constraint terms Where u and v are the horizontal wind components, T is the temperature profile obtained by inversion from the GNSS-RO radio occultation refractive index profile, f is the Coriolis parameter, g is the gravitational acceleration, and x, y, and z are the east-west, north-south, and vertical coordinates, respectively. The temperature profile (temperature field) T is obtained by physical inversion from the input GNSS-RO refractive index profile. This forced network prediction of the wind field's vertical shear and horizontal temperature gradient satisfies the thermo-wind relationship, which is an important equilibrium constraint for large-scale wind fields in mid-to-high latitudes.

[0038] Quality continuity constraint The forced wind field divergence is close to zero, which satisfies the continuity equation for incompressible fluids and suppresses non-physical mass sources / sinks.

[0039] vorticity constraint term The residual term is designed based on the vorticity equation to coordinate the evolution relationship between vorticity and divergence, and further suppress non-physical vortex structures.

[0040] By embedding physical laws as "soft constraints" into the network training using PINN, the model is no longer a black box. It learns statistical features from the data and follows physical laws, thus ensuring that the inverted wind field is physically plausible even in areas of sparse observation. This fundamentally avoids the "illusion" problem of pure AI models and improves the credibility of the inversion results.

[0041] S4. Dynamic weighted fusion: Based on the weather system type or observation data quality, dynamically adjust the weights of the multi-source heterogeneous observation data mentioned in step S3 during the fusion and inversion process.

[0042] This step is implemented collaboratively with step S3 in the model. A dynamic weight fusion module is designed within the network.

[0043] 1. Weather type identification: A lightweight ResNet-18 classification network analyzes the input multi-channel observation tensor (or its feature map) to identify the weather type of the current scene in real time, such as "typhoon", "front", "convective cell" or "clear sky".

[0044] 2. Dynamic Weight Allocation: The classification result serves as a control signal, inputting into an attention module. This module assigns different attention weights to input channels from different data sources. For example, when identified as "typhoon," the system determines that GNSS-R (directly sensing sea surface winds) and radar (capturing eyewall convection) data are more reliable, automatically increasing the attention weight of their corresponding channels to 0.6 (before sum normalization), while reducing the weight of GNSS-RO (large inversion error in strong convection areas) to 0.2. When identified as "clear sky," the weights of GNSS-RO (high accuracy under stable strata) and ground-based GNSS (stable humidity information) are increased, while the radar weight is decreased.

[0045] This mechanism overcomes the limitations of fixed weights in traditional fusion methods, enabling the system to possess "contextual awareness" capabilities. It can intelligently rely on more reliable observation sources based on actual weather conditions, thereby significantly enhancing the robustness and accuracy of inversion under extreme and complex weather systems such as typhoons.

[0046] S5. Product Generation and Uncertainty Quantification: Outputs a three-dimensional wind field grid product with mean and standard deviation.

[0047] 1. Ensemble Inversion and Uncertainty Quantification: During the inference phase, Monte Carlo Dropout is applied to the trained PINN array, performing multiple forward propagations to obtain a series of slightly different 3D wind field samples, forming an ensemble. The mean of this ensemble is used as the final best estimate of the wind field, and its standard deviation is used as a measure of the wind field's uncertainty.

[0048] 2. Derivative Product Generation: Based on 3D wind field grid data, the system automatically calculates: Vertical wind shear: used for aviation turbulence warning; Turbulence intensity: assesses atmospheric mixing and diffusion conditions; Wind energy density: provides input for power prediction of offshore wind farms.

[0049] These products output in standardized formats (such as NetCDF and GRIB2). The output of uncertainty information meets the risk quantification needs of high-impact weather decision-making. Directly generating application-level products enables a one-stop transformation from raw data to decision support information, greatly improving the efficiency of business applications.

[0050] This invention constructs a novel intelligent inversion paradigm for 3D wind fields by organically combining three major technical features: multi-source data access and fusion, physically constrained AI modeling, and dynamic intelligent weighting. It effectively overcomes the challenges of existing technologies in terms of resolution, physical consistency, and robustness to extreme weather. The high-precision 3D wind field products generated, with credibility assessment, are of great value for improving the quality of numerical weather prediction initialization fields, ensuring aviation safety, and optimizing wind energy utilization.

[0051] Based on the same inventive concept, embodiments of the present invention provide a wind field inversion system based on multi-source fusion and physical co-modeling, used to implement the method described above, including: The multi-source data access module is used to receive and preprocess multi-source heterogeneous observation data; it is responsible for interfacing with various observation networks and satellite data links to complete data decoding, format parsing and preliminary time synchronization. Its functions are implemented as described in step S1 of Embodiment 1.

[0052] The physics-AI fusion engine, deployed with the aforementioned physical information neural network, performs spatiotemporal alignment, physical constraint inversion, and dynamic weight fusion. As the core processing unit, it integrates a complete set of algorithms for data spatiotemporal alignment, quality control (step S2), PINN model inference and dynamic weight fusion (steps S3 and S4), and uncertainty quantification (step S5). Deployed on a GPU server cluster, this engine leverages its parallel computing capabilities to complete a full-process inversion update of the covered area every 5-10 minutes.

[0053] The product generation and distribution module is used to generate and output three-dimensional wind field products and derivative products with uncertainty quantification. It receives the results output by the fusion engine, generates standardized products and derivative products as described in step S5 of Example 1, and pushes them in real time to end users such as air traffic control, wind power dispatch centers, and meteorological stations through meteorological information broadcasting systems, dedicated APIs, or data stream interfaces.

[0054] Example embodiments have been disclosed herein, and while specific terminology has been used, it is for illustrative purposes only and should be construed as such, and is not intended to be limiting. In some instances, it will be apparent to those skilled in the art that features, characteristics, and / or elements described in conjunction with particular embodiments may be used alone, or in combination with features, characteristics, and / or elements described in conjunction with other embodiments, unless otherwise expressly indicated. Therefore, those skilled in the art will understand that various changes in form and detail may be made without departing from the scope of the invention as set forth in the appended claims.

Claims

1. A wind field inversion method based on multi-source fusion and physical co-modeling, characterized in that, include: S1. Acquire multi-source heterogeneous observation data; S2. Interpolation and mapping to a unified spatiotemporal grid to construct standardized observation vectors; S3. Input the standardized observation vector into the physical information neural network. The physical information neural network outputs a three-dimensional wind field. The loss function includes a data fitting term and an atmospheric physical constraint term. The inverted three-dimensional wind field is obtained by minimizing the loss function. The atmospheric physical constraint term includes at least constraints based on the thermal wind relationship, constraints based on the mass continuity equation, and constraints based on the vorticity equation. S4. Based on the weather system type or observation data quality, dynamically adjust the weights of the multi-source heterogeneous observation data in the fusion and inversion process; S5. Output three-dimensional wind field grid data with mean and standard deviation.

2. The method according to claim 1, characterized in that, The multi-source heterogeneous observation data includes: radio occultation refractive index profiles of space-based global navigation satellite systems, sea surface wind vectors measured by reflection from spaceborne or shore-based global navigation satellite systems, total zenith delay of ground-based global navigation satellite systems, wind profile data from weather radar or lidar, and dynamic observation data collected by mobile observation platforms.

3. The method according to claim 1, characterized in that, In step S2, the construction of the standardized observation vector specifically includes: interpolating the radio occultation refractive index profile of the Global Navigation Satellite System to the vertical layer, mapping the sea surface wind vector measured by the Global Navigation Satellite System reflection to the bottom grid, converting the total zenith delay of the ground-based Global Navigation Satellite System into prior information on boundary layer humidity, interpolating the radar wind profile data to the grid, and mapping the mobile platform data according to the actual coordinates; performing gross error removal, consistency verification, and signal-to-noise ratio filtering on the mapped data to form a four-dimensional observation tensor containing valid observation values, data source identifiers, and preset uncertainties.

4. The method according to claim 1, characterized in that, In step S3, the physical information neural network adopts an encoder-decoder structure, and the input channels include: the refractive index profile of the radio occultation of the global navigation satellite system, the zenith total delay time series of the ground-based global navigation satellite system, the radar radial velocity, the wind speed and direction measured by the global navigation satellite system reflection, and the background wind field data of the reanalysis.

5. The method according to claim 1, characterized in that, In step S3, the constraint term based on the thermal wind relationship is implemented through the following loss function components: ； Where u and v are horizontal wind components, T is the temperature profile obtained by inversion of the refractive index profile of radio occultation of the Global Navigation Satellite System, f is the Coriolis parameter, g is the gravitational acceleration, and x, y, and z are the coordinates in the east-west, north-south, and vertical directions, respectively.

6. The method according to claim 1, characterized in that, Step S4 specifically includes: using a convolutional neural network to classify the input observation field and identify the weather type as a typhoon, front, convective cell, or clear sky; and adjusting the channel weights corresponding to each observation data source through an attention module based on the classification results.

7. The method according to claim 6, characterized in that, When the weather type is identified as a typhoon, the fusion weight of global navigation satellite system reflectance measurement data and radar data is increased, while the fusion weight of global navigation satellite system radio occultation data is decreased.

8. The method according to claim 1, characterized in that, In step S1, the mobile observation platform is a UAV equipped with a global navigation satellite system reflectance measurement receiver, used for vertical profile observation at a height of 0-2km in the boundary layer.

9. The method according to claim 1, characterized in that, In step S5, the output products also include the vertical wind shear index, turbulence intensity estimate, and wind energy density map calculated based on the three-dimensional wind field grid product.

10. A wind field inversion system based on multi-source fusion and physical co-modeling, used to implement the method as described in any one of claims 1 to 9, characterized in that, include: The multi-source data access module is used to receive and preprocess multi-source heterogeneous observation data; The physics-AI fusion engine is equipped with the aforementioned physical information neural network, which is used to perform spatiotemporal alignment, physical constraint inversion, and dynamic weight fusion. The product generation and distribution module is used to generate and output three-dimensional wind field products and derivative products with uncertainty quantification.

Images