Three-dimensional downscaling wind field two-stage super-resolution reconstruction method and system for complex terrain bridge site

Abstract

This invention proposes a two-stage super-resolution reconstruction method and system for three-dimensional downscaled wind fields in bridge sites with complex terrain. First, the invention utilizes a WRF model and explicitly enables large eddy simulation to construct a high-fidelity training set. To build a spatially three-dimensional reconstruction result, the invention extracts dynamic meteorological sequences and static geographical features for different feature height layers. After filtering out invalid regions using a mountain occlusion mask matrix, three parallel cascaded reconstruction models are independently trained to significantly improve the real-time performance of data acquisition. The model training innovatively integrates terrain-forced physical loss and frequency domain energy constraints. Finally, the two-dimensional reconstruction results from each height layer are combined in vertical space to construct a high-precision three-dimensional microscale meteorological field ensemble, providing reliable technical support for three-dimensional wind energy resource assessment and bridge health monitoring in complex terrain.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of artificial intelligence and atmospheric science, and in particular to a two-stage super-resolution reconstruction method and system for three-dimensional downscaled wind fields in complex terrain bridge sites. Specifically, it relates to a three-dimensional microscale meteorological field super-resolution reconstruction method for bridge health monitoring and wind engineering, which combines physical mechanism constraints, large eddy simulation dataset support, mountain shading masking mechanism, and multi-height layer parallel depth generation network. Background Technology

[0002] In the fields of refined weather forecasting and bridge wind engineering, especially in bridge structural health monitoring, obtaining high-resolution (e.g., 100-meter level) three-dimensional near-surface meteorological field data with vertical gradients around the bridge is crucial for accurately assessing the wind resistance and operational safety of long-span bridges. Traditional methods often rely on multi-layered nested dynamic downscaling using mesoscale numerical weather prediction models (such as WRF models), but generating ultra-high-resolution simulation results is extremely computationally intensive, with very high computational time costs, which cannot meet the urgent need of bridge health monitoring systems for real-time acquisition of high-precision wind field data.

[0003] In recent years, deep learning-based super-resolution technology has been introduced into the meteorological field to improve data acquisition efficiency. However, most existing models are limited to two-dimensional planar downscaling at a single altitude, ignoring the strong vertical shear characteristics of wind speed variation with altitude within the atmospheric boundary layer, and failing to form a three-dimensional flow field structure. Furthermore, existing models also suffer from the following shortcomings: 1. Lack of physical consistency, generating flow fields that often violate atmospheric dynamics; 2. In complex terrain areas, they cannot accurately reflect physical phenomena such as mountain obstruction and flow around obstacles; 3. Difficulty in recovering the high-frequency turbulent energy spectrum of the flow field. Therefore, there is an urgent need for an intelligent meteorological downscaling method that balances physical constraints, realistic terrain considerations, and the ability to construct three-dimensional structural features. Summary of the Invention

[0004] The purpose of this invention is to solve the problems in the prior art and to propose a two-stage super-resolution reconstruction method and system for three-dimensional downscaled wind fields in complex terrain bridge sites.

[0005] This invention is achieved through the following technical solution: This invention proposes a two-stage super-resolution reconstruction method for three-dimensional downscaling wind fields in complex terrain bridge site areas, the method comprising: Step 1: Obtain a high-fidelity simulation dataset from a mesoscale meteorological model and extract low-resolution dynamic meteorological sequences and high-resolution static geographic features. Step 2: For multiple preset height layers, interpolate to construct corresponding multi-source heterogeneous input features, and construct a spatial mask matrix representing the mountain shading effect based on the real terrain height. Step 3: Train the corresponding cascaded super-resolution reconstruction model independently for each altitude layer. The model includes a first-stage fully convolutional network for extracting large-scale meteorological background fields and a second-stage generative adversarial network for predicting high-frequency spatial residual fields by combining the background fields. Step 4: Under the constraints of the spatial mask matrix and physical loss function, output the high-resolution two-dimensional meteorological reconstruction field of each height layer, and stitch them together in the vertical spatial dimension to reconstruct a three-dimensional meteorological field assembly of the target area.

[0006] Furthermore, the different height layers include at least 30 meters, 54.5 meters, and 70 meters; the models trained independently for each layer have the same topology but their weights are optimized independently.

[0007] Furthermore, the high-fidelity simulation dataset is obtained by disabling boundary layer and cumulus parameterization schemes and enabling explicit analytical turbulence analysis using large eddy simulation within the target region of a mesoscale meteorological model.

[0008] Furthermore, the multi-source heterogeneous input features include dynamic sequences of three-dimensional wind speed, temperature, and air pressure at current and historical moments, as well as static data on terrain height, slope, aspect, and surface roughness.

[0009] Furthermore, the spatial mask matrix is ​​generated by marking the grids obscured by the physical terrain as missing values; the model's loss calculation and gradient backpropagation are performed only within the effective area of ​​the mask.

[0010] Furthermore, the physical loss function includes a terrain-forcing loss based on no-slip boundary conditions.

[0011] Furthermore, the loss function of the second-stage generative adversarial network is composed of a weighted sum of generative adversarial loss, pixel-level error loss, and frequency domain energy constraint loss based on fast Fourier transform.

[0012] This invention also proposes a two-stage super-resolution reconstruction system for three-dimensional downscaling wind fields in complex terrain bridge sites, the system comprising: Acquisition module: Acquires high-fidelity simulation datasets from mesoscale meteorological models, and extracts low-resolution dynamic meteorological sequences and high-resolution static geographic features; Construction module: For multiple preset height layers, interpolation is used to construct corresponding multi-source heterogeneous input features, and a spatial mask matrix representing the mountain shading effect is constructed based on the real terrain height. Training module: Independently train the corresponding cascaded super-resolution reconstruction model for each altitude layer. The model includes a first-stage fully convolutional network for extracting large-scale meteorological background fields, and a second-stage generative adversarial network for predicting high-frequency spatial residual fields by combining the background fields. Reconstruction Module: Under the constraints of the spatial mask matrix and physical loss function, outputs high-resolution two-dimensional meteorological reconstruction fields at each altitude level, and stitches them together in the vertical spatial dimension to reconstruct a three-dimensional meteorological field assembly of the target area.

[0013] The present invention also proposes an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the two-stage super-resolution reconstruction method for three-dimensional downscaling wind field in complex terrain bridge site areas.

[0014] The present invention also proposes a computer-readable storage medium for storing computer instructions, which, when executed by a processor, implement the steps of the two-stage super-resolution reconstruction method for three-dimensional downscaling wind fields in complex terrain bridge sites.

[0015] The beneficial effects of this invention are: 1. Accurate Reconstruction of Three-Dimensional Flow Field Structure: This breakthrough overcomes the limitation of traditional deep learning meteorological downscaling, which can only generate two-dimensional planes. By independently training parallel networks for different key height layers and spatially combining them, a three-dimensional meteorological field ensemble with vertical wind shear characteristics was successfully constructed, meeting the stringent requirements of bridge health monitoring.

[0016] 2. Deep mapping of real physical laws: Based on WRF large eddy simulation data and combined with terrain-forced physical loss, the three neural networks are forced to follow the fluid dynamics no-slip boundary conditions, which greatly improves the realism of vertical airflow and hill-crossing airflow.

[0017] 3. Topological consistency under complex terrain: The proposed mountain occlusion masking mechanism is dynamically applicable at different height levels, perfectly reproducing the topological structure of the three-dimensional flow field around the complex terrain.

[0018] 4. Breaking through the computing power bottleneck of traditional modes and achieving real-time high-precision acquisition of wind fields: By replacing the traditional WRF numerical simulation physical integration process with a pre-trained multi-height cascaded neural network, the problem of extremely high computation time cost of the WRF mode is completely solved. It can output high-precision three-dimensional wind field data around the bridge in real time within a very short inference time, which greatly improves the data acquisition efficiency and early warning timeliness of the bridge health monitoring system. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0020] Figure 1 This is a flowchart of the two-stage super-resolution reconstruction method for three-dimensional downscaling wind fields in complex terrain bridge sites, as described in this invention.

[0021] Figure 2 Comparison cloud maps of super-resolution results for flow field at a height of 30m.

[0022] Figure 3 Comparison of super-resolution results for flow field at a height of 54.5m.

[0023] Figure 4 Comparison cloud maps of super-resolution results for the flow field at a height of 70m.

[0024] Figure 5 This is a schematic diagram of the computational region for WRF numerical simulation. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] Specifically, in combination Figures 1-5 This invention proposes a two-stage super-resolution reconstruction method for three-dimensional downscaling wind fields in complex terrain bridge site areas, the method comprising: Step 1: Construction and preprocessing of high-fidelity dataset based on large eddy simulation: Obtain high-fidelity simulation dataset from mesoscale meteorological models, and extract low-resolution dynamic meteorological sequences and high-resolution static geographic features; The high-fidelity simulation dataset is obtained by disabling boundary layer and cumulus parameterization schemes and enabling explicit analytical turbulence analysis in large eddy simulation within the target area of ​​the mesoscale meteorological model.

[0027] Based on a 5-year numerical simulation using a mesoscale meteorological model (WRF), dynamic meteorological sequences were obtained for the low-resolution source region (500m resolution) and the target region (100m resolution). Boundary layer and cumulus parameterization schemes were disabled for the high-resolution target region, and a 3D Smagorinsky first-order closure scheme was used for large eddy simulation. 3D variables such as wind speed, temperature, and air pressure were extracted for the current and adjacent historical moments. Step 2: Multi-height layer feature separation and mountain occlusion mask construction: For multiple preset height layers, interpolation is used to construct corresponding multi-source heterogeneous input features, and a spatial mask matrix representing the mountain occlusion effect is constructed based on the actual terrain height. The different height layers include at least 30 meters, 54.5 meters, and 70 meters. The multi-source heterogeneous input features include dynamic sequences of current and historical three-dimensional wind speed, temperature, and air pressure, as well as static data on terrain height, slope, aspect, and surface roughness. The spatial mask matrix is ​​generated by marking grids occluded by physical terrain as missing values; the model's loss calculation and gradient backpropagation are performed only within the effective area of ​​the mask.

[0028] For different height layers such as 30m, 54.5m, and 70m, corresponding dynamic meteorological data are extracted, and specific input features for each height layer are constructed by combining them with static geographic variables such as terrain height and slope. At the same time, based on the actual terrain height, missing values ​​are assigned to grid nodes that are blocked by mountain entities to generate an effective region mask matrix that represents the boundary of the real three-dimensional flow field. Step 3: Parallel Prediction of Large-Scale Background Field and Parallel Adversarial Generation of High-Frequency Residual Field: For each altitude layer, a cascaded super-resolution reconstruction model is independently trained. This model includes a first-stage fully convolutional network for extracting the large-scale meteorological background field, and a second-stage generative adversarial network for predicting the high-frequency spatial residual field based on the background field. The models trained independently at each layer have the same topology but their weights are independently optimized. The loss function of the second-stage generative adversarial network is a weighted sum of generative adversarial loss, pixel-level error loss, and frequency domain energy constraint loss based on Fast Fourier Transform.

[0029] For different altitude layers, independent shallow fully convolutional neural networks are constructed. The input data for each altitude layer is fed into the corresponding network, and the large-scale average meteorological background field for each altitude layer is predicted by combining Sobel gradient loss and masking mechanism. In the model of each altitude layer, the original low-resolution input and the background field output of stage one are concatenated and fed into the dedicated generative adversarial network generator for each altitude layer to extract the spatial high-frequency residual features of each altitude layer.

[0030] Step 4: Joint Reconstruction Based on Physical Constraints and Construction of a 3D Assemblies: Under the constraints of the spatial mask matrix and the physical loss function, high-resolution two-dimensional meteorological reconstruction fields for each altitude layer are output, and these fields are stitched together in the vertical spatial dimension to reconstruct a 3D meteorological field ensemble for the target area. The physical loss function includes terrain forcing loss based on no-slip boundary conditions.

[0031] The background field and residual field at each altitude level are superimposed to obtain a high-precision two-dimensional reconstructed field for each level. Finally, the high-resolution two-dimensional flow fields at 30 meters, 54.5 meters, and 70 meters are stitched together in the vertical spatial dimension to construct a three-dimensional microscale meteorological field ensemble with vertical profile gradient characteristics.

[0032] This invention also proposes a two-stage super-resolution reconstruction system for three-dimensional downscaling wind fields in complex terrain bridge sites, the system comprising: Acquisition module: Acquires high-fidelity simulation datasets from mesoscale meteorological models, and extracts low-resolution dynamic meteorological sequences and high-resolution static geographic features; Construction module: For multiple preset height layers, interpolation is used to construct corresponding multi-source heterogeneous input features, and a spatial mask matrix representing the mountain shading effect is constructed based on the real terrain height. Training module: Independently train the corresponding cascaded super-resolution reconstruction model for each altitude layer. The model includes a first-stage fully convolutional network for extracting large-scale meteorological background fields, and a second-stage generative adversarial network for predicting high-frequency spatial residual fields by combining the background fields. Reconstruction Module: Under the constraints of the spatial mask matrix and physical loss function, outputs high-resolution two-dimensional meteorological reconstruction fields at each altitude level, and stitches them together in the vertical spatial dimension to reconstruct a three-dimensional meteorological field assembly of the target area.

[0033] This invention addresses the urgent need for high-resolution wind field data around bridges in bridge health monitoring and wind engineering fields, proposing a two-stage super-resolution reconstruction method and system for three-dimensional downscaled wind fields in complex terrain bridge sites. To address the high computational cost and time consumption of traditional mesoscale models for downscaling, which makes it difficult to acquire high-precision wind field data in real time, and the inability of single-plane reconstruction to reflect vertical wind shear, this invention first utilizes a WRF model and explicitly enables Large Eddy Simulation (LES) to construct a high-fidelity training set. To construct a spatially three-dimensional reconstruction result, this invention extracts dynamic meteorological sequences and static geographical features for different feature height layers. After filtering out invalid areas using a mountain occlusion mask matrix, three parallel cascaded reconstruction models are independently trained to significantly improve the real-time performance of data acquisition. The model training innovatively integrates terrain-forced physical loss and frequency domain energy constraints. Finally, the two-dimensional reconstruction results of each height layer are combined in vertical space to construct a high-precision three-dimensional microscale meteorological field ensemble, providing reliable technical support for three-dimensional wind energy resource assessment and bridge health monitoring in complex terrain.

[0034] Example This invention proposes a two-stage super-resolution reconstruction method for three-dimensional downscaling wind fields in complex terrain bridge site areas, the method comprising: 1. Construction of a high-fidelity microscale meteorological dataset based on large eddy simulation In this example, the area near the Xihoumen Bridge in Zhoushan City, Zhejiang Province, was selected as the simulation object. A mesoscale WRF model was used for a five-year (2013-2017) continuous numerical simulation. To avoid data leakage caused by bidirectional feedback from nested meshes and to ensure the physical independence of low-resolution input and high-resolution target, a separate simulation strategy was adopted: three-layer and four-layer nested mesh simulations were run independently, with the computational domain view as shown below. Figure 5 As shown in the diagram. In terms of configuration, the conventional planetary boundary layer and cumulus convection parameterization schemes are used in regions d01 to d03; in the independently simulated region d04, the above schemes are explicitly turned off, and the three-dimensional Smagorinsky first-order closed large eddy simulation (LES) is activated to explicitly resolve microscale turbulence.

[0035] The core of the computational principle in WRF numerical simulations lies in the equations of motion for non-hydrostatic compressible atmospheres. First, the momentum equation (expressed in η coordinates) can be written as:

[0036] in V=(u,v,w) For three-dimensional wind speed components, f Here are the Coriolis parameters and the Φ potential function. This represents the momentum friction and subgrid turbulence terms. The continuity equation (mass conservation) is:

[0037] Where μ is the layer thickness in mass coordinates. The thermodynamic equation (evolution of potential temperature θ) and the water conservation equation are respectively:

[0038] in For heat source items, This refers to the microphysical source and sink terms for various water phases (such as cloud water, raindrops, snow particles, etc.). Regarding the vertical coordinates, WRF-ARW uses a terrain-following η-coordinate defined as:

[0039] Make each layer interact with the surface pressure p s and mode top pressure p top This correlation ensures sufficient linear resolution in complex terrain areas. On the horizontal coordinates, either Cartesian or polar projection can be selected based on the study area.

[0040] For time-discrete conditions, the WRF mode uses a third-order Runge-Kutta (RK3) explicit scheme on the nonlinear term:

[0041] inR( ) This represents the nonlinear term on the right-hand side of the equation. Δt For time step.

[0042] The model outputs dynamic meteorological variables at a 10-minute resolution. Subsequently, the acquired d03 and d04 flow field data are interpolated vertically to heights of 30m, 54.5m, and 70m to extract three-dimensional wind speed, temperature, and sea-level pressure sequences, constructing a high-fidelity heterogeneous dataset.

[0043] 2. Multi-height-layer parallel cascaded super-resolution architecture Based on the above approach, deep learning is introduced for further processing. For example... Figure 1 As shown, for three altitude layers (30m, 54.5m, and 70m), three cascaded neural networks with consistent network topologies were established and trained independently. Taking any altitude layer as an example, the model input concatenates dynamic meteorological variables and static geographical features (22 channels in total) for that layer, and invalid nodes inside the mountain are removed using a mask. The first stage predicts the large-scale average flow field using a convolutional network containing only two 3×3 downsampling operations to preserve the basic spatial morphology; the second stage concatenates the input data with the output of the first stage, and a dedicated generator network predicts the high-frequency spatial residuals. Each altitude layer model is trained independently to fit specific physical distribution patterns from the lower-level topographic friction to the upper-level free atmosphere. The super-resolution results are shown below. Figure 2 , Figure 3 and Figure 4 As shown.

[0044] 3. Construction of three-dimensional assemblies and calculation of terrain-forced physical loss During single-layer model training, terrain-forced physical loss is introduced to constrain vertical wind speed: the elevation physical gradient is calculated based on the central difference method.

[0045] Subsequently, the forced velocity was calculated based on the fluid no-slip boundary condition:

[0046] And calculate the loss under the mask matrix M specification:

[0047] During the inference phase, after the three sets of height layer networks complete their calculations in parallel, the high-resolution two-dimensional results output from 30m, 54.5m, and 70m are stacked and assembled in the vertical direction to finally reconstruct a three-dimensional meteorological field assembly of the target area.

[0048] The present invention also proposes an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the two-stage super-resolution reconstruction method for three-dimensional downscaling wind field in complex terrain bridge site areas.

[0049] The present invention also proposes a computer-readable storage medium for storing computer instructions, which, when executed by a processor, implement the steps of the two-stage super-resolution reconstruction method for three-dimensional downscaling wind fields in complex terrain bridge sites.

[0050] The memory in this application embodiment can be volatile memory or non-volatile memory, or it can include both volatile and non-volatile memory. The non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous linked dynamic random access memory (SLDRAM), and direct rambus RAM (DR RAM). It should be noted that the memory used in the methods described in this invention is intended to include, but is not limited to, these and any other suitable types of memory.

[0051] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When the computer instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available media may be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., high-density digital video discs (DVDs)), or semiconductor media (e.g., solid-state disks (SSDs)).

[0052] In implementation, each step of the above method can be completed by integrated logic circuits in the processor's hardware or by instructions in software. The steps of the method disclosed in the embodiments of this application can be directly implemented by a hardware processor, or by a combination of hardware and software modules in the processor. The software modules can reside in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, detailed descriptions are omitted here.

[0053] It should be noted that the processor in the embodiments of this application can be an integrated circuit chip with signal processing capabilities. During implementation, each step of the above method embodiments can be completed by the integrated logic circuitry in the processor's hardware or by instructions in software form. The processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this application can be directly embodied as execution by a hardware decoding processor, or as a combination of hardware and software modules in the decoding processor. The software modules can be located in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. This storage medium is located in memory, and the processor reads the information in the memory and, in conjunction with its hardware, completes the steps of the above methods.

[0054] The above provides a detailed description of the two-stage super-resolution reconstruction method and system for three-dimensional downscaling wind fields in complex terrain bridge sites proposed in this invention. Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A two-stage super-resolution reconstruction method for three-dimensional downscaling wind fields in complex terrain bridge site areas, characterized in that, The method includes: Step 1: Obtain a high-fidelity simulation dataset from a mesoscale meteorological model and extract low-resolution dynamic meteorological sequences and high-resolution static geographic features. Step 2: For multiple preset height layers, interpolate to construct corresponding multi-source heterogeneous input features, and construct a spatial mask matrix representing the mountain shading effect based on the real terrain height. Step 3: Train the corresponding cascaded super-resolution reconstruction model independently for each altitude layer. The model includes a first-stage fully convolutional network for extracting large-scale meteorological background fields and a second-stage generative adversarial network for predicting high-frequency spatial residual fields by combining the background fields. Step 4: Under the constraints of the spatial mask matrix and physical loss function, output the high-resolution two-dimensional meteorological reconstruction field of each height layer, and stitch them together in the vertical spatial dimension to reconstruct a three-dimensional meteorological field assembly of the target area.

2. The method according to claim 1, characterized in that, The different height layers include at least 30 meters, 54.5 meters, and 70 meters; the models trained independently for each layer have the same topology but their weights are optimized independently.

3. The method according to claim 1, characterized in that, The high-fidelity simulation dataset was obtained by disabling boundary layer and cumulus parameterization schemes and enabling explicit analytical turbulence analysis using large eddy simulation within the target region of a mesoscale meteorological model.

4. The method according to claim 1, characterized in that, The multi-source heterogeneous input features include dynamic sequences of three-dimensional wind speed, temperature, and air pressure at current and historical moments, as well as static data on terrain height, slope, aspect, and surface roughness.

5. The method according to claim 1, characterized in that, The spatial mask matrix is ​​generated by marking the grids obscured by the physical terrain as missing values; the model's loss calculation and gradient backpropagation are performed only within the effective area of ​​the mask.

6. The method according to claim 1, characterized in that, The physical loss function includes terrain forcing loss based on no-slip boundary conditions.

7. The method according to claim 1, characterized in that, The loss function of the second-stage generative adversarial network is composed of a weighted sum of generative adversarial loss, pixel-level error loss, and frequency domain energy constraint loss based on fast Fourier transform.

8. A two-stage super-resolution reconstruction system for three-dimensional downscaling wind fields in complex terrain bridge site areas, characterized in that, The system includes: Acquisition module: Acquires high-fidelity simulation datasets from mesoscale meteorological models, and extracts low-resolution dynamic meteorological sequences and high-resolution static geographic features; Construction module: For multiple preset height layers, interpolation is used to construct corresponding multi-source heterogeneous input features, and a spatial mask matrix representing the mountain shading effect is constructed based on the real terrain height. Training module: Independently train the corresponding cascaded super-resolution reconstruction model for each altitude layer. The model includes a first-stage fully convolutional network for extracting large-scale meteorological background fields, and a second-stage generative adversarial network for predicting high-frequency spatial residual fields by combining the background fields. Reconstruction Module: Under the constraints of the spatial mask matrix and physical loss function, outputs high-resolution two-dimensional meteorological reconstruction fields at each altitude level, and stitches them together in the vertical spatial dimension to reconstruct a three-dimensional meteorological field assembly of the target area.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1-7.

10. A computer-readable storage medium for storing computer instructions, characterized in that, When the computer instructions are executed by the processor, they implement the steps of the method according to any one of claims 1-7.

Images