A progressive wind speed field deep downscaling method and device fusing multi-source data and spatiotemporal perception, electronic equipment, and storage medium

Abstract

This invention relates to the field of meteorological data processing technology, and in particular to a progressive wind speed field depth downscaling method, device, electronic equipment, and storage medium that integrates multi-source data and spatiotemporal perception. This application improves the accuracy of wind field reconstruction in complex terrain areas by introducing multi-scale terrain features and performing adaptive fusion; it adopts a progressive multi-level downscaling network architecture, effectively achieving stable downscaling from kilometer-level to hundred-meter-level scales and controlling error accumulation; and it utilizes a network trained with physical constraint loss for reconstruction, ensuring that the output wind field conforms to atmospheric dynamics in terms of divergence and vorticity, enhancing the physical rationality and scientific credibility of the results, thus providing an efficient solution for obtaining high-resolution, highly physically reliable near-surface wind fields.

Description

Technical Field

[0001] This invention relates to the field of meteorological data processing technology, and in particular to a progressive wind speed field depth downscaling method, device, electronic equipment, and storage medium that integrates multi-source data and spatiotemporal perception. Background Technology

[0002] The refined spatial distribution of near-surface wind speed fields is crucial for applications such as wind energy resource assessment, urban planning, architectural engineering design, and low-altitude flight safety. Currently, the main technical approaches to obtaining second spatial resolution wind speed fields include numerical weather prediction, statistical downscaling, dynamic downscaling, and emerging deep learning methods, but all of these methods have significant limitations.

[0003] First, operational numerical weather prediction models are limited by computational resources, and their output is typically a gridded field with kilometer-level resolution, which cannot meet the direct needs of detailed applications at the hundred-meter level. Second, while statistical downscaling methods can establish statistical relationships between high first-spatial-resolution data, they often ignore the spatial correlation and physical consistency of meteorological elements, resulting in limited model expressive power. Dynamic downscaling methods (such as the WRF model) solve atmospheric dynamic equations through nested grids, theoretically yielding high-precision results, but their computational cost is extremely high, making rapid operational deployment difficult. In recent years, deep learning-based image super-resolution technology has been introduced into meteorological element downscaling research and has made progress in scalar fields such as precipitation and temperature. However, existing methods have significant shortcomings when applied to vector wind fields: First, most models do not fully consider the physical characteristics of wind speed fields as vector fields (such as divergence and vorticity), lacking clear physical constraints, which may lead to reconstruction results that violate the basic principles of atmospheric dynamics; second, they fail to effectively integrate high-precision digital terrain data, while terrain is one of the decisive factors affecting near-surface wind fields, resulting in poor model performance in complex terrain areas; third, they usually adopt a single-resolution one-step reconstruction strategy, which is prone to error accumulation and detail distortion in large-scale downscaling tasks from kilometer to hundred-meter scale.

[0004] In summary, existing technologies struggle to achieve a good balance between computational efficiency, numerical accuracy, physical plausibility, and adaptability to complex terrain. Therefore, there is an urgent need for an efficient downscaling method that can adaptively fuse multi-source data, explicitly introduce physical constraints, and effectively utilize terrain and temporal evolution information to achieve near-surface wind speed field reconstruction with second spatial resolution and high physical reliability. Summary of the Invention

[0005] In view of this, the purpose of the present invention is to provide a progressive wind speed field depth downscaling method, device, electronic device, and storage medium that integrates multi-source data and spatiotemporal perception to improve the efficiency of wind speed field reconstruction.

[0006] In a first aspect, embodiments of the present invention provide a progressive wind speed field depth downscaling method that integrates multi-source data and spatiotemporal awareness, including:

[0007] Acquire gridded wind speed data, station observation data, and DEM data at the first spatial resolution;

[0008] Topographic features are extracted from DEM data to obtain multi-scale topographic features;

[0009] Based on gridded wind speed data at the first spatial resolution, station observation data, and multi-scale terrain features, the data are input into a trained progressive multi-level downscaling network for iterative reconstruction, and the output wind speed field at the second spatial resolution is generated.

[0010] The progressive multi-level downscaling network is trained by minimizing a composite loss function that includes physical constraint loss. The physical constraint loss is used to constrain the divergence of the reconstructed wind field in the horizontal direction and the vorticity in the vertical direction. The second spatial resolution is higher than the first spatial resolution.

[0011] In conjunction with the first aspect, the steps for extracting terrain features from DEM data to obtain multi-scale terrain features include:

[0012] Based on DEM data, a multi-channel topographic feature map is calculated, including elevation features, slope features, aspect features, curvature features, and roughness index features.

[0013] The multi-channel terrain feature map is input into the multi-scale terrain feature extractor, and primary terrain features corresponding to different spatial scales are extracted through at least two parallel convolutional branches with receptive fields of different sizes.

[0014] By using learnable weights, multiple primary terrain features are adaptively weighted and fused to generate a terrain embedding vector, which is then used as a multi-scale terrain feature.

[0015] In conjunction with the first aspect, the steps for calculating a multi-channel topographic feature map based on DEM data, including elevation features, slope features, aspect features, curvature features, and roughness index features, include:

[0016] Elevation features are directly obtained based on the grid point elevation values ​​in the DEM data.

[0017] Based on elevation values, slope and aspect characteristics are calculated through spatial difference.

[0018] Curvature characteristics are calculated using second-order differentials based on elevation values.

[0019] Based on the elevation value, the roughness index characteristic is obtained by calculating the standard deviation of the elevation values ​​within a predetermined local window.

[0020] In conjunction with the first aspect, the parallel convolutional branches include a first branch with a kernel size of 3×3, a second branch with a kernel size of 5×5, and a third branch with a kernel size of 7×7.

[0021] In conjunction with the first aspect, each downscaling stage of the progressive multi-level downscaling network includes a terrain-aware residual dense network; the terrain-aware residual dense network includes a terrain-wind speed coupled attention module.

[0022] The steps involved in iteratively reconstructing a wind speed field at a second spatial resolution, based on gridded wind speed data at a first spatial resolution, station observation data, and multi-scale terrain features, are as follows:

[0023] The system receives gridded wind speed data at the first spatial resolution from multiple consecutive time intervals as input, and performs time-dimension weighted fusion of the feature maps corresponding to multiple time intervals through a time attention module to obtain fused features.

[0024] The fused features, station observation data, and multi-scale terrain features are input into a progressive multi-level downscaling network. The network is iteratively processed through multiple connected downscaling stages to output the final second spatial resolution wind speed field.

[0025] The input to the current downscaling stage consists of the wind speed field output from the previous downscaling stage, intermediate features extracted from the network in the previous downscaling stage, and terrain features at the resolution of the current downscaling stage.

[0026] Combining the first aspect, the steps of inputting fused features, station observation data, and multi-scale terrain features into a progressive multi-level downscaling network, and iteratively processing through multiple serial downscaling stages to output the final second spatial resolution wind speed field include:

[0027] The adaptive fusion module performs weighted fusion of features from gridded wind speed data and features from station observation data to generate multi-source wind speed features for the current stage.

[0028] The multi-source wind speed features and the multi-scale terrain features corresponding to the current stage are input into the terrain-aware residual dense network of this stage. The features are modulated through its terrain-wind speed coupled attention module, and the wind speed features enhanced with terrain information are output.

[0029] Subpixel convolution upsampling is performed on the wind speed features enhanced with terrain information to generate the second spatial resolution wind speed field output at the current stage.

[0030] Combining the first aspect, the composite loss function includes reconstruction loss, multi-scale supervision loss, and physical constraint loss; the physical constraint loss includes divergence constraint loss and vorticity constraint loss calculated based on the two-dimensional spatial partial derivative of the predicted wind.

[0031] Secondly, embodiments of this application also provide a progressive wind speed field depth downscaling device that integrates multi-source data and spatiotemporal perception, the device comprising:

[0032] The acquisition module is used to acquire gridded wind speed data, station observation data, and DEM data at the first spatial resolution.

[0033] The extraction module is used to extract terrain features from DEM data to obtain multi-scale terrain features;

[0034] The wind speed field reconstruction module is used to input gridded wind speed data, station observation data and multi-scale terrain features based on the first spatial resolution into a trained progressive multi-level downscaling network for iterative reconstruction and output a wind speed field with the second spatial resolution.

[0035] The progressive multi-level downscaling network is trained by minimizing a composite loss function that includes physical constraint loss. The physical constraint loss is used to constrain the divergence of the reconstructed wind field in the horizontal direction and the vorticity in the vertical direction. The second spatial resolution is higher than the first spatial resolution.

[0036] Thirdly, this application provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor runs the computer program to cause the electronic device to perform the above-described method.

[0037] Fourthly, this application provides a readable storage medium storing computer program instructions, which are read and executed by a processor to perform the above-described method.

[0038] The embodiments of this invention bring the following beneficial effects: This application provides a progressive wind speed field depth downscaling method that integrates multi-source data and spatiotemporal perception, including: acquiring gridded wind speed data, station observation data, and DEM data at a first spatial resolution; extracting terrain features from the DEM data to obtain multi-scale terrain features; inputting the gridded wind speed data, station observation data, and multi-scale terrain features at the first spatial resolution into a trained progressive multi-level downscaling network for iterative reconstruction, and outputting a wind speed field at a second spatial resolution; wherein, the progressive multi-level downscaling network is trained by minimizing a composite loss function that includes physical constraint loss, the physical constraint loss being used to constrain the divergence in the horizontal direction and the vorticity in the vertical direction of the reconstructed wind field; the second spatial resolution is higher than the first spatial resolution.

[0039] This application improves the accuracy of wind field reconstruction in complex terrain areas by introducing multi-scale terrain features and performing adaptive fusion; it adopts a progressive multi-level downscaling network architecture, which effectively achieves stable downscaling from the kilometer level to the hundred-meter level and controls error accumulation; and it uses a network trained with physical constraint loss for reconstruction, which ensures that the output wind field conforms to the laws of atmospheric dynamics in terms of divergence and vorticity, enhancing the physical rationality and scientific credibility of the results, thus providing an efficient solution for obtaining high-resolution, high-physical-credibility near-surface wind fields.

[0040] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention are realized and obtained in accordance with the structures particularly pointed out in the description, claims and drawings.

[0041] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0042] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0043] Figure 1 This is a flowchart illustrating the near-surface wind speed field downscaling method based on terrain perception and physical constraints provided in an embodiment of the present invention.

[0044] Figure 2 for Figure 1 A schematic diagram of the training process for the progressive multi-level downscaling network in the provided method;

[0045] Figure 3 A schematic diagram of a near-surface wind speed field downscaling device based on terrain perception and physical constraints provided in an embodiment of the present invention;

[0046] Figure 4 This is a schematic diagram of the electronic device structure provided in an embodiment of the present invention.

[0047] Figure label:

[0048] 10 - Acquisition module, 20 - Extraction module, 30 - Wind speed field reconstruction module;

[0049] 130 - Processor, 131 - Memory, 132 - Bus, 133 - Communication interface. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions 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, 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.

[0051] To facilitate understanding of this embodiment, the application scenarios and design concepts of this application embodiment will be briefly introduced below.

[0052] Refined near-surface wind speed fields are crucial for wind power and urban planning. However, existing technologies have significant limitations: traditional numerical models output kilometer-level resolution, which cannot meet the needs of applications at the hundred-meter level; dynamic downscaling, while highly accurate, comes at a huge computational cost; and existing deep learning methods generally do not fully consider the vector physical characteristics of wind fields (such as divergence and vortex constraints) and the influence of high-precision terrain, resulting in insufficient accuracy and poor physical plausibility of reconstruction results in complex terrain areas.

[0053] Based on this, this application provides a progressive wind speed field depth downscaling method, apparatus, electronic device, and storage medium that integrates multi-source data and spatiotemporal awareness.

[0054] Example 1

[0055] This application proposes a progressive wind speed field depth downscaling method that integrates multi-source data and spatiotemporal awareness, combining... Figure 1 As shown, the method includes:

[0056] S110 acquires gridded wind speed data, station observation data, and DEM data at the first spatial resolution.

[0057] S120 extracts terrain features from DEM data to obtain multi-scale terrain features.

[0058] S130, based on gridded wind speed data at the first spatial resolution, station observation data, and multi-scale terrain features, is input into a trained progressive multi-level downscaling network for iterative reconstruction, and outputs a wind speed field at the second spatial resolution.

[0059] The progressive multi-level downscaling network is trained by minimizing a composite loss function that includes physical constraint loss. The physical constraint loss is used to constrain the divergence of the reconstructed wind field in the horizontal direction and the vorticity in the vertical direction. The second spatial resolution is higher than the first spatial resolution.

[0060] In step S110, gridded wind speed data refers to the near-surface wind speed vector field generated by global or regional numerical weather prediction models (such as ECMWF, GFS) and output on a regular latitude and longitude grid. It typically contains two horizontal components: east-west (U) and north-south (V). Its characteristics are wide coverage and good physical consistency, but its spatial resolution is relatively coarse (kilometer level), and it cannot resolve the impact of local details such as topography on the wind field.

[0061] In this embodiment, model products from the European Centre for Medium-Range Weather Forecasts (ECMWF) are used, with a typical spatial resolution of approximately 0.125° × 0.125° (equivalent to approximately 14 km grid spacing). This data represents the large-scale circulation background of the atmospheric dynamic system simulation and is the main input and starting point for the downscaling task.

[0062] Station observation data refers to near-surface (e.g., at a height of 10 meters) wind speed and direction data measured by automatic weather stations distributed on the ground. Its characteristics include high accuracy, but it is spatially unevenly distributed, appearing as discrete points, and may contain missing data or local representativeness errors.

[0063] The observation data from this site, serving as ground truth, plays a crucial role in two aspects: 1) during the model training phase, it is used to construct the ground truth objective for supervised learning or as an important source of it; 2) during the model inference or data fusion phase, it provides high-precision point observations to calibrate and constrain the bias of grid data.

[0064] Digital Elevation Model (DEM) data: A dataset representing ground elevation information stored digitally; it is a digital representation of terrain features. In this embodiment, high spatial resolution DEM data is used, such as SRTM or ASTER GDEM with a resolution of 30 to 90 meters. This resolution is significantly higher than the input gridded wind speed data, aiming to provide the model with sufficient detailed information to resolve key terrain features such as ridges, valleys, and slope variations.

[0065] Understandably, DEM data is the physical foundation for terrain sensing capabilities. Wind fields are strongly modulated by topographic dynamics, obstruction, flow around, and thermal effects, and DEM data is an indispensable input for quantifying and incorporating these influences.

[0066] Understandably, data acquisition is not simply reading; it typically involves a preprocessing step of data cleaning and integration to ensure data availability, guarantee model input quality, and achieve effective fusion of multi-source data. In this embodiment, the preprocessing process includes four steps, specifically:

[0067] Data standardization: Wind speed data is Z-score normalized, and the original wind speed components (U / V wind) are dimensionless, calculated using the following formula:

[0068]

[0069] in, This represents the normalized wind speed component. Represents the original wind speed component (which can be U-wind or V-wind); This represents the average value of the wind speed component calculated across the entire training set; This represents the standard deviation of the wind speed component calculated across the entire training set. This operation aims to scale data of different magnitudes and distributions to a similar numerical range, thereby accelerating the training convergence process of deep learning models and improving training stability.

[0070] Spatial consistency: Unify all input data (grid wind field, station observations, DEM) to the same geographic coordinate system (such as WGS84). For spatially discrete station observation data, map them to a regular grid that matches the low-resolution grid data through nearest neighbor interpolation and other methods, thereby achieving preliminary alignment between grid points and station data in the spatial dimension and laying the foundation for subsequent feature-level fusion.

[0071] Time consistency: To address potential time resolution differences between different data sources (e.g., pattern outputs are updated every 6 hours, while site records are updated every 1 hour), linear interpolation or nearest neighbor methods are used to align all data to a series of unified timestamps, ensuring that the input at each moment is synchronized in the time dimension to support subsequent spatiotemporal evolution modeling.

[0072] Quality control: The raw data were examined according to meteorological standards, and obviously unreasonable outliers (such as extreme wind speeds exceeding the physical limits) were removed. For missing data, interpolation methods based on spatiotemporal correlation were used to fill in the gaps, in order to construct a complete and continuous training sample set.

[0073] In conjunction with the first aspect, step S120 includes:

[0074] S121, based on DEM data, calculates a multi-channel topographic feature map including elevation features, slope features, aspect features, curvature features, and roughness index features.

[0075] In step S121, starting from the grid point elevation values ​​of the original DEM data, a series of standard geospatial algorithms are used to calculate and generate five derived terrain parameters with clear hydrodynamic significance, which together constitute a set of multi-channel feature maps.

[0076] Among them, the elevation feature H(i,j) represents the altitude. This feature directly affects atmospheric pressure and temperature stratification, and is the basic elevation field that generates thermal and dynamic processes such as topographic waves and valley winds.

[0077] The slope characteristic S(i,j) refers to the steepness of the Earth's surface inclination. Slope determines the mechanical obstruction and lifting effect of the ground on airflow. Steep slopes force airflow to rise, which may lead to acceleration; gentle slopes have a smaller impact.

[0078] The slope aspect characteristic A(i,j) refers to the direction the slope faces (e.g., north-facing or south-facing). This characteristic is directly related to sunlight and radiation reception and is a key factor leading to thermal differences (e.g., temperature difference between sunny and shady slopes) and consequently triggering local circulation (e.g., valley winds). It also determines whether the terrain faces the windward or leeward direction.

[0079] The curvature feature C(i,j) describes a measure of the concavity and convexity of the Earth's surface (positive for convexity, such as a ridge, and negative for concavity, such as a valley). Over convex terrain such as a ridge, airflow diverges horizontally, and wind speed may increase; over concave terrain such as a valley, airflow converges horizontally, and wind speed may decrease. It characterizes the guiding effect of the shape of the second derivative of the terrain on the convergence and divergence of airflow.

[0080] The roughness index R(i,j) is obtained by calculating the standard deviation of elevation within a local window (such as a 3x3 grid) and reflects the severity of small-scale surface undulations. This characteristic is directly related to surface friction resistance. Areas with high roughness (such as forests and cities) can dissipate more wind energy, reducing near-surface wind speeds; while smooth areas (such as water surfaces and grasslands) have low friction.

[0081] In step S121, the single height information is expanded into a set of features with clear physical orientation, resulting in a multi-channel terrain feature map [H, S, A, C, R]. This is equivalent to providing the progressive multi-level downscaling network with a structured physical dictionary for understanding how terrain affects wind, so that the learning process of the progressive multi-level downscaling network is no longer guessing from scratch, but is carried out within the framework of physical laws.

[0082] S122, input the multi-channel terrain feature map into the multi-scale terrain feature extractor, and extract primary terrain features corresponding to different spatial scales through at least two parallel convolution branches with receptive fields of different sizes.

[0083] The five-channel feature map [H, S, A, C, R] obtained in the previous step is input into a dedicated convolutional neural network module called a "multi-scale terrain feature extractor". This module uses multiple parallel branches with different kernel sizes to simultaneously extract terrain patterns at different spatial scales from the feature map, thus achieving scale decoupling of terrain information. This enables the progressive multi-level downscaling network to determine whether the terrain influence mainly comes from macroscopic patterns or microscopic details, laying the foundation for subsequent intelligent feature fusion and targeted modulation.

[0084] S123 uses learnable weights to adaptively weight and fuse multiple primary terrain features to generate a terrain embedding vector, which is then used as a multi-scale terrain feature.

[0085] Step S123 fuses the primary terrain features representing different scales extracted from the parallel branches in step S122. This is not a simple addition, but rather the neural network automatically learns a set of weights. Based on the characteristics of the current region and the task objective, it adaptively assigns appropriate importance to features at different scales, ultimately fusing them to generate a high-level terrain embedding vector with a fixed dimension (e.g., 256 dimensions).

[0086] In conjunction with the first aspect, step S121 includes:

[0087] S1211, based on the elevation values ​​of grid points in DEM data, directly obtains elevation features.

[0088] The core of this step is to directly extract the elevation value of each grid point (i, j) from the DEM data to form the most basic terrain feature, namely the elevation feature H(i, j). Essentially, it involves direct data reading and mapping, without complex numerical calculations, because the DEM data itself stores surface elevation information in a regular grid structure, and therefore can be directly accessed.

[0089] Understandably, elevation is the most fundamental and intuitive expression of topography, forming the basis for calculating all subsequent derived topographic features (such as slope and curvature). From an atmospheric physics perspective, altitude is directly related to key meteorological elements such as air pressure distribution and vertical temperature lapse rate, and is the fundamental driving force and thermodynamic factor causing local climate phenomena such as valley winds, mountain airflows, and thermal circulation. By directly transmitting raw elevation data, lossless input of topographic information is ensured. This allows the subsequent progressive multi-level downscaling network to not only perform inference based on derived features, but also to autonomously uncover the potential statistical relationship between elevation and wind speed (such as the average wind speed characteristics corresponding to specific altitude zones), enhancing the model's expressive power and adaptability.

[0090] S1212, based on elevation values, calculates slope and aspect characteristics through spatial difference.

[0091] The core of this step is to calculate the slope characteristics S(i,j) and aspect characteristics A(i,j) of the surface section at each point based on the elevation values ​​of the grid points using the first-order spatial difference method. Usually, the Horn algorithm or the central difference method is used to calculate the slope angle S and aspect angle A by solving the horizontal gradient (dz / dx, dz / dy) of the elevation in the east-west and north-south directions.

[0092] The formulas for calculating the slope angle S and the slope aspect angle A are as follows:

[0093]

[0094]

[0095] The role and effect of this step are significant: First, it characterizes the first-order dynamic morphology of the Earth's surface. Slope directly quantifies the mechanical obstruction strength of the airflow, determining whether the airflow rises, flows around, or accelerates. Slope aspect determines the relative relationship between the terrain surface and the prevailing wind direction (windward or leeward) and reflects the thermal unevenness caused by differences in solar radiation, which is a key factor in triggering local thermal circulation (such as valley winds). Second, it provides accurate dynamic and thermal cues, transforming static elevation into dynamic indicators driving the wind field response. This allows the model to identify dynamic acceleration zones, shading zones, and thermal convergence lines caused by terrain, providing direct physical guidance for refined wind field reconstruction.

[0096] S1213, based on elevation values, calculates curvature characteristics using second-order differentials.

[0097] This step aims to quantify the concavity and convexity of the land surface profile, i.e., the curvature feature C(i,j), based on the elevation field and using second-order spatial derivative calculations. The calculation principle can employ the Laplace operator approximation or obtain the profile curvature and planar curvature through local surface fitting. This step is used to reveal the intrinsic mechanism of airflow convergence and divergence. The curvature feature is directly related to the horizontal convergence and diffusion of airflow; concave terrain (such as valleys and basins) guides horizontal airflow convergence, often accompanied by weakened wind speed or upward motion; convex terrain (such as ridges and hilltops) leads to horizontal airflow divergence, easily causing wind speed acceleration or downward motion. Providing detailed topographic structure information, as the first derivative of slope, curvature can keenly capture key topographic abrupt changes such as ridgelines, valley lines, and slope breakpoints. These areas are physically sensitive zones where wind field separation, acceleration, or turbulence enhancement occurs, thus greatly enhancing the model's ability to reconstruct complex wind field structural details.

[0098] S1214, based on elevation values, obtains the roughness index characteristics by calculating the standard deviation of elevation values ​​within a predetermined local window.

[0099] This step generates a roughness index feature R(i,j) by defining a local moving window (e.g., a 3×3 or 5×5 grid) and calculating the statistical standard deviation of all elevation values ​​within the window. This step quantifies the surface friction effect. The roughness index is an effective proxy for the surface dynamic roughness length, directly reflecting the frictional dissipation capacity of the underlying surface for airflow. High-value areas (e.g., forests, cities, rugged rocks) correspond to strong momentum dissipation and reduced wind speed, while low-value areas (e.g., water surfaces, flat grasslands) correspond to weak friction and higher wind speeds. Sub-grid-scale process parameterization is introduced. Since the input coarse-resolution wind field cannot resolve micro-topography and surface cover details, the frictional effect at the sub-grid scale is explicitly introduced into the model in statistical form. This allows the deep learning system to learn the physical laws governing the drag attenuation of large-scale wind fields under different roughnesses, significantly improving the realism and physical consistency of near-surface wind speed simulation, especially in complex underlying surface areas with strong land-atmosphere interactions.

[0100] In conjunction with the first aspect, the parallel convolutional branches in step S122 include a first branch with a kernel size of 3×3, a second branch with a kernel size of 5×5, and a third branch with a kernel size of 7×7.

[0101] Step S122 includes:

[0102] S1221, input the multi-channel terrain feature map into the multi-scale terrain feature extractor, extract local features through the first branch, extract mesoscale features through the second branch, and extract global features through the third branch.

[0103] In step S122, the multi-channel terrain feature map [H, S, A, C, R] output in step S121 is used as input and simultaneously fed into three parallel convolutional neural network branches. These three branches define different spatial receptive fields based on the difference in their convolutional kernel sizes, thereby focusing on extracting terrain patterns at different spatial scales.

[0104] The first branch (3×3 convolution): uses a small-sized convolution kernel with a limited receptive field, primarily operating on adjacent grid cells. This branch is responsible for extracting local features. This captures microscopic topographic details and abrupt changes, such as the edges of individual hills, small-scale slope transitions, and micro-topographic undulations. These features directly affect the most immediate airflow around, uplift, or micro-turbulence.

[0105] The second branch (5×5 convolution): uses a medium-sized convolution kernel, providing a larger receptive field. This branch is responsible for extracting mesoscale features. It can perceive topographic structures such as the overall tilt trend of a continuous hillside, the outline of a medium-sized valley or ridge, and the shape of a plateau. The impact of these structures on the wind field manifests as regional acceleration (such as windward slopes), blocking, or channeling effects;

[0106] The third branch (7×7 convolution): uses a large-size convolutional kernel, providing a wide global receptive field. This branch is responsible for extracting global features. The purpose of wind spectroscopy is to understand macroscopic topographic patterns, such as the orientation of entire mountain ranges, the extent of large basins, and the overall topographic framework of watersheds. These macroscopic patterns determine the basic paths of prevailing winds, large-scale around-flow or uplift patterns, and provide background constraints for wind fields.

[0107] Understandably, the modulation effect of terrain on near-surface wind fields is multi-scale coupled. A single-scale convolutional kernel cannot fully represent this complexity. In step S122, by designing a multi-branch parallel structure, the basic, explicit physical features (elevation, slope, etc.) are transformed into a set of implicit, multi-scale deep features. This enables the model to observe the influence of different levels of terrain, providing high-quality primary feature raw materials for the intelligent, adaptive multi-scale information fusion in the subsequent step S123. This is a key step in realizing the leap from physical parameterization to data-driven intelligent understanding of terrain perception capabilities.

[0108] In step S123, the feature fusion module then adaptively weights and combines the features to obtain the final terrain embedding vector. . Specific formula:

[0109]

[0110]

[0111]

[0112]

[0113] in, Local features extracted using small convolutional kernels Mesoscale features extracted using medium-sized convolution kernels It is a global feature, the final terrain feature. Then, weights are automatically learned through a learnable attention mechanism. , , It is calculated by combining three features.

[0114] In conjunction with the first aspect, each downscaling stage of the progressive multi-level downscaling network includes a terrain-aware residual dense network; the terrain-aware residual dense network includes a terrain-wind speed coupled attention module. Step S130 includes:

[0115] S131 receives gridded wind speed data at the first spatial resolution from multiple consecutive time intervals as input, and performs time-dimension weighted fusion of the feature maps corresponding to multiple time intervals through a time attention module to obtain fused features.

[0116] S132 inputs the fusion features, station observation data and multi-scale terrain features into a progressive multi-level downscaling network, and iterates through multiple downscaling stages in sequence to output the final second spatial resolution wind speed field.

[0117] The input to the current downscaling stage consists of the wind speed field output from the previous downscaling stage, intermediate features extracted from the network in the previous downscaling stage, and terrain features at the resolution of the current downscaling stage.

[0118] Step S130 is the core execution stage of the entire downscaling method. Its goal is to gradually, stably, and physically reconstruct a high-resolution near-surface wind speed field from the preprocessed multi-source input data through a trained progressive multi-level downscaling network. This process is no longer a simple interpolation or single transformation as in traditional methods, but a complex intelligent reasoning process that integrates spatiotemporal attention, multi-source intelligent fusion, terrain-aware modulation, and progressive refinement.

[0119] In step S131, low-resolution gridded wind speed data from multiple consecutive historical time points (e.g., the past T = 3-6 time points) are received as input. These data form a sequence in the time dimension. Through a time attention module (e.g., SEBlock), the model automatically analyzes the importance of the feature map of each time point in the sequence for the current downscaling task, learns and assigns a set of time attention weights (e.g., SEBlock). Subsequently, a new fused feature is obtained through weighted fusion, which can centrally reflect the recent evolution trend and key state of the wind field.

[0120] Specifically, step S131 includes:

[0121] S1311, For each time period, perform preliminary feature extraction on the U / V wind field data of that time period to obtain the time period feature map.

[0122] For low-resolution gridded U / V wind field data from multiple consecutive time intervals (e.g., T time intervals), preliminary spatial feature extraction is performed. Each time interval's data is treated as an independent spatial field and processed by a shallow convolutional network with shared weights. This transforms the original U / V dual-channel wind speed data into a time-series feature map Ft (where t=1,2…T) containing richer spatial pattern information. Typically, one or more standard convolutional layers are used to process the U / V wind field data for each time interval. Shared weights mean using the same convolutional kernel across different time intervals, ensuring consistency in feature extraction methods and enabling subsequent attention mechanisms to fairly compare the quality or importance of features from different time intervals. This transforms the raw, numerical wind speed field into a deeper feature representation that better represents its spatial structure, gradient information, and local patterns. This provides comparable, high-quality input for subsequent temporal attention calculations. Early feature compression reduces the amount of data that subsequent modules (such as the attention mechanism) need to process, improving overall computational efficiency.

[0123] S1312, the temporal attention module calculates the attention score for each temporal feature through global average pooling, fully connected layers, and activation functions.

[0124] The set of temporal feature maps {F1, F2…FT} obtained in step S1311 is input into the temporal attention module. The core of this module is to evaluate and quantify the relative value of the information contained in each temporal step for the current downscaling task (i.e., predicting the target time or state), and output a series of corresponding attention scores St.

[0125] Taking SEBlock-style attention as an example, firstly, global average pooling is performed on each temporal feature map Ft, compressing it from a two-dimensional spatial feature map into a channel description vector. This operation aggregates information from the entire spatial domain, obtaining the global mean of each feature channel, representing the global state of the wind field at that time. Then, the aforementioned channel description vector is input into a lightweight quantum network consisting of fully connected layers (FC) and nonlinear activation functions (such as sigmoid). The FC learns the complex dependencies and nonlinear interactions between time periods, determining which time periods are more critical for predicting the current target (e.g., nearby moments are generally more important than distant moments, but certain historical turning points may also be highly valuable). Finally, the output of the FC is mapped to the (0, 1) interval using the sigmoid activation function, generating a normalized temporal attention score (weight) St. The magnitude of St directly reflects the importance of the t-th time feature.

[0126] Based on the above operations, the subsequent progressive multi-level downscaling network no longer treats all historical moments equally. Instead, it learns to intelligently focus on the moments with the highest information content according to the specific weather conditions and downscaling task. For example, during the passage of a front, it may pay more attention to the moments before and after the change; under stable weather conditions, it may rely more on the most recent moments. Thus, the progressive multi-level downscaling network trained based on this method can automatically identify and strengthen the key frames that are most indicative of understanding the current wind field evolution from the time series, essentially refining information in the time dimension.

[0127] S1313, combine the feature maps of each time step to calculate the calibration feature map after feature fusion.

[0128] Step S1313 uses the temporal attention score St calculated in step S1312 as a weight to perform a weighted linear summation on the original temporal feature map Ft, generating a single calibration feature map that integrates the essence of information from multiple temporal times. The calculation expression is as follows:

[0129]

[0130] in, Representative time calibration feature map, The feature map representing the t-th time period. This represents element-wise multiplication. This represents the time attention weight at time t.

[0131] Output fusion features Instead of being a portrait of an isolated moment, the wind field is now a contextualized representation that encapsulates its recent dynamic evolution. It contains both direct information about the current state and encodes the process information that led to it. This fused feature, as a key input, is fed into the subsequent progressive downscaling network. This means that the progressive multi-level downscaling network's inference process during spatial super-resolution reconstruction is implicitly constrained by the physical laws of temporal evolution, helping to generate more coherent and reasonable high-resolution wind field sequences in the temporal dimension, particularly improving the reconstruction accuracy for processes such as diurnal variation and weather system movement.

[0132] In conjunction with the first aspect, step S132 includes:

[0133] S1321 uses an adaptive fusion module to perform weighted fusion of features from gridded wind speed data and features from station observation data to generate multi-source wind speed features for the current stage.

[0134] S1322 inputs multi-source wind speed features and multi-scale terrain features corresponding to the current stage into the terrain-aware residual dense network of this stage. The feature is modulated through its terrain-wind speed coupled attention module, and the wind speed features enhanced with terrain information are output.

[0135] S1323 performs sub-pixel convolution upsampling on the wind speed features enhanced with terrain information to generate the second spatial resolution wind speed field output at the current stage.

[0136] Step S132 defines a complete data processing pipeline within a single downscaling stage in a progressive multi-level downscaling network. This pipeline takes the output of the previous stage, inter-stage transfer features, and current resolution terrain features as inputs. Through a series of precise operations such as multi-source fusion, terrain modulation, feature refinement, and spatial upsampling, it generates a high-resolution wind speed field for the current stage. This process will be executed cyclically in K cascaded stages and is the core operation unit for achieving a gradual and stable improvement in resolution.

[0137] In step S1321, at the beginning of each downscaling stage, the gridded wind speed characteristics derived from numerical weather prediction are... Wind speed characteristics at stations derived from ground observations Intelligent fusion is performed to generate unified multi-source wind speed characteristics. Specifically, a light quantum network (data quality assessment network) is used to evaluate two feature sources, outputting two sets of parameters: data weights. , and the quality score characterizing the reliability of each feature at the current location. and The following formula is then used for fusion calculation:

[0138]

[0139] The combined multi-source wind speed characteristics for the current stage. For grid data features, For site data characteristics, Q represents the data weights, and Q represents the quality score.

[0140] In step S1321, in areas with sparse or no meteorological stations, the model automatically relies on grid features with good physical consistency; in areas with dense stations, the role of high-precision observation features is strengthened to achieve local correction of grid system bias. By introducing a quality assessment mechanism, noise interference caused by missing or abnormal observation data is reduced, providing a cleaner and more reliable initial wind speed feature field for subsequent processing.

[0141] In step S1322, the multi-source wind speed features generated in step S1321 are... Multi-scale terrain embedding vector corresponding to the current stage Both are fed into the Terrain-Aware Residual Dense Network (TA-RDN) at each downscaling stage of the progressive multi-level downscaling network. The core component of this network, the Terrain-Wind Speed ​​Coupled Attention Module, is responsible for deep interaction and modulation between the two, outputting wind speed features significantly enhanced by terrain information. .

[0142] In this embodiment, the terrain-wind speed coupled attention module includes a parallel channel attention module and a spatial attention module.

[0143] Channel attention module for terrain features (Right now Global average pooling and fully connected operations are performed, and channel attention weights are generated using the sigmoid function. These weights represent the differences in the importance of different terrain factors (such as elevation, slope, and roughness) to the overall wind speed in the current area, and their expression is as follows:

[0144]

[0145] Where FC() represents a fully connected layer, This indicates global average pooling.

[0146] Spatial attention module for wind speed features F ( Max pooling and average pooling are performed separately, and the results are concatenated and then passed through a 7×7 convolution and a sigmoid function to generate spatial attention weights. This weight map highlights key spatial locations of topographic dynamics, such as ridges, valleys, windward slopes, and leeward slopes. Its expression is as follows:

[0147]

[0148] in, Indicates channel splicing. This indicates max pooling. This indicates average pooling.

[0149] Subsequently, the terrain features are multiplied by both attention weights simultaneously to achieve precise and adaptive modulation of key terrain factors at critical locations, as expressed in the following expression:

[0150]

[0151] in, To enhance wind speed characteristics.

[0152] Finally, the modulated terrain features and wind speed features are further fused in multiple residual dense blocks (RDBs) within the TA-RDN, and the expression is as follows:

[0153]

[0154] in, This is the output feature map of the l-th convolutional layer in the residual dense block. For a series of nonlinear transformations performed on the l-th layer, For feature splicing, This represents the input feature map of the current residual dense block.

[0155] This ensures that all hierarchical features (including original terrain clues) can be effectively reused and guarantees stable gradient propagation, promoting the training of deep networks. As a result, it can selectively and strategically utilize terrain knowledge to correct and enhance wind speed characteristics based on physical laws (terrain dynamics or thermal effects), rather than simply superimposing them. This is beneficial for improving the accuracy and physical rationality of fine-grained wind field reconstruction in complex terrain areas (such as mountains and hills).

[0156] Step S1323 performs a sub-pixel convolution (Pixel Shuffle) operation on the enhanced wind speed features output in step S1322, which have undergone deep terrain modulation and feature extraction, to increase the spatial resolution of the feature map to a preset target multiple (such as 2 times) for the current stage, thereby generating the output wind speed field for the current stage.

[0157] Understandably, subpixel convolution does not directly interpolate in the spatial domain, but rather reorganizes the data along the channel dimension. For example, to increase the resolution by a factor of 2 (r=2), the number of channels in the feature map is first increased to r² (i.e., 4) times the original number. Then, periodic shuffling is used to rearrange these channels, combining them into a high-resolution feature map with the channel count restored to normal. Compared to bilinear or bicubic interpolation, subpixel convolution is a learnable, content-based upsampling method. It can learn how to generate more reasonable details from rich features, effectively mitigating the blurring and checkerboard artifacts caused by traditional methods. This results in a higher-resolution wind speed field that can be used in the next stage or as the final output, a key operation enabling the progressive multi-level downscaling network strategy.

[0158] In this embodiment, the progressive multi-level downscaling network comprises three cascaded downscaling stages. The core unit of each stage is a Terrain-Aware Residual Dense Network (TA-RDN), and each TA-RDN contains six residual dense blocks (RDBs) by default. Each stage achieves a ×2 upsampling through sub-pixel convolution, thereby achieving a total amplification of 8 times from the input low-resolution wind field to the output high-resolution wind field through the cascaded processing of the three stages. This progressive architecture, in conjunction with the terrain modulation mechanism, enables stable and precise adaptive reconstruction of wind speed characteristics.

[0159] Combining the first aspect, the composite loss function includes reconstruction loss, multi-scale supervision loss, and physical constraint loss; the physical constraint loss includes divergence constraint loss and vorticity constraint loss calculated based on the two-dimensional spatial partial derivative of the predicted wind.

[0160] The progressive multi-level downscaling network in this application employs a composite loss function, which is a multi-objective joint optimization loss function specifically designed for wind speed field downscaling tasks. Its core purpose is to guide the model to achieve three key objectives simultaneously during training:

[0161] Numerical high fidelity: making the predicted wind speed as close as possible to the actual observation in numerical terms;

[0162] Process stability: Ensure that the intermediate results at each step are accurate and reliable during the progressive scale-up process;

[0163] Physical rationale: The forced reconstruction of the wind field complies with basic laws of fluid dynamics.

[0164] This function does not use a single loss, but rather incorporates reconstruction loss. Multi-scale monitoring loss and physical constraint loss The integration is performed using a weighted summation method, and its general expression is:

[0165]

[0166] in, It is the total loss function used to optimize the model;

[0167] It is the reconstruction loss, used to ensure that the predicted wind speed is close to the actual wind speed in numerical terms;

[0168] It is a multi-scale supervised loss, used to ensure the accuracy of intermediate results at each downscaling stage and effectively control error accumulation;

[0169] It is a physical constraint loss, used to ensure that the reconstructed wind field conforms to the laws of atmospheric dynamics and improve the physical rationality of the results;

[0170] and These are hyperparameters, set to [default value]. , Used for regulation and relative to reconstruction loss The strength of the impact is determined by selecting a combination of comprehensive indicators through a grid search of the validation set.

[0171] Reconstruction loss is a fundamental loss term that ensures the numerical accuracy of the model. It directly calculates the difference between the high-resolution predicted wind speed field and the corresponding high-resolution true wind speed field (ground value) in the final output of the model. It usually adopts mean squared error (MSE) or mean absolute error (MAE), and its goal is to minimize the error per pixel so that the predicted field approximates the true field in terms of overall distribution and intensity.

[0172] Multi-scale supervised loss is a key setting for progressive multi-stage network architectures. It supervises not only the final output but also the wind speed field output at each intermediate downscaling stage. In each training iteration, the output of each stage of the network is compared with the true wind speed field at the corresponding resolution (obtained by downsampling the full "ground truth") to calculate the loss. This prevents errors generated in lower levels from being amplified in later stages, ensuring that each level produces reasonable results. It provides direct and effective gradient feedback for both deep and shallow layers of the network, stabilizing the training process, and is particularly suitable for high-magnification super-resolution tasks.

[0173] The physical constraint loss does not directly rely on ground truth data, but rather, based on physical laws, imposes constraints on the derivative physical properties of the predicted wind field to ensure its dynamic characteristics are reasonable, thereby improving the physical reliability of the results. This loss consists of two parts: divergence constraint and vorticity constraint. The divergence constraint penalizes the convergence and divergence of the wind field, ensuring that it satisfies the near-surface incompressibility assumption. The expression for this divergence constraint is:

[0174]

[0175] Where u is the predicted U-wind component and v is the predicted V-wind component.

[0176] The vortex constraint requires that the downscaled vortex field maintains the large-scale vortex structure consistent with the true value. The expression for this vortex constraint is:

[0177]

[0178] Combination Figure 2As shown, during the training of the progressive multi-level downscaling network, the prepared historical dataset is divided into a training set (70%), a validation set (15%), and a test set (15%) according to time sequence. During training, the data flows through the complete processing chain shown in the attached figure: the input data undergoes preliminary spatiotemporal alignment and standardization through the data prefetching module; then it enters the core terrain-aware downscaling module (corresponding to the TA-RDN network), which achieves deep coupling between wind field and terrain features through its internal multi-scale terrain feature extractor and terrain-wind speed coupled attention mechanism; the coupled features are progressively upsampled by the progressive upsampling module through multi-level sub-pixel convolution to gradually improve spatial resolution; the entire forward propagation process is supervised by the physical constraint optimization module, which provides the network with gradient signals that combine data fidelity and physical consistency by calculating the composite loss function.

[0179] The network weights are updated using the Adam optimizer, with appropriate learning rates and batch sizes set, aiming to minimize the composite loss function. During training, model performance is continuously monitored on independent validation sets, and an early stopping strategy is employed to prevent overfitting. Finally, through this data-driven and physics-guided optimization mechanism, a progressive multi-level downscaling network with optimal parameters, stable performance, and physical reliability is obtained. This network is then used for subsequent high-resolution wind speed field inference and reconstruction, outputting a second spatial resolution wind speed field in practical applications.

[0180] Secondly, a progressive wind speed field depth downscaling device that integrates multi-source data and spatiotemporal perception, combined with... Figure 3 As shown, the device includes: an acquisition module 10, an extraction module 20, and a wind speed field reconstruction module.

[0181] The acquisition module 10 is used to acquire gridded wind speed data, station observation data, and DEM data at the first spatial resolution.

[0182] The extraction module 20 is used to extract terrain features from the DEM data to obtain multi-scale terrain features.

[0183] The wind speed field reconstruction module 30 is used to input gridded wind speed data, station observation data and multi-scale terrain features based on the first spatial resolution into a trained progressive multi-level downscaling network for iterative reconstruction and output a wind speed field with the second spatial resolution.

[0184] Among them, the progressive multi-level downscaling network is trained by minimizing a composite loss function that includes physical constraint loss, which is used to constrain the divergence and vorticity of the reconstructed wind field; the second spatial resolution is higher than the first spatial resolution.

[0185] Thirdly, embodiments of this application provide an electronic device, combined with Figure 4 As shown, the electronic device includes a memory 131 and a processor 130. The memory 131 stores a computer program, and the processor 130 runs the computer program to make the electronic device perform the above-described method.

[0186] Furthermore, combined Figure 4 The electronic device shown also includes a bus 132 and a communication interface 133, with the processor 130, the communication interface 133 and the memory 131 connected via the bus 132.

[0187] The memory 131 may include high-speed random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device. Communication between this system network element and at least one other network element is achieved through at least one communication interface 133 (which can be wired or wireless), such as the Internet, wide area network, local area network, metropolitan area network, etc. The bus 132 may be an ISA bus, PCI bus, or EISA bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 4 The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.

[0188] Processor 130 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of processor 130 or by instructions in software form. Processor 130 may be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it may also be 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 invention. The general-purpose processor may be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this invention can be directly manifested as execution by a hardware decoding processor, or execution by a combination of hardware and software modules in the decoding processor. The software module can reside in a readily available storage medium in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory 131. The processor 130 reads the information from memory 131 and, in conjunction with its hardware, completes the steps of the method described in the foregoing embodiments.

[0189] Fourthly, embodiments of this application provide a readable storage medium storing computer program instructions, which are read and executed by a processor to perform the above-described method.

[0190] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system and apparatus described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0191] Furthermore, in the description of the embodiments of the present invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in the present invention based on the specific circumstances.

[0192] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0193] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0194] Finally, it should be noted that the above embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A progressive wind speed field depth downscaling method integrating multi-source data and spatiotemporal sensing, characterized in that, include: Acquire gridded wind speed data, station observation data, and DEM data at the first spatial resolution; Topographic features are extracted from the DEM data to obtain multi-scale topographic features; Based on the gridded wind speed data at the first spatial resolution, the station observation data, and the multi-scale terrain features, the data are input into the trained progressive multi-level downscaling network for iterative reconstruction, and the wind speed field at the second spatial resolution is output. The progressive multi-level downscaling network is trained by minimizing a composite loss function that includes a physical constraint loss term, which is used to constrain the divergence of the reconstructed wind field in the horizontal direction and the vorticity in the vertical direction; the second spatial resolution is higher than the first spatial resolution. Each downscaling stage of the progressive multi-level downscaling network includes a terrain-aware residual dense network; the terrain-aware residual dense network includes a terrain-wind speed coupled attention module; the steps of inputting gridded wind speed data at the first spatial resolution, the station observation data, and the multi-scale terrain features into the trained progressive multi-level downscaling network for iterative reconstruction and outputting a wind speed field at the second spatial resolution include: The system receives gridded wind speed data at the first spatial resolution from multiple consecutive time intervals as input, and performs time-dimension weighted fusion of the feature maps corresponding to the multiple time intervals through a time attention module to obtain fused features. The adaptive fusion module performs weighted fusion of features from the gridded wind speed data and features from the station observation data to generate multi-source wind speed features for the current stage. The multi-source wind speed features and the multi-scale terrain features corresponding to the current stage are input into the terrain-aware residual dense network of this stage. The features are modulated through its terrain-wind speed coupled attention module, and the wind speed features enhanced with terrain information are output. Subpixel convolution upsampling is performed on the wind speed features enhanced with terrain information to generate the second spatial resolution wind speed field output at the current stage; The input to the current downscaling stage consists of the wind speed field output from the previous downscaling stage, intermediate features extracted from the network in the previous downscaling stage, and terrain features at the current downscaling stage resolution.

2. The method according to claim 1, characterized in that, The steps of extracting terrain features from the DEM data to obtain multi-scale terrain features include: Based on the DEM data, a multi-channel terrain feature map including elevation features, slope features, aspect features, curvature features, and roughness index features is calculated; The multi-channel terrain feature map is input into a multi-scale terrain feature extractor, and primary terrain features corresponding to different spatial scales are extracted through at least two parallel convolutional branches with receptive fields of different sizes. By using learnable weights, multiple primary terrain features are adaptively weighted and fused to generate a terrain embedding vector, which is then used as the multi-scale terrain feature.

3. The method according to claim 2, characterized in that, Based on the DEM data, the steps for calculating a multi-channel terrain feature map including elevation features, slope features, aspect features, curvature features, and roughness index features include: The elevation feature is obtained directly based on the grid point elevation values ​​in the DEM data; Based on the elevation value, the slope characteristics and the aspect characteristics are calculated through spatial difference; Based on the elevation value, the curvature feature is calculated using second-order differential. Based on the elevation value, the roughness index feature is obtained by calculating the standard deviation of the elevation values ​​within a predetermined local window.

4. The method according to claim 2, characterized in that, The parallel convolutional branches include a first branch with a kernel size of 3×3, a second branch with a kernel size of 5×5, and a third branch with a kernel size of 7×7.

5. The method according to claim 1, characterized in that, The composite loss function includes reconstruction loss, multi-scale supervision loss, and the physical constraint loss term; the physical constraint loss term includes divergence constraint loss and vorticity constraint loss calculated based on the two-dimensional spatial partial derivative of the predicted wind.

6. A progressive wind speed field depth downscaling device integrating multi-source data and spatiotemporal sensing, characterized in that, The device includes: The acquisition module is used to acquire gridded wind speed data, station observation data, and DEM data at the first spatial resolution. The extraction module is used to extract terrain features from the DEM data to obtain multi-scale terrain features; The wind speed field reconstruction module is used to input the gridded wind speed data with the first spatial resolution, the station observation data, and the multi-scale terrain features into the trained progressive multi-level downscaling network for iterative reconstruction and output the wind speed field with the second spatial resolution. The progressive multi-level downscaling network is trained by minimizing a composite loss function that includes a physical constraint loss term, which is used to constrain the divergence of the reconstructed wind field in the horizontal direction and the vorticity in the vertical direction; the second spatial resolution is higher than the first spatial resolution. Each downscaling stage of the progressive multi-level downscaling network includes a terrain-aware residual dense network; the terrain-aware residual dense network includes a terrain-wind speed coupled attention module; the steps of inputting gridded wind speed data at the first spatial resolution, the station observation data, and the multi-scale terrain features into the trained progressive multi-level downscaling network for iterative reconstruction and outputting a wind speed field at the second spatial resolution include: The system receives gridded wind speed data at the first spatial resolution from multiple consecutive time intervals as input, and performs time-dimension weighted fusion of the feature maps corresponding to the multiple time intervals through a time attention module to obtain fused features. The adaptive fusion module performs weighted fusion of features from the gridded wind speed data and features from the station observation data to generate multi-source wind speed features for the current stage. The multi-source wind speed features and the multi-scale terrain features corresponding to the current stage are input into the terrain-aware residual dense network of this stage. The features are modulated through its terrain-wind speed coupled attention module, and the wind speed features enhanced with terrain information are output. Subpixel convolution upsampling is performed on the wind speed features enhanced with terrain information to generate the second spatial resolution wind speed field output at the current stage; The input to the current downscaling stage consists of the wind speed field output from the previous downscaling stage, intermediate features extracted from the network in the previous downscaling stage, and terrain features at the current downscaling stage resolution.

7. An electronic device, characterized in that, The electronic device includes a memory and a processor, the memory storing a computer program and the processor running the computer program to cause the electronic device to perform the method of any one of claims 1 to 5.

8. A storage medium, characterized in that, The storage medium stores computer program instructions, which, when read and executed by a processor, perform the method described in any one of claims 1 to 5.

Images