A downscaling method for near-surface meteorological fields based on terrain-constrained Transformer model

Abstract

The present invention relates to the fields of meteorological data processing and geographic information technology, and in particular to a near-surface meteorological field downscaling method based on a terrain-constrained Transformer model. The method comprises the following steps: extracting multimodal fusion features from reanalyzed meteorological data and digital elevation model data, and then mapping to a target resolution using a terrain-constrained Transformer network to obtain initial downscaled meteorological data; performing latent variable modeling on the initial data through a variational autoencoder to characterize multi-scale uncertainty, and incorporating physical constraints such as energy balance and mass conservation into an iterative update mechanism based on meta-learning to generate core downscaled meteorological data; and the present invention achieves detailed capture and dynamic adaptation of local meteorology through self-supervised pre-training and online updating.

Description

Technical Field

[0001] The present invention relates to the fields of meteorological data processing and geographic information technology, and in particular to a near-surface meteorological field downscaling method based on a terrain-constrained Transformer model. Background Art

[0002] The near-surface meteorological environment is often influenced by complex topography, microscale climate effects, and regional atmospheric circulation. Therefore, relying solely on traditional data interpolation or simple global deep learning methods fails to capture the coupled mechanisms of local characteristics and dynamic changes. Existing technologies generally employ deep learning methods that incorporate physical mechanism constraints (Chinese invention patent, publication number: CN119107229A, title: A method and system for spatial downscaling of meteorological data). After downsampling and reanalyzing meteorological data, spatial downscaling is achieved through a series of convolutional networks or shallow feature extraction modules. However, due to the lack of detailed modeling of local topographic influences and the high coupling of local meteorological elements, the models are prone to problems such as reduced accuracy and blurred boundaries. This results in poor performance when simulating highly localized meteorological phenomena such as complex mountainous areas, coastal areas, and urban heat islands.

[0003] In addition, existing technologies often only make single or simplified treatments of physical constraints, making it difficult to take into account the superimposed effects of multiple physical laws. They lack adaptive correction mechanisms for highly nonlinear near-surface meteorological elements, resulting in the problem of accumulated deviations in prediction results in non-flat terrain areas. Summary of the Invention

[0004] To address the numerous issues with the aforementioned existing technologies, the present invention provides a near-surface meteorological field downscaling method based on a terrain-constrained Transformer model. By integrating the terrain-constrained Transformer architecture with multimodal features, the present invention obtains a high-dimensional representation of preprocessed meteorological and terrain data. A variational autoencoder is then used to capture data uncertainty and multiscale features, supplemented by meta-learning iterative correction to achieve physical consistency. This method significantly improves downscaling accuracy in complex terrain areas and exhibits excellent generalization and online update capabilities.

[0005] A near-surface meteorological field downscaling method based on a terrain-constrained Transformer model includes the following steps:

[0006] Collect reanalysis meteorological data and digital elevation model data of the corresponding area, perform spatial alignment and resampling, extract and fuse features after preprocessing, and generate fused multimodal feature data;

[0007] The fused multimodal graph feature data is mapped to the target downscaling resolution through a downscaling network to form initial downscaled meteorological data, and a variational autoencoder is used to perform latent variable modeling on the initial downscaled meteorological data to obtain a latent variable expression that describes data uncertainty and multi-scale characteristics. At the same time, combined with physical constraints, a feedback loop is established between the data and theoretical physical values ​​through an iterative update mechanism based on meta-learning, thereby generating core downscaled meteorological data;

[0008] The core downscaled meteorological data and the measured meteorological data are used as supervision samples, a hybrid loss function is constructed through self-supervised pre-training and supervised fine-tuning, and an online update mechanism is implemented to achieve the output of downscaled meteorological data.

[0009] Preferably, the reanalysis meteorological data is generated by the Global Climate Reanalysis System or the Regional Climate Reanalysis System, the digital elevation model data is acquired by aerial remote sensing or satellite remote sensing and collected in the same area, and the spatial alignment and resampling steps perform projection conversion and grid resampling through a geographic information system.

[0010] Preferably, the multimodal feature extraction and fusion step adopts a two-branch convolutional network containing a terrain-constrained Transformer structure, one branch extracts features from the preprocessed reanalysis meteorological data to generate embedded meteorological data, and the other branch extracts features from the preprocessed digital elevation model data to generate embedded terrain data, and uses the terrain weight factors generated by the embedded terrain data to construct a dynamic adjacency matrix, and then uses the graph convolutional network to perform multi-level propagation and fusion of node features, thereby forming fused multimodal graph feature data.

[0011] Preferably, the downscaling network adopts a multi-layer convolution structure combined with linear mapping to receive the fused multimodal graph feature data. Each layer of the convolution structure is set with a fixed convolution kernel size and step size, and batch normalization and activation calculation are performed after the output to map the fused multimodal graph feature data to the target downscaling resolution to form initial downscaled meteorological data.

[0012] Preferably, the variational autoencoder consists of an encoder and a decoder, the encoder converts the initial downscaled meteorological data into a representation vector describing the mean and variance of the latent variable, and the decoder generates a reconstructed output containing uncertainty features based on reparameterized sampling, and the encoding and decoding process retains the local gradient information and spatial distribution characteristics of the initial downscaled meteorological data.

[0013] Preferably, the iterative update mechanism based on meta-learning utilizes a lightweight residual network to process the difference between the output of the variational autoencoder and the theoretical physical value. The lightweight residual network regards the difference as input and generates correction data. The correction data is added to the output of the variational autoencoder to form a feedback loop, thereby updating the initial downscaled meteorological data under the action of physical constraints and obtaining the core downscaled meteorological data.

[0014] Preferably, the theoretical physical values ​​are calculated based on the energy balance and mass conservation formulas. The energy balance calculates the energy flux based on the input meteorological elements, and the mass conservation calculates the divergence based on the humidity and wind speed related elements. The two are combined to evaluate the deviations of the initial downscaled meteorological data in dynamics and thermodynamics.

[0015] Preferably, the core downscaled meteorological data and the measured meteorological data obtained by independent observation constitute the supervised samples, and a hybrid loss function is established by combining self-supervised pre-training with supervised fine-tuning. The hybrid loss function includes a reconstruction loss based on difference statistics and a constraint loss that reflects physical consistency, and the gradient is calculated by summing the two losses according to preset weights.

[0016] Preferably, self-supervised pre-training generates masked data samples by performing regional masking on the reanalyzed meteorological data, and reconstructs the masked data samples using a multimodal network containing a terrain-constrained Transformer structure; supervised fine-tuning corrects parameters by comparing the spatial distribution and meteorological element value gaps between the core downscaled meteorological data and the measured observation meteorological data, and both self-supervised pre-training and supervised fine-tuning adopt the hybrid loss function.

[0017] Preferably, the online update mechanism periodically calculates the output deviation of the core downscaled meteorological data based on the continuously received measured observation meteorological data. When the deviation exceeds the threshold, the supervised fine-tuning process is triggered to incrementally train the parameters of the network containing the terrain constrained Transformer structure and the variational autoencoder, so as to gradually correct the weights of the downscaling network and output the downscaled meteorological data within multiple update cycles.

[0018] Compared with the prior art, the advantages and beneficial effects of the present invention are:

[0019] This paper combines the terrain-constrained Transformer structure with graph convolution fusion technology to achieve an explicit modeling effect of the influence of local complex terrain;

[0020] The present invention obtains the latent variable distribution through variational autoencoders to characterize the uncertainty of meteorological data at multiple scales;

[0021] The present invention incorporates physical constraints such as energy balance and mass conservation through an iterative update mechanism based on meta-learning to achieve adaptive correction of the initial downscaled meteorological data;

[0022] This paper introduces self-supervised pre-training and online update mechanisms to enhance the generalization ability and real-time adaptability of the model under data scarcity or local anomalies.

[0023] Through the above-mentioned technical means, the present invention achieves the comprehensive effects of capturing local meteorological details and satisfying physical consistency under multimodal data fusion, characterizing uncertainty in multi-scale latent variable modeling, and gradually approaching real physical laws in cyclic iterative updates. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 Schematic diagram of the process of the present invention;

[0025] Figure 2 Schematic diagram of fusion multimodal graph feature extraction in the present invention;

[0026] Figure 3 Schematic diagram of supervised fine-tuning and online updating in the present invention. DETAILED DESCRIPTION

[0027] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are merely exemplary and are not intended to limit the scope of the present disclosure. In the following detailed description, for ease of explanation, many specific details are set forth to provide a comprehensive understanding of the embodiments of the present disclosure. However, it is apparent that one or more embodiments may also be implemented without these specific details. In addition, in the following description, descriptions of well-known structures and technologies are omitted to avoid unnecessary confusion of the concepts of the present disclosure.

[0028] The terms used herein are only for describing specific embodiments and are not intended to limit the present disclosure. The terms "comprise," "include," etc. used herein indicate the presence of the features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0029] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art unless otherwise defined. It should be noted that the terms used herein should be interpreted as having a meaning consistent with the context of this specification and should not be interpreted in an idealized or overly rigid manner.

[0030] like Figure 1 As shown in FIG, a near-surface meteorological field downscaling method based on a terrain-constrained Transformer model includes the following steps:

[0031] Collect reanalysis meteorological data and digital elevation model data of the corresponding area, perform spatial alignment and resampling, extract and fuse features after preprocessing, and generate fused multimodal feature data;

[0032] The key to achieving downscaling of the near-surface meteorological field in the present invention as a whole is to collect meteorological data from atmospheric reanalysis and digital elevation model data in the corresponding area, perform strict spatial alignment and resampling processing on the two, and rely on advanced feature extraction and multimodal data fusion technology after preprocessing to finally form fused multimodal graph feature data. Specifically, the reanalysis meteorological data first collected by the present invention include meteorological elements such as temperature, humidity, and wind speed. These data are usually generated by global or regional climate reanalysis systems, and the data format is mostly NetCDF or GRIB, and its spatial resolution is relatively low; at the same time, the digital elevation model data in the corresponding area is stored in GeoTIFF format, reflecting terrain undulations, slope, aspect, curvature and other features. The acquisition stage ensures that the two cover the same area, so that subsequent data processing can achieve accurate matching.

[0033] After data collection, the collected reanalysis meteorological data and digital elevation model data are spatially aligned and resampled using a geographic information system. This step involves coordinate system conversion, projection correction, and grid unification of the original data to ensure that the meteorological and terrain data are completely consistent in coordinate system, spatial extent, and resolution. Next, the aligned data are preprocessed separately. The reanalysis meteorological data undergoes noise suppression and information enhancement through Gaussian smoothing and local statistics (such as local mean, variance, and gradient calculations), while the digital elevation model data undergoes terrain indices calculated using edge detection operators and normalization. The resulting preprocessed data provides a stable and consistent foundation for subsequent feature extraction.

[0034] After data preprocessing, the present invention employs advanced feature extraction and fusion methods to extract high-dimensional features from the preprocessed data and establish associations between multimodal data. Specifically, the present invention utilizes a strategy for fusing multimodal graph feature data, inputting reanalysis meteorological data and digital elevation model data into their respective feature extraction modules and fusing their features to form a comprehensive information representation that reflects the spatial distribution of the meteorological field and the influence of terrain. This fusion process not only preserves the spatiotemporal dynamics of the meteorological data but also incorporates the local environmental constraints embodied in the terrain data, providing rich input for subsequent downscaling models.

[0035] In general, the present invention realizes a closed loop of the entire process from data collection, preprocessing to feature extraction and fusion, in which each step ensures data consistency through precise spatial matching and data normalization, and uses advanced graphical feature fusion technology to build a solid input foundation for the downscaling model. Ultimately, the fused multimodal graph feature data becomes the core intermediate product in the entire downscaling process. It not only contains the macroscopic atmospheric characteristics of the reanalysis meteorological data, but also integrates the local detail information reflected by the terrain data, providing the necessary high-dimensional feature support for the subsequent use of the terrain-constrained Transformer structure and the variational autoencoder and meta-learning iterative update mechanism. The overall design of the present invention not only ensures the rigor of data processing, but also realizes the efficient use of multiple data sources in downscaling processing through cross-modal fusion, thereby laying the foundation for the refined expression of the near-surface meteorological field.

[0036] During implementation, the present invention first relies on advanced data acquisition equipment and sensors to obtain reanalysis meteorological data and digital elevation model data. These two types of data undergo strict regional division and time matching before acquisition to ensure that the data have high consistency and comparability. Subsequently, the spatial alignment process uses mature geographic information system algorithms, such as nearest neighbor interpolation and bilinear interpolation, to eliminate systematic errors between acquisition equipment and sensors; at the same time, when resampling the data, a fixed parameter resampling strategy is used to ensure the uniformity of the grid data. In the preprocessing stage, the reanalysis meteorological data is smoothed by a Gaussian filter before local statistical calculations are performed. Specific parameters such as the standard deviation of the filter and the window size are determined in experiments to ensure numerical stability; the digital elevation model data uses filtering methods such as Sobel and Laplacian to calculate terrain gradient information, and the data scale is standardized in combination with normalization methods. Ultimately, these preprocessing steps provide high-quality raw input for subsequent feature extraction.

[0037] Through the above steps, the overall design of this invention integrates multimodal information from different data sources into a unified high-dimensional representation. Its core lies in data consistency and multimodal fusion strategies, ensuring sufficient information content at the input level of the downscaling model. This overall design not only conforms to the development trend of existing technologies, but also lays a solid data foundation for the further introduction of advanced Transformer and meta-learning iterative update mechanisms.

[0038] Preferably, the reanalysis meteorological data is generated by the Global Climate Reanalysis System or the Regional Climate Reanalysis System, the digital elevation model data is acquired by aerial remote sensing or satellite remote sensing and collected in the same area, and the spatial alignment and resampling steps perform projection conversion and grid resampling through a geographic information system.

[0039] In the present invention, the preferred implementation further stipulates the specific requirements for data collection and spatial alignment. In the preferred embodiment, the reanalysis meteorological data is explicitly generated by the global climate reanalysis system or the regional climate reanalysis system. These systems use multi-source observation data and numerical model assimilation technology to generate meteorological data with high temporal and spatial consistency and historical continuity. In order to ensure the accuracy and availability of the data, the present invention requires that the collected reanalysis meteorological data must cover key meteorological elements such as temperature, humidity and wind speed, and be stored in a standardized data format (such as NetCDF or GRIB) for subsequent processing. When collecting data, it is also particularly required that the collected digital elevation model data be obtained from aerial remote sensing or satellite remote sensing. These data are provided by specialized remote sensing equipment, have a fine resolution, and can truly reflect the details of the terrain in the target area. The digital elevation model data is stored in GeoTIFF format, and its accuracy and grid density meet the requirements of subsequent graphics processing and feature extraction in terms of technical indicators.

[0040] In a preferred embodiment, data collection requires that reanalysis meteorological data and digital elevation model data be collected simultaneously in the same area to ensure that the two are strictly spatially corresponding. For example, when studying a specific area, the collected reanalysis meteorological data and digital elevation model data should be consistent in geographic coordinates, regional scope and collection time, so as to ensure that no additional errors are introduced in the subsequent spatial alignment process. The spatial alignment and resampling steps are automatically processed by the geographic information system, and the two types of data are converted to the same coordinate system through projection conversion and grid resampling. The projection conversion algorithms used in the specific implementation include the commonly used Mercator projection, UTM projection, etc., while the grid resampling method uses bilinear interpolation or cubic convolution interpolation technology. In order to ensure the numerical continuity and smoothness of the data, the interpolation algorithm parameters used (such as interpolation window size, weight distribution) are determined in the preliminary experiment and strictly implemented during the implementation process.

[0041] In addition, the preferred embodiment emphasizes that during the spatial alignment and resampling process, standardized processes are used for coordinate system conversion, spatial range interception, and resolution adjustment of the reanalysis meteorological data and the digital elevation model data to ensure that the processing results have high consistency and reproducibility. For example, if the original resolution of the collected reanalysis meteorological data is 0.25°×0.25°, and the resolution of the digital elevation model data is 10 meters, when performing projection conversion and resampling through GIS software, the bounding box of the target area is first determined and a unified grid is generated at a fixed resolution; on this basis, an interpolation algorithm is used to convert the digital elevation model data into a grid format that matches the reanalysis meteorological data. Each step in this process is accompanied by detailed parameter setting records and calibration methods to ensure that the final output data is strictly aligned in terms of spatial position, resolution, and numerical range.

[0042] In the specific operation, a standard software platform (such as the open source GIS software QGIS or the commercial GIS software ArcGIS) is used for spatial alignment and resampling. The operation steps include: first loading the original reanalysis meteorological data and digital elevation model data, then selecting the appropriate projection conversion tool to perform coordinate conversion, then setting the target grid parameters, and finally performing the resampling operation to generate data in a unified format. During the resampling process, the interpolation formula used is the bilinear interpolation formula:

[0043]

[0044] in, , and , are the coordinates of adjacent grid points, represents the interpolated data value. The above formula ensures that the resampled data is numerically smooth and continuous. Through strict parameter settings and standard operations, this preferred embodiment can provide high-quality and consistent input data in the data acquisition and preprocessing stages, providing a strong guarantee for subsequent processing.

[0045] Preferably, the multimodal feature extraction and fusion step adopts a two-branch convolutional network containing a terrain-constrained Transformer structure, one branch extracts features from the preprocessed reanalysis meteorological data to generate embedded meteorological data, and the other branch extracts features from the preprocessed digital elevation model data to generate embedded terrain data, and uses the terrain weight factors generated by the embedded terrain data to construct a dynamic adjacency matrix, and then uses the graph convolutional network to perform multi-level propagation and fusion of node features, thereby forming fused multimodal graph feature data.

[0046] like Figure 2 As shown, the multimodal feature extraction and fusion steps use a dual-branch convolutional network with a terrain-constrained Transformer structure to fully utilize the complementary information of reanalysis meteorological data and digital elevation model data. Specifically, the preprocessed reanalysis meteorological data is first processed using a dedicated convolutional network channel. This channel maps the original input to a high-dimensional feature space through continuous convolution layers, activation functions, and batch normalization modules to generate embedded meteorological data. The convolution kernel size, step size, and activation function (such as ReLU) in this process are determined based on experimental data to ensure the effectiveness and stability of feature extraction. For example, in one embodiment, the convolution kernel size used is 3×3, the step size is fixed at 1, and after three layers of convolution, the number of output channels increases step by step to ensure that spatial features from low-level to high-level are captured.

[0047] Simultaneously, another dedicated convolutional network channel extracts terrain features from the preprocessed digital elevation model data to generate embedded terrain data. This channel also employs multi-layer convolution operations, but incorporates a feature extraction module tailored to the specific characteristics of terrain data, highlighting terrain features such as local slope, aspect, and curvature. During this process, the embedded terrain data undergoes a nonlinear transformation to generate terrain weight factors. These factors reflect the sensitivity of each region to terrain change and are normalized using a sigmoid function to maintain a value between 0 and 1.

[0048] Furthermore, in the preferred embodiment, a dynamic adjacency matrix is ​​constructed using terrain weight factors generated by embedding terrain data. The construction of the dynamic adjacency matrix takes into account the terrain differences at each spatial location, and its calculation method can be expressed as:

[0049] in, Representation node With node The weight between and Represent the terrain weight factor vectors at the corresponding positions, is a smoothing parameter. This formula ensures that the weight is higher between similar areas and decays rapidly between areas with large differences. Through the present invention, the dynamic adjacency matrix can truly reflect the impact of terrain changes on the correlation of spatial data.

[0050] After constructing the dynamic adjacency matrix, the preferred embodiment further uses a graph convolutional network to perform multi-level propagation and fusion of node features. The graph convolutional network extracts global and local fusion features by performing a weighted sum operation on each node and its neighboring nodes. Its propagation formula is usually:

[0051] in, Indicates the The node feature matrix of the layer, is the matrix formed by adding the dynamic adjacency matrix to the self-connection matrix, for The degree matrix of For the The weight matrix of the layer, is a nonlinear activation function. Through the processing of continuous multi-layer graph convolutional networks, the feature information from reanalysis meteorological data and digital elevation model data can be fully integrated, ultimately forming fused multimodal graph feature data that reflects regional spatiotemporal characteristics and terrain constraints.

[0052] The entire multimodal feature extraction and fusion process fully utilizes the advantages of the Transformer architecture with terrain constraints. This architecture combines the characteristics of traditional convolutional neural networks and the Transformer self-attention mechanism. Through a dual-branch design, it processes data of two different modalities separately, and realizes efficient information interaction through a dynamic adjacency matrix and graph convolutional network, ensuring the capture of local details while establishing global dependencies. This method is highly operational in practice, and related parameters (such as the number of convolutional layers, convolution kernel size, activation function type, adjacency matrix smoothing parameter) are very simple. etc.) can be debugged and optimized according to the characteristics of the target area and the amount of data, so as to ensure that the final fusion of multimodal map feature data can truly reflect the complex relationship between the meteorological field and terrain in the region.

[0053] In summary, by employing a dual-branch convolutional network with a terrain-constrained Transformer structure and combining it with a graph convolutional network for feature fusion, the preferred embodiment extracts highly representative high-dimensional features from preprocessed reanalysis meteorological data and digital elevation model data. Based on this, a dynamic adjacency matrix is ​​constructed, and multi-level feature propagation and fusion are achieved through a graph convolutional network. This process provides rich input data for the subsequent downscaling stage, ensuring that the entire downscaling method achieves a deep fusion of meteorological and terrain data while maintaining regional spatial consistency, providing a solid data support and theoretical foundation for the downscaling of near-surface meteorological fields.

[0054] The fused multimodal graph feature data is mapped to the target downscaling resolution through a downscaling network to form initial downscaled meteorological data, and a variational autoencoder is used to perform latent variable modeling on the initial downscaled meteorological data to obtain a latent variable expression that describes data uncertainty and multi-scale characteristics. At the same time, combined with physical constraints, a feedback loop is established between the data and theoretical physical values ​​through an iterative update mechanism based on meta-learning, thereby generating core downscaled meteorological data;

[0055] The overall idea of ​​the present invention to achieve downscaling of the near-surface meteorological field is to map the fused multimodal graph feature data obtained from the preprocessing stage to the target downscaling resolution through a downscaling network to obtain initial downscaled meteorological data; then use a variational autoencoder to perform latent variable modeling on the initial downscaled meteorological data, so as to obtain a latent variable expression that can describe data uncertainty and multi-scale characteristics; finally, combined with theoretical physics constraints, through an iterative update mechanism based on meta-learning, a feedback loop is established between the initial downscaled meteorological data and the theoretical physical values, and the downscaled data is gradually updated and corrected to generate core downscaled meteorological data.

[0056] Overall, this method utilizes key technologies such as data preprocessing, feature extraction, multimodal information fusion, downscaling network mapping, variational autoencoder latent variable modeling, and meta-learning iterative updates to form a complete data processing loop. This ensures consistent data processing across space, time, and modality, while also introducing physical constraints to ensure that the downscaled data conforms to the requirements of dynamics and thermodynamics theory. In this scheme, fused multimodal graph feature data is generated by fusing preprocessed reanalysis meteorological data with digital elevation model data using advanced graph convolutional network technology. This data contains both macroscopic background information of the meteorological field and local detailed features of the terrain, providing a high-dimensional, rich feature representation for subsequent downscaling operations. Subsequently, this fused feature data is mapped through the downscaling network to achieve data resolution conversion and generate initial downscaled meteorological data. This stage primarily involves convolution and linear mapping operations, and its design parameters have been rigorously debugged in practical applications.

[0057] On this basis, a variational autoencoder is used for latent variable modeling. The encoder converts the initial downscaled meteorological data into a low-dimensional representation of the latent variable mean and variance. The decoder, combined with reparameterized sampling, then obtains a latent variable representation that reflects the data's uncertainty and multi-scale characteristics. This process ensures that the inherent statistical properties of the downscaled data are preserved. Finally, the scheme incorporates theoretical physical constraints (such as energy balance and mass conservation) and utilizes an iterative update mechanism based on meta-learning. The difference between the initial downscaled meteorological data and the theoretical calculated value is used as feedback. Corrected data is generated through a residual network, forming a closed-loop feedback correction system. After multiple iterative updates, the data is more consistent with physical principles, and the generated core downscaled meteorological data has higher reliability and precision. The overall method has a compact structure and clear hierarchy. The seamless integration of modules ensures that information is fully transferred and utilized across different scales and modes, thereby achieving effective downscaling of the near-surface meteorological field.

[0058] Preferably, the downscaling network adopts a multi-layer convolution structure combined with linear mapping to receive the fused multimodal graph feature data. Each layer of the convolution structure is set with a fixed convolution kernel size and step size, and batch normalization and activation calculation are performed after the output to map the fused multimodal graph feature data to the target downscaling resolution to form initial downscaled meteorological data.

[0059] In a preferred embodiment, the downscaling network utilizes a multi-layer convolutional structure combined with linear mapping to receive the aforementioned fused multimodal graph feature data and map it to the target downscaled resolution. Specifically, the design of the downscaling network primarily considers the use of a multi-layer convolutional neural network to fully extract local information from the input features. Each layer utilizes a fixed-size convolution kernel (e.g., 3×3 or 5×5) and a fixed step size (e.g., 1 or 2) to ensure parameter stability of the convolution operation and uniformity of the network structure. After the convolution operation, the network undergoes batch normalization at each layer to eliminate data distribution bias and improve convergence speed during model training. The convolution results are then processed using nonlinear activation functions (e.g., ReLU or Leaky ReLU) to enhance the network's expressive power.

[0060] The linear mapping process involves reducing or mapping the multidimensional convolution output to a predetermined target resolution using fully connected layers or linear transformations after multiple convolution operations. This process ensures uniform spatial resolution of the output data while preserving the information transfer of the original multimodal features. To achieve numerical smoothing and feature preservation during the downscaling process, a residual connection mechanism is incorporated into the network design. This mitigates the vanishing gradient problem in deep network structures and ensures the effective transfer of low-level features. For example, in a specific implementation, the following steps can be used: the input fused multimodal graph feature data is first passed through a series of convolutional layers to gradually extract features, followed by dimensionality reduction through a fully connected layer. The final output of the initial downscaled meteorological data maintains continuity and spatial consistency within the target area. During the operation, the weights, step sizes, activation functions, and normalization parameters of each convolution kernel layer are calibrated using pre-experimental data to ensure stable and highly reproducible output of the entire downscaling network. The design of the downscaling network follows the architecture of a conventional convolutional neural network, while also customizing its parameters based on the data characteristics of the present invention, ensuring that the data resolution can be effectively reduced during the mapping process without losing key meteorological and topographic feature information, thereby providing a reliable initial data foundation for further processing of subsequent data.

[0061] Preferably, the variational autoencoder consists of an encoder and a decoder, the encoder converts the initial downscaled meteorological data into a representation vector describing the mean and variance of the latent variable, and the decoder generates a reconstructed output containing uncertainty features based on reparameterized sampling, and the encoding and decoding process retains the local gradient information and spatial distribution characteristics of the initial downscaled meteorological data.

[0062] The variational autoencoder consists of an encoder and a decoder. Its main function is to perform latent variable modeling on the initial downscaled meteorological data to obtain a latent variable expression that can reflect the uncertainty and multi-scale characteristics of the data. Specifically, the encoder part uses a multi-layer fully connected network or a convolutional neural network structure to map the initial downscaled meteorological data into a low-dimensional latent variable vector, which also contains the mean and variance information of the latent variable. In the design of the encoder, each layer is equipped with an activation function (such as ReLU) and a batch normalization layer to ensure the stability of the output feature distribution, and fixed parameters are used in actual operations to keep the statistical distribution of the data consistent. Obtain the mean vector of the latent variable and variance vector Then, using the reparameterization technique, the latent variables are sampled from the normal distribution , its mathematical expression is:

[0063] in, is the noise term sampled from a standard normal distribution, represents the mean of the latent variable, is the standard deviation, and this formula ensures that the gradient can be transferred during the back propagation process. The decoder also uses a multi-layer convolution or fully connected structure to convert the sampled latent variables Mapping back to the data space generates a reconstructed output containing uncertainty features. During the encoding and decoding process, the network structure is designed to preserve local gradient information and spatial distribution characteristics, ensuring that the reconstructed data can reflect the subtle changes and multi-scale correlations of the initial downscaled meteorological data. During operation, the weights of the encoder and decoder are optimized through joint training, with the goal of minimizing the sum of the reconstruction error and the KL divergence term. The hybrid loss function is usually expressed as:

[0064] in, is the reconstruction error, is the KL divergence, To adjust the parameters, is the posterior distribution generated by the encoder, is the prior distribution. Through this method, the variational autoencoder can capture the inherent uncertainty of the data while retaining multi-scale feature information. The resulting latent variable representation provides a solid data representation foundation for subsequent iterative updates. In practical applications, the number of layers and nodes in the encoder and decoder structures, as well as the choice of activation function, have been rigorously debugged in experiments to ensure the high stability and representativeness of the latent variable representation. This implementation provides an effective statistical basis and uncertainty description for the generation of core downscaled meteorological data.

[0065] Preferably, the iterative update mechanism based on meta-learning utilizes a lightweight residual network to process the difference between the output of the variational autoencoder and the theoretical physical value. The lightweight residual network regards the difference as input and generates correction data. The correction data is added to the output of the variational autoencoder to form a feedback loop, thereby updating the initial downscaled meteorological data under the action of physical constraints and obtaining the core downscaled meteorological data.

[0066] The meta-learning-based iterative update mechanism aims to use a lightweight residual network to correct the difference between the output of the variational autoencoder and the theoretical physical value, thereby gradually updating the initial downscaled meteorological data under physical constraints and generating core downscaled meteorological data. In specific operations, the physical constraint error is first calculated based on the difference between the initial downscaled meteorological data and the theoretical physical value (calculated by the energy balance and mass conservation formulas). The calculation of the theoretical physical value is based on the input meteorological factors, where the energy balance is obtained by measuring thermal radiation and sensible heat flux, while the mass conservation is obtained by divergence calculation based on factors such as humidity and wind speed. The calculation formula can be, for example:

[0067] in, represents the physical quantity value corresponding to the initial downscaled meteorological data, represents the theoretical physical value, represents the absolute difference between the two. This difference is then fed as input into a lightweight residual network. The residual network architecture uses a fixed-parameter design and consists of several convolutional layers, each of which undergoes batch normalization and nonlinear activation. The output, corrected data, represents a correction to the initial downscaled meteorological data. This corrected data is added to the output of the variational autoencoder, forming a feedback loop that achieves adaptive correction between the data and the theoretical physical value. This feedback loop is implemented through iterative training, with each iteration updating the initial downscaled meteorological data to gradually approach the theoretical physical requirements.

[0068] For example, in one embodiment, by setting the number of iterations to N, each iteration calculates the physical error and generates correction data through a residual network until the error drops to a preset threshold. This meta-learning iterative update mechanism fully leverages the efficient computing power and adaptive characteristics of lightweight residual networks, enabling continuous refinement of data updates under physical constraints. The resulting core downscaled meteorological data is more consistent with theoretical requirements in terms of dynamics and thermodynamics. During operation, various parameters (such as convolution kernel size, number of residual network layers, and preset thresholds) are determined based on multiple sets of experimental data, ensuring the iterative update mechanism is highly stable and reliable, and can adapt to the complexity and changing trends of diverse meteorological data.

[0069] Preferably, the theoretical physical values ​​are calculated based on the energy balance and mass conservation formulas. The energy balance calculates the energy flux based on the input meteorological elements, and the mass conservation calculates the divergence based on the humidity and wind speed related elements. The two are combined to evaluate the deviations of the initial downscaled meteorological data in dynamics and thermodynamics.

[0070] The calculation of theoretical physical values ​​is based on two basic principles: energy balance and mass conservation. The energy balance part uses input meteorological factors such as temperature, radiation flux, and wind speed to establish an energy flux calculation formula to obtain an estimated value of energy transfer within the region. Specifically, the energy flux calculation can use the following standard formula:

[0071] in, represents the regional average energy flux, represents the air density, represents the specific heat capacity, represents the temperature of each grid point, is the reference temperature, is the number of grid points. This formula calculates the energy transfer value by calculating the difference between each grid point and the reference temperature, thereby obtaining an estimated energy balance result. The mass conservation part is obtained by calculating the divergence of the airflow based on the factors related to humidity and wind speed. Its basic calculation formula is:

[0072] in, represents the mass conservation deviation, represents the wind speed vector field, which describes the distribution of gas mass flow within a region. In practical applications, the theoretical physical value is a combination of the aforementioned energy flux and mass divergence data according to preset weights, forming an assessment criterion for the physical consistency of the initial downscaled meteorological data. Specifically, meteorological data collected by sensors provides numerical information on temperature, humidity, and wind speed. The theoretical physical values ​​are then calculated using a numerical model. The corresponding physical quantities in the initial downscaled meteorological data for the same region are compared, and the difference between the two is calculated. This difference serves as input for the subsequent meta-learning iterative update mechanism, ensuring that the feedback correction process strictly adheres to physical principles. This calculation process introduces explicit physical constraints during data processing, ensuring that the downscaling results are consistent with both statistical laws and basic kinetic and thermodynamic theories. In practice, the parameters in each calculation formula (such as air density, specific heat capacity, reference temperature, and weight coefficients) are predetermined based on regional meteorological characteristics and empirical data and used as fixed parameters in the downscaling model, ensuring the accuracy and reproducibility of the theoretical physical value calculations. This theoretical physical value calculation implementation method provides a physical basis for the entire downscaling process and is an important link in ensuring that the downscaled meteorological data are reasonable in terms of numerical and physical performance.

[0073] like Figure 2 As shown, the core downscaled meteorological data and the measured meteorological data are used as supervision samples, a hybrid loss function is constructed through self-supervised pre-training and supervised fine-tuning, and an online update mechanism is implemented to achieve the output of downscaled meteorological data.

[0074] This embodiment involves using core downscaled meteorological data and measured observational meteorological data to form a supervised sample, then constructing a hybrid loss function through a combination of self-supervised pre-training and supervised fine-tuning, and implementing an online update mechanism to achieve the overall process of outputting downscaled meteorological data. The overall solution first obtains core downscaled meteorological data through a pre-processing module. This data is generated through multimodal feature fusion, a downscaling network, a variational autoencoder, and an iterative update mechanism based on meta-learning. In terms of spatial resolution and data representation, it effectively preserves the macroscopic atmospheric characteristics of the reanalysis meteorological data and the terrain constraints provided by the digital elevation model data.

[0075] In order to further improve the consistency between the downscaled meteorological data and the measured observation data in terms of spatial distribution and meteorological element values, the present invention forms a supervision sample with the core downscaled meteorological data and the measured observation meteorological data obtained through independent observation equipment, and uses this combined data to carry out targeted training and fine-tuning of the deep network. Through self-supervised pre-training, this method performs regional masking on the reanalysis meteorological data, generates masked data samples, and uses a multimodal network containing a terrain-constrained Transformer structure to reconstruct and train the masked samples, so that the model can learn the spatiotemporal correlation and inter-modal interaction within the data in an unsupervised state. Then, in the supervised fine-tuning stage, the core downscaled meteorological data and the measured observation meteorological data are compared in terms of spatial distribution and the values ​​of various meteorological elements, the deviation between the two is quantified and the deviation is fed back into the network parameters, so as to realize the gradual correction of the model in a supervised environment.

[0076] The entire process involves constructing a hybrid loss function consisting of a reconstruction loss based on difference statistics and a constraint loss reflecting physical consistency. The gradients of these loss terms are summed according to preset weights to guide model parameter updates. Finally, through an online update mechanism, the system periodically calculates the core downscaled meteorological data output by the model based on the continuous receipt of measured meteorological data. When the output deviation exceeds a preset threshold, incremental training is triggered, gradually correcting the model parameters to ensure that the downscaled meteorological data output always maintains a high degree of consistency with the measured observations. This overall approach not only achieves innovations in data fusion, downscaling mapping, and latent variable modeling, but also establishes a complete feedback loop in supervised learning and online adaptive updating, ensuring that the output downscaled meteorological data has good continuity and physical consistency in time and space, thus providing reliable technical support for the downscaling of near-surface meteorological fields.

[0077] Preferably, the core downscaled meteorological data and the measured meteorological data obtained by independent observation constitute the supervised samples, and a hybrid loss function is established by combining self-supervised pre-training with supervised fine-tuning. The hybrid loss function includes a reconstruction loss based on difference statistics and a constraint loss that reflects physical consistency, and the gradient is calculated by summing the two losses according to preset weights.

[0078] The core downscaled meteorological data and the measured meteorological data obtained by independent observation constitute the supervision samples. The core downscaled meteorological data described here are jointly generated by the aforementioned downscaling network, variational autoencoder and meta-learning iterative update mechanism, while the measured meteorological data are collected in real time through ground automatic weather stations, radar or satellite observation systems, and are preprocessed to obtain stable observation values. The composition of the supervision samples requires that the two data sources have strict correspondence in time, space and data expression to ensure the effective transmission of supervision information. To achieve this goal, this embodiment establishes a hybrid loss function, which is composed of a reconstruction loss based on data difference statistics and a constraint loss that reflects physical consistency. The reconstruction loss is intended to measure the deviation in numerical value and spatial distribution between the core downscaled meteorological data and the measured meteorological data. Specifically, the mean square error (MSE) can be used as a metric, and its formula is:

[0079]

[0080] in, Indicates that the core downscaled meteorological data is The values ​​at the grid points, Represents the value of the observed meteorological data at the corresponding location, Represents the total number of grid points. Constraint loss reflects the gap between the thermodynamic and kinetic values ​​of the initial downscaled meteorological data and the theoretical physical values ​​by introducing physical consistency indicators, such as energy balance and mass conservation constraints. Its calculation method is based on the corresponding physical formulas, for example, the formula:

[0081]

[0082] in, The core downscaled meteorological data is The estimated energy flux in the region, is the theoretical value calculated based on the energy balance formula, is the number of regional divisions. The final hybrid loss function is a weighted summation of the above two losses using a preset weight coefficient:

[0083]

[0084] in, and These are weight parameters preset based on experimental data, used to balance the effects of data reconstruction and physical constraints. During implementation, the system optimizes the hybrid loss function using gradient descent to guide network parameter updates, enabling the downscaling model to effectively reduce the gap between the core downscaled meteorological data and the measured observational data during training. This process is highly operational, with parameter settings and specific loss function construction calibrated based on experimental data. This ensures the stability and consistency of the gradient direction during supervised training, enabling efficient and stable network training and model updates.

[0085] Preferably, Figure 3 As shown in the figure, self-supervised pre-training generates masked data samples by performing regional masking on the reanalyzed meteorological data, and reconstructs the masked data samples using a multimodal network containing a terrain-constrained Transformer structure; supervised fine-tuning corrects parameters by comparing the spatial distribution and meteorological element value gaps between the core downscaled meteorological data and the measured observation meteorological data. Both self-supervised pre-training and supervised fine-tuning use the aforementioned hybrid loss function.

[0086] The self-supervised pre-training component generates masked data samples by applying regional masking to the reanalysis meteorological data. This method creates missing data samples by masking certain regions or data channels in a fixed manner, forcing the model to learn the inherent spatiotemporal correlations in the data. Specifically, regional masking uses rectangular or random masks. The masking ratio is determined in pre-experiments to a certain range (e.g., 20% to 30%) to ensure that the generated masked samples adequately simulate missing data without making the model overly reliant on the masking structure. The masked data samples are then reconstructed using a multimodal network with a terrain-constrained Transformer architecture. This network combines multimodal information from the reanalysis meteorological data and the digital elevation model data, and utilizes a self-attention mechanism to enhance the interaction between local and global information, enabling the model to capture the complex spatiotemporal patterns inherent in the data in an unsupervised manner. During the reconstruction training process, the network uses the same loss calculation as the hybrid loss function, simultaneously considering both reconstruction error and physical constraint loss to update network parameters, achieving the self-supervised pre-training goal.

[0087] During the supervised fine-tuning phase, the preferred implementation performs parameter correction by comparing the spatial distribution and numerical values ​​of meteorological elements in the core downscaled meteorological data with independently collected field observations. Specifically, this process compares the core downscaled meteorological data with the field observations point by point on the same spatial grid and calculates a difference matrix. This difference matrix is ​​then input as feedback into the deep network, and the network parameters are updated based on the gradient calculated from the hybrid loss function. This supervised fine-tuning process relies on both a precisely constructed hybrid loss function and continuous observation data as an external supervision signal, allowing the downscaling model to continuously converge to an optimal state during training. In practice, the supervised fine-tuning process uses a gradient descent method (such as the Adam or AdamW optimization algorithm) to iteratively update the network parameters until the reconstruction error and physical constraint loss between the model output and the field observations fall to a preset threshold. The entire self-supervised pre-training and supervised fine-tuning process adopts a unified hybrid loss function design to ensure the consistency of the model's training objectives in both self-supervised and supervised states, and by adjusting the preset weight parameters, the two parts of training are made to reach the optimal balance, thereby effectively improving the output quality and physical consistency of the downscaled meteorological data.

[0088] Preferably, the online update mechanism periodically calculates the output deviation of the core downscaled meteorological data based on the continuously received measured observation meteorological data. When the deviation exceeds the threshold, the supervised fine-tuning process is triggered to incrementally train the parameters of the network containing the terrain constrained Transformer structure and the variational autoencoder, so as to gradually correct the weights of the downscaling network and output the downscaled meteorological data within multiple update cycles.

[0089] By periodically calculating the deviation between the measured meteorological data and the core downscaled meteorological data, and triggering incremental training when the deviation exceeds a preset threshold, the model parameters are dynamically updated, thereby achieving adaptive adjustment of the downscaling model during the continuous data reception process. In specific operations, the system continuously collects measured meteorological data and compares it with the core downscaled meteorological data in real time on the same spatial grid. The output deviation is calculated, which can be measured using the root mean square error or mean absolute error. For example, the system can be set to perform a statistical calculation of the model output data every 24 hours and compare the calculated error value with a pre-set threshold.

[0090] When the monitored error value exceeds the set threshold, the system automatically starts the online update module, which performs incremental training on the parameters of the multimodal network and variational autoencoder containing the terrain-constrained Transformer structure. During the incremental training process, the online update module uses small batches of data as training samples and fine-tunes the network parameters through the gradient descent method. The updated model can better adapt to the latest collected measured meteorological data. Specifically, the online update module first constructs a supervision sample based on the latest observation data, calculates the new gradient through the hybrid loss function, and then uses the gradient to update the network parameters; the update process adopts a learning rate decay strategy to avoid overfitting and ensure the stability of the update process.

[0091] Through multi-cycle online updates, the model parameters gradually approach the optimal value within each update cycle, and the output downscaled meteorological data can continuously maintain consistency with the measured meteorological data. For example, in a certain embodiment, the system sets the update cycle to 24 hours, and calculates the root mean square error between the downscaled meteorological data and the measured meteorological data within each cycle. When the error exceeds the set allowable range, the online incremental training module is automatically called to perform parameter correction. This online update mechanism not only realizes the real-time adaptive adjustment of model parameters, but also builds a data feedback loop, so that the downscale model can always output meteorological data that meets physical constraints and actual observation conditions during long-term operation. In actual operation, the implementation of the online update mechanism relies on a distributed computing platform and a real-time data acquisition system to ensure that the system can respond to data changes in a relatively short time, and realize continuous optimization of model parameters through an automated training process, thereby providing long-term and stable technical support for the downscaling of near-surface meteorological fields.

[0092] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware.

[0093] The above are merely embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various modifications and variations. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application should be included within the scope of the claims of the present application.

Claims

1. A near-surface meteorological field downscaling method based on a terrain-constrained Transformer model, characterized in that: The following steps are involved: Collect reanalysis meteorological data and digital elevation model data of the corresponding area, perform spatial alignment and resampling, extract and fuse features after preprocessing, and generate fused multimodal feature data; The feature extraction and fusion step uses a two-branch convolutional network with a terrain-constrained Transformer structure. One branch extracts features from the preprocessed reanalysis meteorological data to generate embedded meteorological data, and the other branch extracts features from the preprocessed digital elevation model data to generate embedded terrain data. The terrain weight factors generated by the embedded terrain data are used to construct a dynamic adjacency matrix. The node features are then propagated and fused at multiple levels through a graph convolutional network to form fused multimodal graph feature data. The fused multimodal graph feature data is mapped to the target downscaling resolution through a downscaling network to form initial downscaled meteorological data, and a variational autoencoder is used to perform latent variable modeling on the initial downscaled meteorological data to obtain a latent variable expression that describes data uncertainty and multi-scale characteristics. At the same time, combined with physical constraints, a feedback loop is established between the data and theoretical physical values ​​through an iterative update mechanism based on meta-learning, thereby generating core downscaled meteorological data; Using the core downscaled meteorological data and the measured meteorological data as supervision samples, a hybrid loss function is constructed through self-supervised pre-training and supervised fine-tuning, and an online update mechanism is implemented to achieve the output of the downscaled meteorological data; The core downscaled meteorological data and the measured meteorological data obtained by independent observations constitute the supervised samples. A hybrid loss function is established by combining self-supervised pre-training with supervised fine-tuning. The hybrid loss function includes a reconstruction loss based on difference statistics and a constraint loss that reflects physical consistency. The gradient is calculated by summing the two losses according to preset weights. Self-supervised pre-training generates masked data samples by performing regional masking on the reanalyzed meteorological data, and reconstructs and trains the masked data samples using a multimodal network containing a terrain-constrained Transformer structure. Supervised fine-tuning corrects parameters by comparing the spatial distribution and meteorological element value differences between the core downscaled meteorological data and the measured observation meteorological data. Both self-supervised pre-training and supervised fine-tuning use the aforementioned hybrid loss function. The online update mechanism periodically calculates the output deviation of the core downscaled meteorological data based on the continuously received measured observation meteorological data. When the deviation exceeds the threshold, the supervised fine-tuning process is triggered to incrementally train the parameters of the network containing the terrain constrained Transformer structure and the variational autoencoder, so as to gradually correct the weights of the downscaling network and output the downscaled meteorological data within multiple update cycles.

2. The method according to claim 1, characterized in that Reanalysis meteorological data are generated by the Global Climate Reanalysis System or the Regional Climate Reanalysis System, and digital elevation model data are obtained by aerial remote sensing or satellite remote sensing and collected in the same area. The spatial alignment and resampling steps perform projection conversion and grid resampling through the geographic information system.

3. The method according to claim 1, characterized in that The downscaling network adopts a multi-layer convolution structure combined with linear mapping to receive the fused multimodal graph feature data. Each layer of the convolution structure sets a fixed convolution kernel size and step size, and performs batch normalization and activation calculation after the output to map the fused multimodal graph feature data to the target downscaling resolution to form the initial downscaled meteorological data.

4. The method according to claim 1, characterized in that The variational autoencoder consists of an encoder and a decoder. The encoder converts the initial downscaled meteorological data into a representation vector describing the mean and variance of the latent variable. The decoder generates a reconstructed output containing uncertainty features based on reparameterized sampling. The encoding and decoding process retains the local gradient information and spatial distribution characteristics of the initial downscaled meteorological data.

5. The method according to claim 1, characterized in that: The meta-learning-based iterative update mechanism uses a lightweight residual network to process the difference between the output of the variational autoencoder and the theoretical physical value. The lightweight residual network regards the difference as input and generates correction data. The correction data is added to the output of the variational autoencoder to form a feedback loop, thereby updating the initial downscaled meteorological data under the action of physical constraints and obtaining the core downscaled meteorological data.

6. The method according to claim 1, characterized in that The theoretical physical values ​​are calculated based on the energy balance and mass conservation formulas. The energy balance infers the energy flux based on the input meteorological factors, and the mass conservation calculates the divergence based on the humidity and wind speed related factors. The two are combined to evaluate the deviations of the initial downscaled meteorological data in dynamics and thermodynamics.

Images