Near-surface meteorological field downscaling method based on topographic constraint Transform model

Abstract

The invention relates to the technical field of meteorological data processing and geographic information, in particular to a near-surface meteorological field downscaling method based on a topographic constraint Transform model, which comprises the following steps of: extracting multi-modal fusion features from reanalysis meteorological data and digital elevation model data; mapping to a target resolution by means of a topographic constraint Transform network to obtain initial downscaling meteorological data; latent variable modeling is carried out on the initial data through a variational auto-encoder to describe multi-scale uncertainty, physical constraints such as energy balance and mass conservation are incorporated in combination with an iterative updating mechanism based on meta-learning, and core downscaling meteorological data are generated; according to the invention, detail capture and dynamic adaptation of local meteorology are realized through self-supervised pre-training and online updating.

Description

Technical Field

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

[0002] The near-surface meteorological environment is often jointly affected by complex terrain, microscale climate effects, and regional atmospheric circulation. Therefore, it is difficult to capture the coupling mechanism of local characteristics and dynamic changes only relying on traditional data interpolation or simple global deep learning methods. Existing technologies generally adopt deep learning methods that integrate 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, a series of convolutional networks or shallow feature extraction modules are used to complete spatial downscaling. However, due to the lack of fine modeling of the influence of local terrain and the problem that the model is prone to problems such as reduced accuracy or blurred boundaries when facing highly coupled local meteorological elements, the performance is poor when simulating local meteorological phenomena with strong locality such as complex mountains, coastal areas, or urban heat islands.

[0003] In addition, existing technologies often only make single or simplified treatments on physical constraints, it is difficult to take into account the superposition effects of multiple physical laws, and lack an adaptive correction mechanism for highly nonlinear near-surface meteorological elements, resulting in the problem of deviation accumulation in the prediction results in non-flat terrain areas. Summary of the Invention

[0004] In view of the above-mentioned many problems existing in the prior art, the present invention provides a method for downscaling the near-surface meteorological field based on a terrain-constrained Transformer model. The present invention obtains a high-dimensional representation of preprocessed meteorological and terrain data through a terrain-constrained Transformer structure and multi-modal feature fusion, then uses a variational autoencoder to capture data uncertainty and multi-scale features, and is supplemented by meta-learning iterative correction to achieve physical consistency. The present invention can significantly improve the downscaling accuracy in complex terrain areas and has good generalization and online update capabilities.

[0005] A method for downscaling the near-surface meteorological field based on a terrain-constrained Transformer model includes the following steps: Collect reanalysis meteorological data and digital elevation model data of its corresponding area, perform spatial alignment and resampling, and generate fused multi-modal map feature data through feature extraction and fusion after preprocessing; Map the fused multi-modal graph feature data to the target downscaled resolution through a downscaling network to form initial downscaled meteorological data, and use a variational autoencoder to perform latent variable modeling on the initial downscaled meteorological data to obtain a latent variable representation that describes data uncertainty and multi-scale features. At the same time, combined with physical constraints, a feedback closed-loop is established between the data and the theoretical physical value through an iterative update mechanism based on meta-learning, thereby generating core downscaled meteorological data; Use the core downscaled meteorological data and the measured observed meteorological data as supervised samples, construct a hybrid loss function through self-supervised pre-training and supervised fine-tuning, and implement an online update mechanism to achieve the output of downscaled meteorological data.

[0006] Preferably, the reanalysis meteorological data is generated by a global climate reanalysis system or a regional climate reanalysis system, the digital elevation model data is obtained by aerial remote sensing or satellite remote sensing, and is collected in the same area. The spatial alignment and resampling steps perform projection transformation and grid resampling through a geographic information system.

[0007] Preferably, the multi-modal feature extraction and fusion step uses a dual-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. A dynamic adjacency matrix is constructed using the terrain weight factor generated by the embedded terrain data, and then the node features are propagated and fused at multiple levels through a graph convolutional network to form fused multi-modal graph feature data.

[0008] Preferably, the downscaling network receives the fused multi-modal graph feature data in a manner combining a multi-layer convolutional structure and a linear mapping. Each convolutional structure is set with a fixed convolutional kernel size and stride, and batch normalization and activation calculations are performed after the output to map the fused multi-modal graph feature data to the target downscaled resolution to form initial downscaled meteorological data.

[0009] Preferably, the variational autoencoder consists of an encoder and a decoder. The encoder converts the initial downscaled meteorological data into a representation vector that describes the mean and variance of the latent variables, and the decoder generates a reconstructed output containing uncertainty features based on reparameterized sampling. The encoding and decoding processes retain the local gradient information and spatial distribution characteristics of the initial downscaled meteorological data.

[0010] Preferably, the iterative update mechanism based on meta-learning 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 takes the difference as input and generates correction data, and the correction data is added to the output of the variational autoencoder to form a feedback closed-loop, thereby updating the initial downscaled meteorological data under the action of physical constraints and obtaining the core downscaled meteorological data.

[0011] Preferably, the theoretical physics 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 deviation of the initial downscaled meteorological data in terms of dynamics and thermodynamics.

[0012] Preferably, the core downscaled meteorological data and the measured meteorological data obtained by independent observation are used to form a supervised sample, and a hybrid loss function is established by combining self-supervised pre-training and supervised fine-tuning. The hybrid loss function includes a reconstruction loss based on difference statistics and a constraint loss reflecting physical consistency, and the gradients are calculated by summing the two losses according to a preset weight.

[0013] Preferably, in self-supervised pre-training, masked data samples are generated by performing regional masking on the reanalysis meteorological data, and a multi-modal network including a terrain-constrained Transformer structure is used to perform reconstruction training on the masked data samples; in supervised fine-tuning, the parameters are corrected by comparing the spatial distribution and meteorological element value differences between the core downscaled meteorological data and the measured meteorological data. Both self-supervised pre-training and supervised fine-tuning use the hybrid loss function.

[0014] Preferably, the online update mechanism periodically calculates the output deviation of the core downscaled meteorological data based on the continuously received measured meteorological data. When the deviation exceeds the threshold, the supervised fine-tuning process is triggered to perform incremental training on the parameters of the network including 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.

[0015] Compared with the prior art, the advantages and beneficial effects of the present invention are as follows: The present invention combines the terrain-constrained Transformer structure and the graph convolution fusion technology to achieve an explicit modeling effect of the influence of local complex terrain; The present invention obtains the latent variable distribution through a variational autoencoder to characterize the uncertainty of meteorological data at multiple scales; 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; The present invention strengthens the generalization ability and real-time adaptability of the model in the case of data scarcity or local anomalies by introducing self-supervised pre-training and an online update mechanism.

[0016] Through the above technical means, the present invention achieves the comprehensive effects of capturing local meteorological details under multi-modal data fusion and meeting physical consistency, characterizing uncertainty in multi-scale latent variable modeling, and gradually approaching the true physical laws in iterative updates. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 is a schematic flowchart of the method of the present invention; Figure 2 is a schematic diagram of the extraction of fused multi-modal graph features in the present invention; Figure 3 is a schematic diagram of supervised fine-tuning and online update in the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0018] 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 the sake of explanation, many specific details are set forth in order to provide a thorough understanding of the embodiments of the present disclosure. However, obviously, one or more embodiments may be implemented without these specific details. In addition, in the following description, descriptions of well-known structures and technologies are omitted to avoid unnecessarily obscuring the concepts of the present disclosure.

[0019] The terms used herein are merely for describing specific embodiments and are not intended to limit the present disclosure. The terms "including", "comprising", etc. used herein indicate the presence of the described features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0020] 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.

[0021] As Figure 1 shown, a near-surface meteorological field downscaling method based on a terrain-constrained Transformer model includes the following steps: Collect reanalysis meteorological data and digital elevation model data of its corresponding area, perform spatial alignment and resampling, and generate fused multi-modal graph feature data through feature extraction and fusion after preprocessing; The key to realizing the downscaling of the near-surface meteorological field in this invention lies in, by collecting meteorological data from atmospheric reanalysis and digital elevation model data in the corresponding area, strictly aligning and resampling the two in space, and relying on advanced feature extraction and multi-modal data fusion technologies after preprocessing, finally forming fused multi-modal map feature data. Specifically, the reanalysis meteorological data collected first in this invention includes 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, with relatively low spatial resolution. At the same time, the digital elevation model data in the corresponding area is stored in GeoTIFF format, reflecting features such as terrain undulation, slope, aspect, and curvature. The collection stage ensures that the two cover the same area, enabling accurate matching in subsequent data processing.

[0022] After data collection, the collected reanalysis meteorological data and digital elevation model data are spatially aligned and resampled through a geographic information system. This step includes coordinate system conversion, projection correction, and grid unification processing of the original data, so as to ensure that the meteorological data and terrain data are completely consistent in the coordinate system, spatial range, and resolution. Next, preprocessing is respectively performed on the aligned data. Among them, the reanalysis meteorological data is subjected to noise suppression and information enhancement through Gaussian smoothing and local statistics (such as local mean, variance, and gradient calculation), while the digital elevation model data calculates terrain indices through edge detection operators and performs normalization processing. The data obtained after preprocessing provides a stable and consistent basis for subsequent feature extraction.

[0023] After the data preprocessing is completed, this invention adopts advanced feature extraction and fusion methods, aiming to extract high-dimensional features from the preprocessed data and establish associations between multi-modal data. Specifically, this invention uses a strategy of fusing multi-modal map feature data, inputs the reanalysis meteorological data and digital elevation model data into their respective feature extraction modules respectively, and fuses their features, thereby forming a comprehensive information expression that can reflect the spatial distribution of the meteorological field and the influence of the terrain. This fusion process not only retains the spatio-temporal dynamic information of the meteorological data, but also incorporates the local environmental constraints reflected by the terrain data, providing rich inputs for the subsequent downscaling model.

[0024] Generally speaking, the present invention realizes a full - process closed - loop from data acquisition, pre - processing to feature extraction and fusion. Each step ensures data consistency through precise spatial matching and data normalization, and adopts advanced graphical feature fusion technology to build a solid input foundation for the downscaling model. Finally, the fused multi - modal graph feature data becomes the core intermediate product in the entire downscaling process. It not only contains the macroscopic atmospheric features of re - analysis meteorological data but also fuses the local detail information reflected by topographic data, providing the necessary high - dimensional feature support for the subsequent use of the terrain - constrained Transformer structure, variational auto - encoder, 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 utilization of multiple data sources for downscaling through cross - modal fusion, thus laying a foundation for the refined expression of the near - surface meteorological field.

[0025] In the implementation process, the present invention first relies on advanced data acquisition devices and sensors to obtain re - analysis meteorological data and digital elevation model data. These two types of data are strictly region - divided and time - matched before acquisition to ensure high consistency and comparability. Subsequently, for spatial alignment processing, mature geographic information system algorithms such as nearest - neighbor interpolation and bilinear interpolation are used to eliminate systematic errors between data acquisition devices and sensors; at the same time, when resampling the data, a resampling strategy with fixed parameters is adopted to ensure the uniformity of grid data. In the pre - processing stage, the re - analysis meteorological data is smoothed by a Gaussian filter and then 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 calculates terrain gradient information using filtering methods such as Sobel and Laplacian, and combines normalization methods to standardize the data scale. Finally, these pre - processing steps provide high - quality original inputs for subsequent feature extraction.

[0026] Through the above steps, the overall design of the present invention realizes the integration of multi - modal information from different data sources into a unified high - dimensional expression. The core lies in data consistency and multi - modal fusion strategies, ensuring sufficient information content at the input data 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.

[0027] Preferably, the re - analysis meteorological data is generated by a global climate re - analysis system or a regional climate re - analysis system, and the digital elevation model data is obtained by aerial remote sensing or satellite remote sensing and collected in the same region. The spatial alignment and resampling steps are performed through the geographic information system to execute projection transformation and grid resampling.

[0028] In the present invention, a preferred embodiment further specifies the specific requirements for data acquisition and spatial alignment. Further analysis shows that in the preferred embodiment, the meteorological data is clearly generated by the global climate reanalysis system or the regional climate reanalysis system. These systems adopt multi-source observation data and numerical model assimilation technology to generate meteorological data with high spatio-temporal consistency and historical continuity. To ensure the accuracy and usability of the data, the present invention requires that the reanalysis meteorological data collected 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. During data acquisition, it is also particularly required that the digital elevation model data collected is sourced from aerial remote sensing or satellite remote sensing. These data are provided by specialized remote sensing equipment, with a finer resolution and can truly reflect the details of the terrain in the target area. The digital elevation model data is stored in the GeoTIFF format, and its accuracy and grid density meet the requirements of subsequent graphic processing and feature extraction in terms of technical indicators.

[0029] In the preferred embodiment, the data acquisition requires that the reanalysis meteorological data and the digital elevation model data be collected simultaneously in the same area to ensure strict spatial correspondence between the two. For example, when studying a specific area, the reanalysis meteorological data and the digital elevation model data collected should be consistent in terms of geographical coordinates, regional scope, and collection time, so as to ensure that no additional errors are introduced during the subsequent spatial alignment process. The spatial alignment and resampling steps are processed automatically by a geographic information system. The two types of data are converted to the same coordinate system through projection transformation and grid resampling. The projection transformation algorithms used in specific implementation include common Mercator projection, UTM projection, etc., while the grid resampling method uses bilinear interpolation or cubic convolution interpolation technology. To ensure the numerical continuity and smoothness of the data, the parameters of the interpolation algorithm used (such as the interpolation window size, weight distribution) are determined in pre-experiments and strictly implemented during the implementation process.

[0030] In addition, the preferred embodiment emphasizes that during the spatial alignment and resampling process, the coordinate system conversion, spatial range truncation, and resolution adjustment of the reanalysis meteorological data and the digital elevation model data all adopt a standardized process to ensure that the processing results have high consistency and reproducibility. For example, if the original resolution of the reanalysis meteorological data collected is 0.25°×0.25°, and the resolution of the digital elevation model data is 10 meters, when performing projection transformation and resampling through GIS software, first determine the bounding box of the target area and generate a unified grid with a fixed resolution; on this basis, then use the interpolation algorithm to convert the digital elevation model data into a grid format matching 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 corresponding in terms of spatial position, resolution, and numerical range.

[0031] In specific operations, 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 processing. The operation steps include: first, loading the original reanalysis meteorological data and digital elevation model data, then selecting an appropriate projection conversion tool for coordinate conversion, subsequently 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, for example, the bilinear interpolation formula:

[0032] where , and , are the coordinates of adjacent grid points, and represents the interpolated data value. The above formula ensures that the resampled data is numerically smooth and continuous. Through strict parameter setting and standard operations, this preferred embodiment can provide high-quality and consistent input data in the data acquisition and preprocessing stage, providing strong guarantee for subsequent processing.

[0033] Preferably, the multi-modal 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. A dynamic adjacency matrix is constructed using the terrain weight factor generated by the embedded terrain data, and then the node features are propagated and fused at multiple levels through a graph convolutional network to form fused multi-modal graph feature data.

[0034] As Figure 2 shown, the multi-modal feature extraction and fusion step adopts a two-branch convolutional network containing a terrain-constrained Transformer structure to make full use of the complementary information of the reanalysis meteorological data and the digital elevation model data. Specifically, first, a dedicated convolutional network channel is used to process the preprocessed reanalysis meteorological data. This channel maps the original input to a high-dimensional feature space through continuous convolutional layers, activation functions, and batch normalization modules to generate embedded meteorological data. The selection of the convolutional kernel size, stride, and activation function (such as ReLU) in this process is determined based on experimental data to ensure the effectiveness and stability of feature extraction. For example, in one embodiment, the convolutional kernel size used is 3×3, the stride is fixed at 1, and the number of output channels increases gradually after 3 layers of convolution to ensure capturing spatial features from low-level to high-level.

[0035] Meanwhile, for the preprocessed digital elevation model data, another dedicated convolutional network channel is used to extract terrain features and generate embedded terrain data. This channel also employs multi-layer convolutional operations, but a feature extraction module designed according to the characteristics of terrain data is added to the network structure, which can highlight terrain features such as local slope, aspect, and curvature. During this process, the embedded terrain data generates a terrain weight factor after non-linear transformation. This weight factor reflects the sensitivity of terrain changes in each region, and its generation process is normalized using the Sigmoid function, making the weight values between 0 and 1.

[0036] Furthermore, in the preferred embodiment, a dynamic adjacency matrix is constructed using the terrain weight factor generated from the embedded terrain data. The construction of the dynamic adjacency matrix takes into account the terrain differences at each spatial position, and its calculation method can be expressed as:

[0037] where, represents the weight between node and node , and represent the terrain weight factor vectors at the corresponding positions respectively, is the smoothing parameter. This formula ensures that the weights are high between similar regions and decay rapidly between regions 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.

[0038] 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 neighbor nodes, and its propagation formula usually adopts:

[0039] where, represents the node feature matrix of the th layer, is the matrix formed by adding the self-connection to the dynamic adjacency matrix, is the degree matrix of , is the weight matrix of the th layer, is the non-linear activation function. Through the processing of consecutive multi-layer graph convolutional networks, the feature information from the reanalysis meteorological data and the digital elevation model data can be fully fused, and finally, fusion multi-modal graph feature data reflecting the regional spatio-temporal characteristics and terrain constraints is formed.

[0040] The entire multi-modal feature extraction and fusion process fully utilizes the advantages of the terrain-constrained Transformer structure. This structure 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 a graph convolutional network, ensuring both the capture of local details and the establishment of global dependencies. This method has high operability in actual operation, and relevant parameters (such as the number of convolutional layers, the size of convolutional kernels, the type of activation function, the smoothing parameter of the adjacency matrix 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 finally fused multi-modal graph feature data can truly reflect the complex relationship between the meteorological field and the terrain within the area.

[0041] In summary, by adopting a dual-branch convolutional network with a terrain-constrained Transformer structure and combining a graph convolutional network to achieve feature fusion, the preferred embodiment can extract highly representative high-dimensional features from the preprocessed reanalysis meteorological data and digital elevation model data, and on this basis, construct a dynamic adjacency matrix, and then complete multi-level feature propagation and fusion through a graph convolutional network. This process provides rich input data for the subsequent downscaling stage, ensuring that the entire downscaling method realizes the deep fusion of meteorological data and terrain data while maintaining regional spatial consistency, and provides a solid data support and theoretical basis for the downscaling of the near-surface meteorological field.

[0042] Mapping the fused multi-modal graph feature data to the target downscaling resolution through a downscaling network to form initial downscaled meteorological data, and using a variational autoencoder to perform latent variable modeling on the initial downscaled meteorological data to obtain a latent variable expression that describes data uncertainty and multi-scale features. At the same time, combined with physical constraints, a feedback closed-loop is established between the data and the theoretical physical values through an iterative update mechanism based on meta-learning, thereby generating core downscaled meteorological data; The overall idea of the present invention to realize the downscaling of the near-surface meteorological field is to map the fused multi-modal 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 to obtain a latent variable expression that can describe data uncertainty and multi-scale features; finally, combined with theoretical physical constraints, through an iterative update mechanism based on meta-learning, a feedback closed-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.

[0043] Generally, this method uses key technologies such as data preprocessing, feature extraction, multi-modal information fusion, downscaling network mapping, variational autoencoder latent variable modeling, and meta-learning iterative update to form a complete data processing closed-loop. This not only ensures consistent processing of data in space, time, and modality, but also guarantees that the downscaled data meets the requirements of dynamics and thermodynamics theories by introducing physical constraints. In the solution, the fused multi-modal graph feature data is formed by fusing the preprocessed reanalysis meteorological data and digital elevation model data through advanced graph convolutional network technology. The data contains both the macroscopic background information of the meteorological field and the local detailed features of the terrain, providing a high-dimensional and rich feature representation for subsequent downscaling operations. Subsequently, the downscaling network maps this fused feature data to achieve the conversion of data resolution and generate initial downscaled meteorological data. This stage mainly involves convolutional and linear mapping operations, and its design parameters are strictly debugged in practical applications.

[0044] On this basis, a variational autoencoder is used for latent variable modeling. The initial downscaled meteorological data is converted into a low-dimensional representation of latent variable mean and variance through the encoder, and then the decoder is used in combination with reparameterized sampling to obtain a latent variable expression reflecting data uncertainty and multi-scale features. This process ensures that the intrinsic statistical characteristics of the downscaled data are retained. Finally, the solution combines theoretical physical constraints (such as energy balance and mass conservation), and uses 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, and correction data is generated through a residual network, thereby forming a closed-loop feedback correction system. This makes the data more compliant with physical principle requirements after multiple iterative updates, and the generated core downscaled meteorological data has higher reliability and fineness. The overall method has a compact structure and distinct levels, and ensures full transmission and utilization of information between different scales and modalities through seamless docking between modules, thereby achieving effective downscaling of the near-surface meteorological field.

[0045] Preferably, the downscaling network adopts a combination of a multi-layer convolutional structure and linear mapping to receive the fused multi-modal graph feature data. Each layer of the convolutional structure is set with a fixed convolutional kernel size and stride, and batch normalization and activation calculations are performed after the output to map the fused multi-modal graph feature data to the target downscaled resolution, forming the initial downscaled meteorological data.

[0046] In a preferred embodiment, the downscaling network adopts a combination of a multi-layer convolutional structure and a linear mapping to receive the aforementioned fused multi-modal graph feature data and map it to the target downscaled resolution. In specific operations, when designing this downscaling network, multi-layer convolutional neural networks are first considered to fully extract the local information of the input features. Each layer uses a convolutional kernel of a fixed size (such as 3×3 or 5×5) and a fixed stride (such as 1 or 2) to ensure the parameter stability of the convolutional operation and the unity of the network structure. After the convolutional operation, the network performs batch normalization on each layer to eliminate the bias of the data distribution, improve the convergence speed during model training, and uses a non-linear activation function (such as ReLU or Leaky ReLU) to process the convolutional result to enhance the expression ability of the network.

[0047] The linear mapping part refers to, after multi-layer convolutional operations, using a fully connected layer or a linear transformation to reduce the dimensionality of the multi-dimensional convolutional output or map it to a predetermined target resolution. This process ensures the unity of the output data in terms of spatial resolution while maintaining the information transfer of the original multi-modal features. To achieve numerical smoothing and feature preservation during the downscaling process, a residual connection mechanism is also introduced in the network design, which can alleviate the vanishing gradient problem in the deep network structure and ensure the effective transmission of low-level features. For example, in a specific implementation, the following steps can be adopted: first, the input fused multi-modal graph feature data is used to gradually extract features through a series of convolutional layers, then dimensionality reduction is achieved through a fully connected layer, and the finally output initial downscaled meteorological data maintains continuity and spatial consistency within the target area. During the operation process, the weights, strides, activation functions, and normalization parameters of each layer of convolutional kernels are calibrated through pre-experiment data to ensure the stability of the output of the entire downscaling network and high reproducibility. The design of this downscaling network follows the architecture of a conventional convolutional neural network while customizing parameters according to 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 terrain feature information, thereby providing a reliable initial data basis for further processing of subsequent data.

[0048] 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 variables, and the decoder generates a reconstructed output containing uncertainty features based on reparameterized sampling. The encoding and decoding processes retain the local gradient information and spatial distribution characteristics of the initial downscaled meteorological data.

[0049] 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 representation that can reflect the data's uncertainty and multi-scale characteristics. 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 contains both 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 data statistical distribution consistent. After obtaining the latent variable mean vector and variance vector , using the reparameterization trick, sample the latent variable from the normal distribution, and its mathematical expression is:

[0050] where is the noise term sampled from the standard normal distribution, represents the latent variable mean, is the standard deviation, and this formula ensures that the gradient can be transmitted during the backpropagation process. The decoder part also uses a multi-layer convolutional or fully connected structure to map the sampled latent variable back to the data space and generate a reconstructed output containing uncertainty features. During the encoding and decoding processes, by designing the network structure, the local gradient information and spatial distribution features are retained to ensure 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, and the goal is to minimize the sum of the reconstruction error and the KL divergence term. The mixed loss function is usually expressed as:

[0051] where is the reconstruction error, is the KL divergence, is the adjustment parameter, 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 the multi-scale feature information, so that the generated latent variable representation provides a solid data representation basis for subsequent iterative updates. In practical applications, the number of layers, the number of nodes, and the selection of activation functions of the encoder and decoder structures are all strictly debugged in experiments to ensure that the latent variable representation has high stability and representativeness. This implementation method provides an effective statistical basis and uncertainty description for the generation of core downscaled meteorological data.

[0052] Preferably, the iterative update mechanism based on meta - learning uses a lightweight residual network to process the difference between the output of the variational auto - encoder and the theoretical physical value. The lightweight residual network takes the said difference as input and generates correction data. The correction data is added to the output of the variational auto - encoder to form a feedback closed - loop, thereby updating the initial downscaled meteorological data under the action of physical constraints and obtaining the core downscaled meteorological data.

[0053] The iterative update mechanism based on meta - learning aims to correct the difference between the output of the variational auto - encoder and the theoretical physical value by using a lightweight residual network, so as to gradually update the initial downscaled meteorological data under physical constraints and generate the core downscaled meteorological data. In specific operations, first, according to the difference between the initial downscaled meteorological data and the theoretical physical value (obtained by calculating the energy balance and mass conservation formulas), the physical constraint error is calculated. The calculation of the theoretical physical value is based on the input meteorological elements, where the energy balance is obtained by measuring thermal radiation and sensible heat flux, and the mass conservation is calculated by divergence based on elements such as humidity and wind speed. The calculation formula can be, for example:

[0054] where, represents the physical quantity value corresponding to the initial downscaled meteorological data, represents the theoretical physical value, represents the absolute difference between the two. Subsequently, this difference is fed into the lightweight residual network as input. The residual network structure is designed with fixed parameters and consists of several convolutional layers. Each layer performs batch normalization and non - linear activation operations. The output correction data represents the correction amount for the initial downscaled meteorological data. The correction data is added to the output of the variational auto - encoder, that is, a feedback closed - loop is formed to achieve the adaptive correction between the data and the theoretical physical value. This feedback closed - loop process is achieved through iterative training. Each iteration updates the initial downscaled meteorological data to make it gradually approach the theoretical physical requirements.

[0055] For example, in an embodiment, by setting the number of iterations to N, the physical error is calculated in each iteration and the correction data is generated through the residual network until the error drops to a preset threshold. This meta - learning iterative update mechanism makes full use of the high - efficiency computing power and adaptive characteristics of the lightweight residual network, realizes the continuous refinement update of data under the action of physical constraints, so that the generated core downscaled meteorological data can be more in line with the theoretical requirements in dynamics and thermodynamics. During the operation process, each parameter (such as the convolutional kernel size, the number of residual network layers, the preset threshold, etc.) is determined based on multiple sets of experimental data to ensure that the iterative update mechanism has high stability and reliability and can adapt to the complexity and change trend of different meteorological data.

[0056] Preferably, the theoretical physical values are calculated based on the energy balance and mass conservation formulas. The energy balance estimates the energy fluxes from the input meteorological elements, and the mass conservation calculates the divergence based on humidity and wind speed related elements. The two are combined to evaluate the deviations of the initial downscaled meteorological data in terms of dynamics and thermodynamics.

[0057] The calculation of the theoretical physical values is determined based on two basic principles: energy balance and mass conservation. For the energy balance part, based on the input meteorological elements such as temperature, radiation flux, and wind speed, an estimation of the energy transfer within the region is obtained by establishing an energy flux calculation formula. Specifically, the energy flux calculation can use the following standard formula:

[0058] where, represents the regional average energy flux, represents the air density, represents the specific heat capacity, represents the temperature at each grid point, is the reference temperature, is the number of grid points. This formula estimates the energy transfer value by calculating the difference between each grid point and the reference temperature, thus obtaining an estimation result of the energy balance. The mass conservation part is obtained by calculating the divergence of the air flow based on humidity and wind speed related elements, and its basic calculation formula is:

[0059] where, represents the mass conservation deviation, represents the wind speed vector field, and this formula describes the distribution of the gas mass flow within the region. In practical applications, the theoretical physical values are formed by combining the above two parts of data, namely the energy flux and mass divergence, according to a preset weight, thus forming an evaluation standard for the physical consistency of the initial downscaled meteorological data. Specifically, when operating, the meteorological data collected by sensors provides the numerical information of temperature, humidity, and wind speed. Then, the theoretical physical values are calculated using a numerical model, and the corresponding physical quantities in the initial downscaled meteorological data in the same region are compared to calculate the difference between the two. This difference is used as the input for the subsequent meta - learning iterative update mechanism to ensure that the feedback correction process is strictly based on physical principles. By implementing this calculation process, clear physical constraints can be introduced during the data processing, making the downscaled results conform to statistical laws and satisfy the basic dynamics and thermodynamics theories. In practice, the parameters in each calculation formula (such as air density, specific heat capacity, reference temperature, weight coefficient, etc.) are pre - determined according to the regional meteorological characteristics and empirical data and used as fixed parameters in the downscaling model, thus ensuring the accuracy and reproducibility of the calculation of the theoretical physical values. This implementation method of calculating the theoretical physical values provides a physical basis for the entire downscaling process and is an important link to ensure the rationality of the numerical and physical performance of the downscaled meteorological data.

[0060] As shown Figure 2 in the figure, the core downscaled meteorological data and the measured observed meteorological data are used as supervised samples, and a hybrid loss function is constructed through self-supervised pre-training and supervised fine-tuning, and an online update mechanism is implemented to realize the output of the downscaled meteorological data.

[0061] This embodiment relates to the overall process of using the core downscaled meteorological data and the measured observed meteorological data to form supervised samples, and then constructing a hybrid loss function through a method combining self-supervised pre-training and supervised fine-tuning, and implementing an online update mechanism to realize the output of the downscaled meteorological data. The overall solution first obtains the core downscaled meteorological data through a preprocessing module. This data is generated by multi-modal feature fusion, a downscaling network, a variational autoencoder, and an iterative update mechanism based on meta-learning. It better retains the macroscopic atmospheric characteristics of the reanalysis meteorological data and the topographic constraint information provided by the digital elevation model data in terms of spatial resolution and data representation.

[0062] In order to further improve the consistency between the downscaled meteorological data and the measured observed data in terms of spatial distribution and meteorological element values, the present invention forms supervised samples from the core downscaled meteorological data and the measured observed meteorological data obtained through independent observation devices, and uses this combined data to conduct targeted training and fine-tuning on the deep network. Through self-supervised pre-training, the method performs regional masking processing on the reanalysis meteorological data to generate masked data samples, and uses a multi-modal network containing a terrain-constrained Transformer structure to reconstruct and train the masked samples, so that the model learns the internal spatio-temporal correlation and inter-modal interaction of the data in an unsupervised state. Then, in the supervised fine-tuning stage, by comparing the core downscaled meteorological data with the measured observed meteorological data in terms of spatial distribution and various meteorological element values, the deviation between the two is quantified and fed back into the network parameters to realize the gradual correction of the model in a supervised environment.

[0063] The whole process is carried out by constructing a hybrid loss function, which consists of a reconstruction loss based on difference statistics and a constraint loss that reflects physical consistency. The gradient of each loss term is calculated by summing up the preset weights to guide the update of model parameters. Finally, through the online update mechanism, the system periodically calculates the core downscaled meteorological data output by the model on the basis of continuously receiving the measured observation meteorological data. When the output deviation exceeds the preset threshold, incremental training is triggered, and the model parameters are gradually corrected to ensure that the output of the downscaled meteorological data always maintains a high consistency with the measured observation data. The overall method not only achieves innovation in data fusion, downscaling mapping, and latent variable modeling, but also builds 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, thereby providing reliable technical support for the downscaling of near-surface meteorological fields.

[0064] Preferably, the core downscaled meteorological data and the measured meteorological data obtained by independent observation constitute a supervised sample, 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.

[0065] The core downscaled meteorological data and the measured meteorological data obtained by independent observation constitute the supervised 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, radars or satellite observation systems, and stable observation values ​​are obtained after preprocessing. The composition of the supervised samples requires that the two data sources have strict correspondence in time, space and data expression to ensure the effective transmission of supervisory information. To achieve this goal, this embodiment establishes a mixed 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 aims to measure the deviation between the core downscaled meteorological data and the measured meteorological data in numerical and spatial distribution. Specifically, the mean square error (MSE) can be used as a metric, and its formula is:

[0066] 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, denotes the total number of grid points. The constraint loss reflects the gap between the initial downscaled meteorological data and the theoretical physical values in terms of thermodynamics and dynamics 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, using the formula:

[0067] where is the estimated value of the energy flux of the core downscaled meteorological data in the th region, is the theoretical value calculated according to the energy balance formula, is the number of regional divisions. The final mixed loss function is expressed as a weighted sum of the above two losses with preset weight coefficients:

[0068] where and are weight parameters preset according to experimental data to balance the effects of data reconstruction and physical constraints. During the implementation process, the system optimizes the mixed loss function through the gradient descent method to guide the update of network parameters, so that the downscaling model can effectively reduce the gap between the core downscaled meteorological data and the measured observed meteorological data during the training process. This process has high operability, and the parameter settings and the specific composition of the loss function can be calibrated according to experimental data, and the stability and consistency of the gradient direction can be ensured during the supervised training process, so as to achieve efficient and stable network training and model update.

[0069] Preferably, as Figure 3 shows, self-supervised pre-training generates masked data samples by performing regional masking on reanalysis meteorological data, and uses a multi-modal network containing a terrain-constrained Transformer structure to perform reconstruction training on the masked data samples; supervised fine-tuning corrects parameters by comparing the spatial distribution and meteorological element value differences between the core downscaled meteorological data and the measured observed meteorological data. Both self-supervised pre-training and supervised fine-tuning use the above-mentioned mixed loss function.

[0070] The self-supervised pre-training part uses regional masking processing on reanalysis meteorological data to generate masked data samples. This processing method constructs missing data samples by masking some regions or data channels in the reanalysis meteorological data in a fixed manner, forcing the model to learn the inherent spatio-temporal correlation of the data. In specific operations, the regional masking processing adopts a rectangular mask or a random mask method, and the mask ratio is determined within a certain range (e.g., 20% to 30%) in the pre-experiment to ensure that the generated masked samples can fully simulate the data missing situation without making the model overly dependent on the mask structure. Then, the masked data samples are reconstructed and trained by a multi-modal network containing a terrain-constrained Transformer structure. This network combines the multi-modal information of reanalysis meteorological data and digital elevation model data during the processing, and uses the self-attention mechanism to strengthen the interaction between local and global information, enabling the model to capture the complex spatio-temporal patterns inherent in the data in an unsupervised state. During the reconstruction training process, the network uses the same loss calculation method as the hybrid loss function, that is, considering both the reconstruction error and the physical constraint loss, and then updating the network parameters to achieve the self-supervised pre-training goal.

[0071] In the supervised fine-tuning part, the preferred implementation method corrects the parameters by comparing the differences in the spatial distribution and meteorological element values between the core downscaled meteorological data and the independently collected measured observed meteorological data. Specifically, in this process, the core downscaled meteorological data and the measured observed meteorological data are compared point by point on the same spatial grid, and the difference matrix is calculated; subsequently, this difference matrix is input into the deep network as feedback information, and the network parameters are updated according to the gradient calculated by the hybrid loss function. This supervised fine-tuning process depends on both the precisely constructed hybrid loss function and the continuous observed data as an external supervision signal, enabling the downscaling model to continuously converge to the optimal state during the training process. In actual operation, the gradient descent method (such as the Adam or AdamW optimization algorithm) used in the supervised fine-tuning process is used to iteratively update the network parameters until the reconstruction error and the physical constraint loss between the model output and the measured observed meteorological data are reduced to the preset threshold. The entire self-supervised pre-training and supervised fine-tuning processes adopt a unified hybrid loss function design to ensure the consistency of the training objectives of the model in the self-supervised and supervised states, and to achieve the best balance between the two parts of the training by adjusting the preset weight parameters, thereby effectively improving the output quality and physical consistency of the downscaled meteorological data.

[0072] Preferably, the online update mechanism periodically calculates the output deviation of the core downscaled meteorological data based on continuously received measured observed meteorological data. When the deviation exceeds the threshold, it triggers a supervised fine-tuning process 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.

[0073] By periodically calculating the deviation between the measured observed meteorological data and the core downscaled meteorological data, and triggering incremental training when the deviation exceeds the preset threshold, the model parameters are dynamically updated, thus realizing the adaptive adjustment of the downscaling model during the continuous data reception process. In specific operations, the system continuously collects measured observed meteorological data, makes real-time comparison with the core downscaled meteorological data on the same spatial grid, and calculates the output deviation. The deviation can use the root mean square error or the mean absolute error as the measurement index; for example, the system can set to perform a statistical calculation on the model output data every 24 hours and compare the calculated error value with the preset threshold.

[0074] When the monitored error value exceeds the set threshold, the system automatically starts the online update module, which incrementally trains the parameters of the multimodal network containing the terrain-constrained Transformer structure and the variational autoencoder. During the incremental training process, the online update module uses small batches of data as training samples and fine-tunes the network parameters by the gradient descent method. The updated model can better adapt to the latest collected measured observed meteorological data. Specifically, the online update module first constructs a supervised sample according to the latest observation data, calculates a new gradient through the hybrid loss function, and then uses this 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.

[0075] Through multi - cycle online updates, the model parameters gradually tend to be optimal within each update cycle. As a result, the downscaled meteorological data output can continuously maintain consistency with the measured observed 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 observed meteorological data within each cycle. When the error exceeds the set allowable range, the online incremental training module is automatically called to correct the parameters. This online update mechanism not only realizes the real - time adaptive adjustment of the model parameters but also constructs a data feedback closed - loop, enabling the downscaled model to always output meteorological data that conforms to physical constraints and actual observation conditions during long - term operation. In actual operation, the implementation of the online update mechanism depends on a distributed computing platform and a real - time data acquisition system, ensuring that the system can respond to data changes in a short time and achieve continuous optimization of the model parameters through an automated training process, thus providing long - term and stable technical support for the downscaling of the near - surface meteorological field.

[0076] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product. Therefore, the present application can take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects.

[0077] The above are only the embodiments of the present application and are not used to limit the present application. For those skilled in the art, various changes and modifications can be made to the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall 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 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 the theoretical physical value through an iterative update mechanism based on meta-learning, thereby generating core downscaled meteorological data; 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.

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 acquired 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 multimodal feature extraction and fusion step adopts 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, and the node features are then propagated and fused at multiple levels through a graph convolutional network to form fused multimodal graph feature data.

4. 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.

5. 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 characteristics based on reparameterized sampling. The encoding and decoding process retains the local gradient information and spatial distribution characteristics of the initial downscaled meteorological data.

6. The method according to claim 1, characterized in that: The iterative update mechanism based on meta-learning 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 physical constraints and obtaining the core downscaled meteorological data.

7. 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 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.

8. The method according to claim 1, characterized in that: The core downscaled meteorological data and the measured meteorological data obtained by independent observation constitute 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. The gradient is calculated by summing the two losses according to the preset weights.

9. The method according to claim 1, characterized in that: Self-supervised pre-training generates masked data samples by performing regional masking on the reanalyzed meteorological data, and uses a multimodal network containing a terrain-constrained Transformer structure to perform reconstruction training on the masked data samples; supervised fine-tuning corrects parameters by comparing the spatial distribution and meteorological element value gap between the core downscaled meteorological data and the measured observation meteorological data. Both self-supervised pre-training and supervised fine-tuning use the hybrid loss function.

10. The method according to claim 1, characterized in that: 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.

Images