Multi-modal meteorological data downscaling generation method and device and computer equipment

Abstract

The invention relates to a multi-modal meteorological data downscaling generation method and device and computer equipment. The method comprises the following steps: acquiring topographic data and meteorological data in a to-be-measured area; inputting the topographic data and the meteorological data into a pre-trained deep learning model, carrying out feature coding on the topographic data and the meteorological data through the deep learning model to obtain a topographic feature vector and a meteorological feature vector, carrying out feature interaction on the topographic feature vector and the meteorological feature vector to obtain interaction features, and carrying out feature extraction on the interaction features; and carrying out downscaling decoding on the interaction features to obtain downscaling meteorological data. By adopting the method, the downscaling generation efficiency and the physical rationality of the meteorological data can be considered.

Description

Technical Field

[0001] This application relates to the interdisciplinary field of meteorological data processing and artificial intelligence, and in particular to a method, apparatus and computer equipment for downscaling multimodal meteorological data. Background Technology

[0002] Spatial downscaling of meteorological data is crucial for improving the spatial resolution of weather forecasts, supporting local decision-making, enhancing environmental impact assessments, studying local climate change, and improving the quality of meteorological services. Large-scale climate models typically have low spatial resolution, making it difficult to capture local meteorological phenomena (such as urban climate effects and mountain climate effects). Spatial downscaling transforms the meteorological data output from these large-scale models into higher-resolution data, thereby improving the accuracy of local weather event predictions. High-resolution meteorological data can be used for disaster prevention and mitigation planning, infrastructure design, and environmental protection.

[0003] Traditional meteorological spatial downscaling methods mainly include dynamic downscaling and statistical downscaling. Dynamic downscaling primarily utilizes regional climate models (RCMs) to generate high-resolution (e.g., 5-50 km) regional climate data by solving physical equations (such as fluid dynamics and thermodynamics) under large-scale boundary conditions provided by global climate models (GCMs). Its advantage is its ability to capture physical processes such as topographic effects and local circulation. Its limitation is high computational resource consumption and reliance on the accuracy of the GCM input. Statistical downscaling uses statistical methods to transform low-resolution climate data output from global climate models (GCMs) into high-resolution regional or local climate data. Its advantage is low computational cost, making it suitable for long-term simulations. Its limitation is its inability to resolve unobserved physical mechanisms and its limited generalization ability. Therefore, traditional meteorological data downscaling methods suffer from low efficiency and poor physical reliability. Summary of the Invention

[0004] Therefore, it is necessary to provide a method, apparatus, computer equipment, computer-readable storage medium, and computer program product for downscaling multimodal meteorological data to address the above-mentioned technical problems, which can balance the efficiency of meteorological data downscaling and generation with physical rationality.

[0005] Firstly, this application provides a method for downscaling multimodal meteorological data, including:

[0006] Acquire topographic and meteorological data for the area to be measured;

[0007] Topographic and meteorological data are input into a pre-trained deep learning model. The deep learning model encodes the topographic and meteorological data separately to obtain topographic feature vectors and meteorological feature vectors. The topographic and meteorological feature vectors are then interacted to obtain interactive features. The interactive features are then downscaled and decoded to obtain downscaled meteorological data.

[0008] In one embodiment, the training steps of the deep learning model include:

[0009] Acquire terrain data samples, meteorological data samples, and meteorological data labels;

[0010] Topographic and meteorological data samples are input into the initial model. The initial model encodes the topographic and meteorological data samples respectively to obtain topographic feature vector samples and meteorological feature vector samples. The topographic and meteorological feature vector samples are then interacted to obtain interactive feature samples. The interactive feature samples are then downscaled and decoded to obtain downscaled meteorological forecast data.

[0011] Based on downscaled meteorological forecast data and meteorological data labels, the data fitting loss is determined, and the physical constraint loss is determined based on the continuity equation constraint and the orthogonality constraint of the topographic wind field.

[0012] The sum of the data fitting loss and the physical constraint loss is used as the model loss. The parameters of the initial model are updated based on the model loss until a preset stopping condition is met, thus obtaining a pre-trained deep learning model.

[0013] In one embodiment, a deep learning model is used to encode features from terrain data and meteorological data respectively, resulting in terrain feature vectors and meteorological feature vectors, including:

[0014] The terrain data is feature-encoded using a terrain encoder based on a deep learning model to obtain a terrain feature vector.

[0015] A meteorological encoder based on a deep learning model performs fully connected processing on meteorological data to obtain meteorological feature vectors.

[0016] In one embodiment, feature interaction is performed on the terrain feature vector and the meteorological feature vector to obtain interactive features, including:

[0017] The attention module of the deep learning model determines the attention weights of the terrain feature vector and the meteorological feature vector respectively. The interaction features are obtained by weighting the terrain feature vector and the meteorological feature vector according to the attention weights.

[0018] In one embodiment, the interactive features are downscaled and decoded to obtain downscaled meteorological data, including:

[0019] Downscaled meteorological data are obtained by upsampling interactive features through the downscaling decoding module of a deep learning model.

[0020] In one embodiment, terrain data and meteorological data are input into a pre-trained deep learning model, including:

[0021] The terrain data and meteorological data are standardized separately, and the standardized terrain data and meteorological data are then input into the pre-trained deep learning model.

[0022] Secondly, this application also provides a multimodal meteorological data downscaling generation device, comprising:

[0023] The acquisition module is used to acquire topographic and meteorological data of the area to be measured.

[0024] The downscaling module is used to input terrain data and meteorological data into a pre-trained deep learning model. The deep learning model encodes the terrain data and meteorological data respectively to obtain terrain feature vectors and meteorological feature vectors. The terrain feature vectors and meteorological feature vectors are interacted to obtain interactive features. The interactive features are downscaled and decoded to obtain downscaled meteorological data.

[0025] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0026] Acquire topographic and meteorological data for the area to be measured;

[0027] Topographic and meteorological data are input into a pre-trained deep learning model. The deep learning model encodes the topographic and meteorological data separately to obtain topographic feature vectors and meteorological feature vectors. The topographic and meteorological feature vectors are then interacted to obtain interactive features. The interactive features are then downscaled and decoded to obtain downscaled meteorological data.

[0028] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the following steps:

[0029] Acquire topographic and meteorological data for the area to be measured;

[0030] Topographic and meteorological data are input into a pre-trained deep learning model. The deep learning model encodes the topographic and meteorological data separately to obtain topographic feature vectors and meteorological feature vectors. The topographic and meteorological feature vectors are then interacted to obtain interactive features. The interactive features are then downscaled and decoded to obtain downscaled meteorological data.

[0031] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, performs the following steps:

[0032] Acquire topographic and meteorological data for the area to be measured;

[0033] Topographic and meteorological data are input into a pre-trained deep learning model. The deep learning model encodes the topographic and meteorological data separately to obtain topographic feature vectors and meteorological feature vectors. The topographic and meteorological feature vectors are then interacted to obtain interactive features. The interactive features are then downscaled and decoded to obtain downscaled meteorological data.

[0034] The aforementioned method, apparatus, computer equipment, computer-readable storage medium, and computer program product for generating multimodal meteorological data downscaling acquire topographic and meteorological data from the area to be measured. The topographic and meteorological data are then input into a pre-trained deep learning model. The pre-trained deep learning model performs feature encoding and feature interaction on the topographic and meteorological data respectively, achieving the fusion of multimodal data. This fully explores the inherent correlation between topographic and meteorological data and accurately captures the modulating effect of topographic data on meteorological data. Downscaling and decoding through the deep learning model to obtain downscaled meteorological data improves the efficiency of meteorological data downscaling generation. The synergistic effect of feature interaction and downscaling decoding enhances the spatial resolution of meteorological data, better preserving large-scale patterns and local details of the meteorological field, which is beneficial for generating physically reasonable high-resolution meteorological data. Attached Figure Description

[0035] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0036] Figure 1 This is an application environment diagram of a multimodal meteorological data downscaling generation method in one embodiment;

[0037] Figure 2 This is a flowchart illustrating a method for generating multimodal meteorological data downscaling in one embodiment;

[0038] Figure 3 This is a schematic diagram of the overall process of a method for generating multimodal meteorological data downscaling in one embodiment;

[0039] Figure 4This is a structural block diagram of a multimodal meteorological data downscaling generation device in one embodiment;

[0040] Figure 5 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0042] It should be noted that the terms "first," "second," etc., used in this application can be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "comprising" and "having," and any variations thereof, used in this application, are intended to cover non-exclusive inclusion. The term "multiple" used in this application refers to two or more. The term "and / or" used in this application refers to one of the embodiments, or any combination of multiple embodiments.

[0043] The multimodal meteorological data downscaling generation method provided in this application embodiment can be applied to, for example... Figure 1The application environment shown is illustrated. Terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104 or placed on a cloud or other network server. This embodiment uses the method applied to server 104 as an example; it is understood that this method can also be applied to terminal 102, and can also be applied to systems including terminals and servers, and implemented through the interaction between the terminal and server. Server 104 acquires terrain data and meteorological data from the area to be measured; it inputs the terrain data and meteorological data into a pre-trained deep learning model, and uses the deep learning model to perform feature encoding on the terrain data and meteorological data respectively, obtaining terrain feature vectors and meteorological feature vectors. It then performs feature interaction on the terrain feature vectors and meteorological feature vectors to obtain interactive features, and performs downscaling decoding on the interactive features to obtain downscaled meteorological data. Terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, drones, low-altitude aircraft, IoT devices, and portable wearable devices. IoT devices can be smart speakers, smart TVs, smart air conditioners, smart vehicle devices, projection devices, etc. Portable wearable devices can include smartwatches, smart bracelets, and head-mounted displays. Head-mounted displays can be virtual reality (VR) devices, augmented reality (AR) devices, smart glasses, etc. Server 104 can be a standalone physical server, a server cluster or distributed system consisting of multiple physical servers, or a cloud server providing cloud computing services.

[0044] In one exemplary embodiment, such as Figure 2 As shown, a method for downscaling multimodal meteorological data is provided, which can be applied to... Figure 1 Taking server 104 as an example, the explanation includes the following steps 202 to 204. Wherein:

[0045] Step 202: Obtain topographic and meteorological data for the area to be measured.

[0046] The area to be measured refers to the region where meteorological data needs to be spatially downscaled, which can be complex terrain areas such as mountainous areas and coastal zones. The area to be measured can be flexibly set according to the actual application scenario.

[0047] Topographic data refers to data that describes the surface morphology of the area to be measured. For example, it can be a digital elevation model (DEM), generated by satellite remote sensing or geographic information system, which includes topographic features such as elevation and slope.

[0048] Meteorological data refers to meteorological observation or simulation data of the area to be measured. For example, it can be global meteorological reanalysis data, output by meteorological observation stations or climate models, which includes large-scale circulation information such as temperature, wind speed, and air pressure.

[0049] In some embodiments, the server acquires high-resolution terrain data and low-resolution meteorological data from the area to be measured, where the resolution of the terrain data is higher than that of the meteorological data. For example, the resolution of the terrain data is 1 km × 1 km, and the resolution of the meteorological data is 50 km × 50 km.

[0050] Step 204: Input the terrain data and meteorological data into the pre-trained deep learning model. The deep learning model encodes the terrain data and meteorological data respectively to obtain terrain feature vectors and meteorological feature vectors. The terrain feature vectors and meteorological feature vectors are interacted to obtain interactive features. The interactive features are downscaled and decoded to obtain downscaled meteorological data.

[0051] Among them, the pre-trained deep learning model refers to a neural network model trained with a large number of terrain data samples, meteorological data samples and corresponding high-resolution observation data labels. It has learned the mapping rules between low-resolution meteorological data, high-resolution terrain data and high-resolution meteorological data, and can be directly used for downscaling generation of meteorological data.

[0052] The pre-trained deep learning model can be a hybrid architecture of Transformer and CNN, including at least an encoder, an attention module, and a downscaling decoding module. Topographic and meteorological data are input into the pre-trained deep learning model. The encoder of the deep learning model encodes the features of the topographic and meteorological data respectively, extracting key features to obtain topographic feature vectors and meteorological feature vectors, thus transforming the raw data into more expressive feature vectors. For example, key features of topographic data include slope gradient, while key features of meteorological data include the vortex structure of the meteorological field.

[0053] The topographic feature vector and meteorological feature vector are the output results of feature encoding, which are high-dimensional representations of the original data after feature extraction. The topographic feature vector focuses on the microstructure of the terrain (such as the spatial distribution of slope and aspect), while the meteorological feature vector focuses on the macroscopic laws of the meteorological field (such as atmospheric circulation and pressure systems).

[0054] Feature interaction refers to the process of cross-modal fusion of terrain feature vectors and meteorological feature vectors. Interactive features are high-dimensional features that integrate micro-topographic features and macro-meteorological laws, preserving both the local modulation effect of terrain on meteorology (such as terrain following) and the overall consistency of the meteorological field (such as atmospheric circulation trends).

[0055] Downscaling decoding refers to mapping interactive features from a high-dimensional feature space back to the actual meteorological variable space through a decoder, generating high-resolution meteorological data with the same resolution as the terrain data, while ensuring the rationality of the generated downscaling meteorological data through physical constraints (such as mass conservation).

[0056] Downscaled meteorological data has a higher resolution than the meteorological data before model processing. Through the above process, end-to-end generation from raw data to high-resolution meteorological fields is achieved, balancing efficiency and physical plausibility, and providing support for scenarios such as mountain microclimates and extreme weather forecasting.

[0057] In the aforementioned method for generating multimodal meteorological data downscaling, topographic and meteorological data from the area to be measured are acquired and input into a pre-trained deep learning model. The pre-trained deep learning model performs feature encoding and feature interaction on the topographic and meteorological data respectively, realizing the fusion of multimodal data. This method can fully explore the inherent correlation between topographic and meteorological data and accurately capture the modulating effect of topographic data on meteorological data. Downscaling and decoding through the deep learning model to obtain downscaled meteorological data can improve the efficiency of meteorological data downscaling generation. The synergistic effect of feature interaction and downscaling decoding improves the spatial resolution of meteorological data, better preserves the large-scale regularity and local details of the meteorological field, and is conducive to generating physically reasonable high-resolution meteorological data.

[0058] In an exemplary embodiment, the training steps of the deep learning model include: acquiring terrain data samples, meteorological data samples, and meteorological data labels; inputting the terrain data samples and meteorological data samples into an initial model, performing feature encoding on the terrain data samples and meteorological data samples respectively through the initial model to obtain terrain feature vector samples and meteorological feature vector samples, performing feature interaction on the terrain feature vector samples and meteorological feature vector samples to obtain interactive feature samples, performing downscaling decoding on the interactive feature samples to obtain downscaled meteorological prediction data; determining the data fitting loss based on the downscaled meteorological prediction data and meteorological data labels, determining the physical constraint loss based on the continuity equation constraint and the orthogonality constraint of the terrain wind field; using the sum of the data fitting loss and the physical constraint loss as the model loss, updating the parameters of the initial model according to the model loss until a preset stopping condition is met, to obtain a pre-trained deep learning model.

[0059] Among them, the terrain data sample, meteorological data sample, and meteorological data label are derived from historical data of the preset area. The terrain data sample can be high-resolution terrain data of the preset area, the meteorological data sample can be low-resolution meteorological data of the preset area, and the meteorological data label refers to the real high-resolution meteorological data of the preset area.

[0060] The initial model refers to the deep learning model before training, which has the same model structure but different model parameters as the trained deep learning model. For example, the initial model can be a hybrid architecture of Transformer and CNN. The encoder of the initial model encodes features from terrain data samples and meteorological data samples respectively, obtaining terrain feature vector samples and meteorological feature vector samples. Then, the attention module performs feature interaction between the terrain feature vector samples and meteorological feature vector samples. Finally, the downscaling decoding module performs downscaling decoding on the interactive features to obtain downscaled meteorological prediction data.

[0061] Downscaled weather forecast data represents predictions made before model optimization and may deviate significantly from the labels. Based on the differences between the downscaled weather forecast data and the weather data labels, a pre-defined loss function is used to calculate the data fitting loss. For example, the pre-defined loss function could be the cross-entropy function.

[0062] Continuity equation constraints refer to the constraints imposed on the physical rationality of predicted wind fields based on the law of conservation of mass in atmospheric dynamics. Topographic wind field orthogonal constraints refer to the constraints imposed on the components of the wind field perpendicular to contour lines, ensuring they conform to topographical characteristics (e.g., airflow ascends / descends along the slope), thus avoiding unreasonable physical results such as wind passing through mountains.

[0063] For example, wind field divergence , For example, div=0.3. Here, u and v represent the horizontal and vertical wind speed components, respectively. The continuity equation constraints can be:

[0064]

[0065] The orthogonal constraints of topographic wind field can be:

[0066]

[0067] in, Indicates terrain gradient, This reflects the rate and direction of change of the terrain; Obtained through dot product The dot product measures the alignment between the terrain gradient direction and the wind velocity direction. If the dot product is close to 0, it indicates that the wind direction is close to the contour lines (because the terrain gradient is perpendicular to the contour lines, the two vectors are perpendicular when the dot product is 0), meeting the terrain's constraint requirements on the wind field. If the dot product is greater than a preset value, it indicates that the wind field has a significant component perpendicular to the contour lines, which does not conform to the wind field characteristics under the influence of actual terrain. γ is a penalty coefficient used to control the strength of this terrain-wind field orthogonal constraint. The larger γ is, the heavier the penalty is for wind fields that do not conform to the contour line flow during model training, and the more it forces the wind field to conform to the terrain constraint. If γ is too small, the constraint effect may be insignificant, and the wind field may exhibit vertical contour line flow that does not conform to the physical laws of the terrain.

[0068] The physical constraint loss is obtained by adding the continuity equation constraint and the topographic wind field orthogonal constraint. The model loss includes both the data fitting loss and the physical constraint loss. During model training, optimization techniques such as backpropagation are used to minimize the model loss, allowing the wind field simulation results to better conform to the topographic constraints and improving the physical rationality of the results in applications such as meteorological downscaling.

[0069] In this embodiment, by constraining the wind field to flow along contour lines, the downscaled meteorological data conforms to the physical laws of how terrain influences atmospheric motion, avoiding situations where the wind field passes through mountains, which violates physical rationality. This makes the high-resolution meteorological data more credible in terms of physical mechanisms. For example, in mountain microclimate simulation, it can accurately present the guiding and blocking effects of terrain on airflow. The constructed orthogonal constraint of the terrain wind field can help the deep learning model learn a more realistic coupling relationship between terrain and meteorology during training. During the model parameter update process, the physical constraint loss and the data fitting loss work together to ensure that the model output matches the observation data and follows physical laws, thereby improving the model's adaptability to complex terrain and meteorological scenarios.

[0070] In an exemplary embodiment, terrain data and meteorological data are feature-encoded using a deep learning model to obtain terrain feature vectors and meteorological feature vectors, respectively. This includes: using a terrain encoder of a deep learning model to encode the terrain data to obtain a terrain feature vector; and using a meteorological encoder of a deep learning model to perform fully connected processing on the meteorological data to obtain a meteorological feature vector.

[0071] The encoders in the deep learning model include a terrain encoder and a weather encoder. The terrain encoder processes terrain data, extracts key features from the terrain data, and outputs a terrain feature vector. For example, the terrain encoder can be a convolutional neural network structure.

[0072] A meteorological encoder is used to process meteorological data, extract key features from the data, and output a meteorological feature vector. For example, a meteorological encoder can be a fully connected layer that maps meteorological data to a high-dimensional space using a weight matrix, then introduces a nonlinear transformation through an activation function to output the meteorological feature vector.

[0073] In this embodiment, the terrain encoder performs feature encoding on the terrain data, which can accurately extract key physical features of the terrain (such as slope and aspect) and avoid redundant information in the original terrain data; the meteorological encoder performs fully connected processing on the meteorological data, which can capture the complex correlation between meteorological variables and improve the feature expression capability; by converting data of different modalities into feature vectors of the same dimension, it is beneficial to subsequent cross-modal feature fusion.

[0074] In an exemplary embodiment, feature interaction is performed on terrain feature vectors and meteorological feature vectors to obtain interactive features. This includes: determining the attention weights of the terrain feature vectors and meteorological feature vectors respectively through the attention module of the deep learning model; and weighting the terrain feature vectors and meteorological feature vectors according to the attention weights to obtain the interactive features. The attention module is the core component in the deep learning model for implementing feature interaction. The correlation strength between different modal features is calculated based on an attention mechanism (such as multi-head attention), allowing the model to automatically focus on key features that significantly affect the results.

[0075] Attention weights represent the degree of attention a specific location in one modality pays to a corresponding location in another modality. Higher attention weights indicate a stronger correlation between the two features; for example, a grid with high wind speeds will have a higher attention weight to a steep slope.

[0076] For example, a multi-head attention mechanism is used to construct an attention module. The inputs are terrain feature vectors and meteorological feature vectors. The terrain feature vectors are linearly transformed using multiple sets of different linear projection matrices to obtain a query matrix Qt, a key matrix Kt, and a value matrix Vt. Similarly, the meteorological feature vectors are linearly transformed using multiple sets of different linear projection matrices to obtain a query matrix Qm, a key matrix Km, and a value matrix Vm. For each head, an attention weight At is calculated based on the query matrix Qt and the key matrix Km. This attention weight At is then multiplied by the value matrix Vm to obtain the weighted value Zt for the current head. Finally, an attention weight Am is calculated based on the query matrix Qm and the key matrix Kt. This attention weight Am is then multiplied by the value matrix Vt to obtain the weighted value Zm for the current head. The formula for calculating the attention weights is as follows:

[0077]

[0078] The attention weights of multiple heads are concatenated and linearly transformed to obtain the interaction features.

[0079] In this embodiment, attention weights are obtained by performing attention processing on terrain data and meteorological data, thereby achieving deep integration of terrain and meteorological features. This ensures that the interactive features conform to both large-scale meteorological laws and local terrain influences, improving physical rationality. The modulation of meteorological data by terrain data is preserved, while the parallel computation of the attention mechanism improves processing efficiency, balancing accuracy and speed.

[0080] In one exemplary embodiment, downscaling the interactive features to obtain downscaled meteorological data includes: upsampling the interactive features using a downscaling decoding module of a deep learning model to obtain downscaled meteorological data.

[0081] The downscaling decoding module refers to the sub-network in the deep learning model responsible for transforming high-dimensional interactive features into high-resolution meteorological data. Its core function is to improve spatial resolution through upsampling and map the features back to the physical space of actual meteorological variables (such as wind speed and temperature).

[0082] For example, a downscaling decoding module is constructed using multiple transposed convolutional layers and 1×1 convolutional layers. The input is interactive features (1km resolution, shape [1000, 1000, 512]). The first transposed convolutional layer (kernel=4×4, stride=2) improves the feature resolution to 500m (intermediate transition). The second transposed convolutional layer further improves it to 250m. The third transposed convolutional layer finally outputs features with a resolution of 100m. Finally, the 512-dimensional features are mapped to three meteorological variables (wind speed u, v, and temperature) through 1×1 convolution.

[0083] In this embodiment, the interactive features are upsampled by the downscaling decoding module to gradually improve the spatial resolution, thereby upscaling low-resolution meteorological data (e.g., 50km) to a high resolution (e.g., 1km) that matches the terrain, capturing the subtle influence of micro-topography on meteorology; ensuring that the generated high-resolution meteorological data retains the micro-details of the coupling between terrain and meteorology.

[0084] In one exemplary embodiment, inputting terrain data and meteorological data into a pre-trained deep learning model includes: standardizing the terrain data and meteorological data respectively, and inputting the standardized terrain data and standardized meteorological data into the pre-trained deep learning model.

[0085] Standardization refers to scaling or offsetting terrain and meteorological data respectively, transforming features of different magnitudes and units to a unified numerical range (such as a normal distribution with a mean of 0 and a standard deviation of 1, or the [0,1] interval), thereby eliminating the impact of dimensional differences on model training and inference.

[0086] Standardized terrain data consists of standardized terrain features (such as elevation, slope, and aspect), preserving the spatial distribution patterns and relative differences of the original terrain, but with a uniform numerical range, making it more suitable for deep learning models. Standardized meteorological data consists of standardized meteorological variables (such as wind speed, temperature, and air pressure), preserving the spatial gradient and temporal trends of the meteorological field, but with a uniform numerical range, facilitating its integration with standardized terrain data into the model.

[0087] The standardized data has the same distribution as the data during model pre-training, avoiding a decrease in prediction accuracy due to distribution shift.

[0088] In this embodiment, by standardizing the terrain data and meteorological data respectively, the distribution of standardized data is stable, the interference of extreme values ​​is reduced, and the model training is more efficient and the inference is more stable. The difference in units is eliminated, and the terrain and meteorological features interact in the same numerical space, ensuring the effectiveness of the attention mechanism and feature fusion, and avoiding misjudgment of association.

[0089] To illustrate the multimodal meteorological data downscaling generation method and its effects in this scheme in detail, a specific embodiment is described below:

[0090] This application proposes a high-resolution downscaling generation system for multimodal meteorological data based on the Transformer architecture and physical constraints. The system aims to generate physically plausible high-resolution meteorological fields by fusing high-resolution topographic data with low-resolution meteorological data and incorporating atmospheric dynamics principles. The system includes a preprocessing module, an encoder module, a cross-modal fusion module, a physical constraint module, and a downscaling decoding module, covering core processes such as data input, feature encoding, cross-modal interaction, physical constraint embedding, and downscaling decoding.

[0091] like Figure 3 The diagram illustrates the overall workflow of multimodal meteorological data downscaling generation methods in some embodiments. Meteorological and topographic data are preprocessed and encoded separately. The processed data undergoes feature interaction through a cross-modal fusion module. A downscaling decoding module upsamples the interactive features. Physical constraints are introduced into the loss function to train a deep learning model. The trained deep learning model outputs high-resolution meteorological data.

[0092] 1. Data flow:

[0093] Input data generation: High-resolution topographic maps (such as DEM digital elevation models): generated by satellite remote sensing or geographic information systems, containing topographic features such as elevation and slope. Low-resolution meteorological data (such as global meteorological reanalysis data): output by meteorological observation stations or climate models, containing large-scale circulation information such as temperature, wind speed, and air pressure.

[0094] 2. Intermediate processing:

[0095] Data encoding: Topographic and meteorological data are encoded into feature vectors respectively; Cross-modal attention fusion: Multi-head attention mechanism is used to achieve interaction between topographic and meteorological features; Physical constraint embedding: Atmospheric dynamics equations and topographic-wind field orthogonal constraints are incorporated into model training.

[0096] 3. Final output:

[0097] Downscaled meteorological datasets: contain high-resolution (e.g., 1km grid) meteorological variables such as wind speed and temperature, suitable for regional climate simulation or extreme weather prediction.

[0098] Execution Entities: Server-side (cloud or local cluster): Responsible for data encoding, cross-modal fusion, physical constraint calculation, and model training. Client-side (meteorological research institution terminal): Provides the user interface (UI) for parameter configuration, model invocation, and result visualization.

[0099] 4. Execution sequence

[0100] Data preprocessing: Topographic data is normalized (0-1 normalization), and meteorological data is interpolated to a unified grid coordinate system. Encoding stage: Topographic encoder: Uses a convolutional neural network (CNN) to extract multi-scale topographic features. Meteorological encoder: Maps meteorological data to a high-dimensional space through fully connected layers. Cross-modal attention fusion: After feature concatenation, the data is input into the Transformer module to calculate the multi-head attention weights between topographic and meteorological data.

[0101] 5. Embedding of physical constraints: Introduce continuous equation constraints and orthogonal constraints of terrain and wind field into the loss function.

[0102] 6. Downscaling Decoding: Use transposed convolutional layers to progressively upsample and generate a high-resolution meteorological field.

[0103] 7. Application Scenarios

[0104] a. Regional Climate Simulation: Scenario Description: Refined wind field prediction for complex terrain regions such as mountainous areas or coastal zones. Technology Adaptation: The terrain encoder focuses on extracting features of canyons and peaks, and cross-modal attention captures the effects of terrain on wind speed enhancement or obstruction.

[0105] b. Extreme Weather Warning: Scenario Description: Localized enhancement prediction of typhoon paths or heavy precipitation events. Technical Adaptation: Physical constraints ensure that the wind field generated by the model conforms to atmospheric dynamics, avoiding spurious vortices.

[0106] The Transformer architecture captures long-range dependencies through a self-attention mechanism, adapting to multimodal data interactions. The model structure includes: an encoder layer where terrain and meteorological data are encoded by independent encoders; and an attention module that calculates the weights of terrain features on meteorological features (e.g., the weight of wind speed in canyon areas).

[0107] The physical constraint formulas include continuity equation constraints and topographic wind field orthogonal constraints.

[0108] Continuity equation constraints:

[0109]

[0110] For the horizontal wind speed component, To constrain the weighting coefficients, the wind field divergence is forced to approach zero.

[0111] Topographic wind field orthogonal constraints:

[0112]

[0113] ∇h represents the topographic gradient, constraining surface winds to flow along contour lines, and γ is the penalty coefficient.

[0114] Transposed convolution upsampling: This method progressively enlarges the feature map size through deconvolution operations to restore high-resolution details. The steps include inputting a low-resolution feature map, zero-padding expansion, kernel sliding computation, and outputting a high-resolution feature map.

[0115] 8. Interface Design (Client-Side)

[0116] Parameter configuration interface: Input: Select terrain file (GeoTIFF format), meteorological data (NetCDF format), target resolution (e.g., 1km). Controls: Slider adjusts physical constraint weights (...). , Visualization Interface: Display: High-resolution wind field vector map overlaid with terrain contour lines, supporting time-series animation playback. Interaction: Click on an area to view numerical details of variables such as wind speed and temperature.

[0117] Through modular design and physics-guided deep learning, the generated high-resolution meteorological data is ensured to have both numerical accuracy and physical plausibility, providing a reliable tool for meteorological research.

[0118] The aforementioned method for downscaling multimodal meteorological data acquires topographic and meteorological data from the target area. This data is then input into a pre-trained deep learning model, which performs feature encoding and feature interaction on both data, achieving multimodal data fusion. This method fully explores the inherent correlation between topographic and meteorological data, accurately capturing the modulating effect of topographic data on meteorological data. Downscaling and decoding via the deep learning model improves the efficiency of meteorological data downscaling. The synergistic effect of feature interaction and downscaling decoding enhances the spatial resolution of meteorological data, better preserving large-scale patterns and local details of the meteorological field, thus facilitating the generation of physically plausible high-resolution meteorological data. Furthermore, by dynamically allocating the correlation between topographic and meteorological data through multi-head attention weights, the method captures, for example, the wind speed-lifting effect of canyons and the airflow-blocking effect of mountains. This enables accurate modeling of wind field changes at different altitudes, avoiding local errors caused by single-scale features in traditional methods. Continuity equation constraints ensure that the generated horizontal wind field divergence approaches zero, eliminating false airflow convergence or divergence. The orthogonal constraint of the topographic wind field forces surface winds to flow along contour lines, avoiding the generation of unreasonable wind fields that traverse mountains. The generated high-resolution meteorological field strictly conforms to the laws of atmospheric dynamics, making it suitable for scientific research and operational forecasting. The transposed convolutional upsampling decoder downscales low-resolution meteorological data (e.g., 50km grid) to high-resolution (e.g., 1km grid), restoring topographically driven local details (e.g., coastal wind speed gradient, urban heat island effect). Compared to traditional bilinear interpolation, transposed convolution preserves the nonlinear relationships in the feature map through learnable parameters, avoiding ambiguity or distortion. The modular design (encoder-fusion-decoder) is adapted to physical constraints, ensuring that the generated wind field accurately reflects local phenomena such as topographic lifting and funneling effects in complex terrain areas such as mountainous regions and coastal zones. Physical constraints ensure that the predicted results of key events such as typhoon paths and rainstorm centers conform to the actual atmospheric evolution laws, reducing the risk of false alarms. By integrating deep learning with physical laws, this method not only overcomes the limitations and physical inconsistencies of traditional meteorological downscaling methods in data fusion, but also significantly improves the generation accuracy and practicality of high-resolution meteorological fields, providing reliable technical support for climate research, disaster early warning, and other scenarios.

[0119] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps. It is understood that the steps in different embodiments can be freely combined as needed, and all non-contradictory solutions formed by such combinations are within the scope of protection of this application.

[0120] Based on the same inventive concept, this application also provides a multimodal meteorological data downscaling generation apparatus for implementing the multimodal meteorological data downscaling generation method described above. The solution provided by this apparatus is similar to the implementation scheme described in the above method. Therefore, the specific limitations in one or more embodiments of the multimodal meteorological data downscaling generation apparatus provided below can be found in the limitations of the multimodal meteorological data downscaling generation method described above, and will not be repeated here.

[0121] In one exemplary embodiment, such as Figure 4 As shown, a multimodal meteorological data downscaling generation device 400 is provided, including: an acquisition module 420 and a downscaling module 440, wherein:

[0122] The acquisition module 420 is used to acquire topographic and meteorological data of the area to be measured.

[0123] The downscaling module 440 is used to input terrain data and meteorological data into a pre-trained deep learning model. The deep learning model encodes the terrain data and meteorological data respectively to obtain terrain feature vectors and meteorological feature vectors. The terrain feature vectors and meteorological feature vectors are interacted to obtain interactive features. The interactive features are downscaled and decoded to obtain downscaled meteorological data.

[0124] The aforementioned multimodal meteorological data downscaling generation device acquires topographic and meteorological data from the area to be measured, inputs the topographic and meteorological data into a pre-trained deep learning model, and performs feature encoding and feature interaction on the topographic and meteorological data respectively through the pre-trained deep learning model, realizing the fusion of multimodal data. It can fully explore the inherent correlation between topographic and meteorological data and accurately capture the modulating effect of topographic data on meteorological data. Downscaling and decoding through the deep learning model to obtain downscaled meteorological data can improve the efficiency of meteorological data downscaling generation. The synergistic effect of feature interaction and downscaling decoding improves the spatial resolution of meteorological data, can better preserve the large-scale regularity and local details of the meteorological field, and is conducive to generating physically reasonable high-resolution meteorological data.

[0125] In one embodiment, the multimodal meteorological data downscaling generation device 400 further includes a training module for training the deep learning model. The training module is used to: acquire terrain data samples, meteorological data samples, and meteorological data labels; input the terrain data samples and meteorological data samples into an initial model; encode the features of the terrain data samples and meteorological data samples using the initial model to obtain terrain feature vector samples and meteorological feature vector samples; perform feature interaction on the terrain feature vector samples and meteorological feature vector samples to obtain interactive feature samples; perform downscaling decoding on the interactive feature samples to obtain downscaled meteorological prediction data; determine the data fitting loss based on the downscaled meteorological prediction data and meteorological data labels; determine the physical constraint loss based on the continuity equation constraint and the orthogonality constraint of the terrain wind field; use the sum of the data fitting loss and the physical constraint loss as the model loss; update the parameters of the initial model based on the model loss until a preset stopping condition is met to obtain a pre-trained deep learning model.

[0126] In one embodiment, the terrain data and meteorological data are feature-encoded by a deep learning model to obtain terrain feature vectors and meteorological feature vectors respectively. The downscaling module 440 is also used to: encode the terrain data by a terrain encoder of a deep learning model to obtain a terrain feature vector; and perform fully connected processing on the meteorological data by a meteorological encoder of a deep learning model to obtain a meteorological feature vector.

[0127] In one embodiment, the terrain feature vector and the meteorological feature vector are subjected to feature interaction to obtain interactive features. The downscaling module 440 is further used to: determine the attention weights of the terrain feature vector and the meteorological feature vector respectively through the attention module of the deep learning model, and weight the terrain feature vector and the meteorological feature vector according to the attention weights to obtain interactive features.

[0128] In one embodiment, the interaction features are downscaled and decoded to obtain downscaled meteorological data. The downscaling module 440 is further configured to: upsample the interaction features through the downscaling decoding module of the deep learning model to obtain downscaled meteorological data.

[0129] In one embodiment, terrain data and meteorological data are input into a pre-trained deep learning model. The downscaling module 440 is further configured to: standardize the terrain data and meteorological data respectively, and input the standardized terrain data and standardized meteorological data into the pre-trained deep learning model.

[0130] Each module in the aforementioned multimodal meteorological data downscaling generation device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of a computer device in hardware form or independent of it, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0131] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 5 As shown, the computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs in the non-volatile storage media to run. The database stores topographic data, meteorological data, and downscaled meteorological data. The I / O interfaces are used for information exchange between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for generating multimodal meteorological data downscaling.

[0132] Those skilled in the art will understand that Figure 5 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0133] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0134] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps in the above method embodiments.

[0135] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0136] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0137] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0138] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0139] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for downscaling multimodal meteorological data, characterized in that, The method includes: Acquire topographic and meteorological data for the area to be measured; The terrain data and meteorological data are input into a pre-trained deep learning model. The deep learning model encodes the terrain data and meteorological data respectively to obtain terrain feature vectors and meteorological feature vectors. The terrain feature vectors and meteorological feature vectors are interacted to obtain interactive features. The interactive features are downscaled and decoded to obtain downscaled meteorological data.

2. The method according to claim 1, characterized in that, The training steps of the deep learning model include: Acquire terrain data samples, meteorological data samples, and meteorological data labels; The terrain data sample and the meteorological data sample are input into the initial model. The initial model encodes the terrain data sample and the meteorological data sample respectively to obtain terrain feature vector sample and meteorological feature vector sample. The terrain feature vector sample and the meteorological feature vector sample are interacted to obtain interactive feature sample. The interactive feature sample is downscaled and decoded to obtain downscaled meteorological forecast data. Based on the downscaled meteorological forecast data and the meteorological data labels, the data fitting loss is determined, and the physical constraint loss is determined based on the continuity equation constraint and the orthogonality constraint of the topographic wind field. The sum of the data fitting loss and the physical constraint loss is used as the model loss. The parameters of the initial model are updated according to the model loss until a preset stopping condition is met, thus obtaining a pre-trained deep learning model.

3. The method according to claim 1, characterized in that, The step of using the deep learning model to encode features of the terrain data and the meteorological data to obtain terrain feature vectors and meteorological feature vectors includes: The terrain data is feature-encoded using the terrain encoder of the deep learning model to obtain a terrain feature vector; The meteorological data is processed by a fully connected method using the meteorological encoder of the deep learning model to obtain meteorological feature vectors.

4. The method according to claim 1, characterized in that, The step of performing feature interaction on the terrain feature vector and the meteorological feature vector to obtain interactive features includes: The attention module of the deep learning model determines the attention weights of the terrain feature vector and the meteorological feature vector respectively, and the terrain feature vector and the meteorological feature vector are weighted according to the attention weights to obtain the interaction features.

5. The method according to claim 1, characterized in that, The downscaling decoding of the interactive features to obtain downscaled meteorological data includes: The interactive features are upsampled using the downscaling decoding module of the deep learning model to obtain downscaled meteorological data.

6. The method according to any one of claims 1 to 5, characterized in that, The step of inputting the terrain data and meteorological data into a pre-trained deep learning model includes: The terrain data and meteorological data are standardized respectively, and the standardized terrain data and standardized meteorological data are input into the pre-trained deep learning model.

7. A device for downscaling multimodal meteorological data, characterized in that, The device includes: The acquisition module is used to acquire topographic and meteorological data of the area to be measured. The downscaling module is used to input the terrain data and meteorological data into a pre-trained deep learning model, and to encode the terrain data and meteorological data into feature vectors and meteorological feature vectors respectively through the deep learning model. The terrain feature vectors and meteorological feature vectors are then interacted to obtain interactive features. Finally, the interactive features are downscaled and decoded to obtain downscaled meteorological data.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images