Meteorological Data Downscaling Method and Device Based on Spatiotemporal Graph Neural Network

Abstract

This application discloses a method and apparatus for downscaling meteorological data based on a spatiotemporal graph neural network. The method includes: acquiring meteorological observation datasets and gridded datasets from meteorological stations within a study area; preprocessing the datasets and determining corresponding feature variables using a pre-defined correlation analysis algorithm; constructing a dataset for downscaling based on the meteorological observation datasets, gridded datasets, and corresponding feature variables; inputting the dataset for downscaling into a trained first spatiotemporal graph neural network to downscale the low-resolution meteorological grid data, obtaining predicted high-resolution meteorological grid data; and using a trained second spatiotemporal graph neural network to correct the bias in the predicted high-resolution meteorological grid data, obtaining corrected high-resolution meteorological grid data. This application can capture the details of local meteorological phenomena more quickly and accurately, improving the resolution and accuracy of the downscaled meteorological data.

Description

Technical Field

[0001] This application relates to the field of meteorological data downscaling technology, and more specifically, to a meteorological data downscaling method and apparatus based on a spatiotemporal graph neural network. Background Technology

[0002] In meteorological research, high-resolution meteorological data is crucial for weather forecasting, climate change analysis, and agricultural production. Existing meteorological data typically originates from meteorological station observations, satellite remote sensing, and reanalysis products, resulting in relatively low spatial resolution, usually ranging from tens to hundreds of kilometers. To meet the needs of regional and local research, downscaling techniques have become an important means of improving the spatial resolution of meteorological data.

[0003] Currently, meteorological data downscaling mainly includes three types of methods: dynamic downscaling, statistical downscaling, and coupled downscaling. Dynamic downscaling captures the physical processes of meteorological variables through numerical simulation to generate high-resolution meteorological data, but it suffers from high computational complexity, is time-consuming, and is easily affected by input data and numerical models. Statistical downscaling establishes a mapping relationship between high-resolution and low-resolution meteorological data through statistical models, offering low computational cost and ease of application, but it relies on long-term observational data with high spatiotemporal resolution. Coupled downscaling combines the advantages of dynamic and statistical downscaling methods, providing high-precision and physically consistent results, but still faces problems such as computational complexity and data dependence, resulting in limited application.

[0004] In statistical downscaling, convolutional neural networks (CNNs), recurrent neural networks (RNNs), and generative adversarial networks (GANs) have been successfully applied to image super-resolution tasks and are gradually being introduced into the field of meteorological data downscaling. However, these current network models still have shortcomings in the application of meteorological data downscaling, making it difficult to accurately capture the details of local meteorological phenomena, thus affecting the resolution and accuracy of the downscaled meteorological data. Summary of the Invention

[0005] The purpose of this application is to provide a method, apparatus, electronic device and storage medium for meteorological data downscaling based on spatiotemporal graph neural networks, so as to more accurately capture the details of local meteorological phenomena and improve the resolution and accuracy of downscaled meteorological data.

[0006] To achieve the above objectives, firstly, this application provides a meteorological data downscaling method based on a spatiotemporal graph neural network, comprising:

[0007] Meteorological observation datasets and gridded datasets from meteorological stations within the study area are obtained. The gridded datasets include low-resolution meteorological grid data, high-resolution meteorological grid data, and high-resolution static land grid data.

[0008] The meteorological observation dataset and the gridded dataset are preprocessed, and the corresponding feature variables are determined using a preset correlation analysis algorithm.

[0009] Based on the meteorological observation dataset, the gridded dataset, and the corresponding feature variables, a dataset for downscaling is constructed.

[0010] The dataset used for downscaling is input into a trained first spatiotemporal graph neural network, which is then used to downscale the low-resolution meteorological grid data to obtain predicted high-resolution meteorological grid data.

[0011] In a preferred embodiment of this application, the acquisition of meteorological observation datasets and gridded datasets from meteorological stations within the study area includes:

[0012] Acquire meteorological observation datasets from meteorological stations within the study area, as well as low-resolution meteorological data and high-resolution land static data;

[0013] The study area is divided into grids to obtain the corresponding grid.

[0014] Based on the grid division, the low-resolution meteorological data, and the high-resolution land static data, low-resolution meteorological grid data, high-resolution land static grid data, and high-resolution meteorological grid data are obtained, and the low-resolution meteorological grid data, the high-resolution meteorological grid data, and the high-resolution land static grid data are used as a gridded dataset.

[0015] In a preferred embodiment of this application, the step of obtaining low-resolution meteorological grid data, high-resolution land static grid data, and high-resolution meteorological grid data based on the grid division, the low-resolution meteorological data, and the high-resolution land static data includes:

[0016] The low-resolution meteorological data is interpolated onto the divided grid to obtain low-resolution meteorological grid data;

[0017] Based on the grid division and the high-resolution static land data, high-resolution static land grid data is calculated; the high-resolution static land grid data and the low-resolution meteorological grid data are input into a preset meteorological research and forecasting modeling system to obtain high-resolution meteorological grid data.

[0018] In a preferred embodiment of this application, the step of constructing a dataset for downscaling based on the meteorological observation dataset, the gridded dataset, and the corresponding feature variables includes:

[0019] Based on the determined number of lag steps, obtain the time lag variable corresponding to the current time;

[0020] Based on the meteorological observation dataset, the gridded dataset, the corresponding feature variables, and the corresponding time lag variables, a lag time feature dataset is constructed as a dataset for downscaling.

[0021] In a preferred embodiment of this application, the first spatiotemporal graph neural network includes an embedding layer, a graph structure construction module, a spatiotemporal convolution module, an attention module, and an output layer.

[0022] The embedding layer of the first spatiotemporal graph neural network has the functions of spatiotemporal encoding, average encoding and feature embedding, and the output layer of the first spatiotemporal graph neural network includes two-dimensional convolutional layers and fully connected layers in the time and space dimensions.

[0023] In a preferred embodiment of this application, after inputting the dataset for downscaling into a trained first spatiotemporal graph neural network, and using the first spatiotemporal graph neural network to downscale the low-resolution meteorological grid data to obtain predicted high-resolution meteorological grid data, the method further includes:

[0024] The dataset used for downscaling and the predicted high-resolution meteorological grid data are input into a trained second spatiotemporal graph neural network. The second spatiotemporal graph neural network is then used to correct the bias in the predicted high-resolution meteorological grid data to obtain the corrected high-resolution meteorological grid data.

[0025] In a preferred embodiment of this application, the second spatiotemporal graph neural network includes a data processing module, a spatiotemporal representation learning module, an uncertainty modeling module, and an extrapolation generation module.

[0026] Secondly, this application provides a meteorological data downscaling device based on a spatiotemporal graph neural network, comprising: a data acquisition unit for acquiring meteorological observation datasets and gridded datasets of meteorological stations within a study area, wherein the gridded datasets include low-resolution meteorological grid data, high-resolution meteorological grid data, and high-resolution land static grid data; and a data preprocessing unit for performing data preprocessing on the meteorological observation datasets and the gridded datasets, and determining the corresponding feature variables using a preset correlation analysis algorithm.

[0027] The dataset construction unit is used to construct a dataset for downscaling based on the meteorological observation dataset, the gridded dataset, and the corresponding feature variables.

[0028] The data downscaling unit is used to input the dataset for downscaling into a trained first spatiotemporal graph neural network, and use the first spatiotemporal graph neural network to downscale the low-resolution meteorological grid data to obtain predicted high-resolution meteorological grid data.

[0029] Thirdly, this application provides an electronic device, including a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to perform the above-described meteorological data downscaling method based on spatiotemporal graph neural networks.

[0030] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described meteorological data downscaling method based on a spatiotemporal graph neural network.

[0031] This application proposes a meteorological data downscaling method, device, electronic device, and storage medium based on spatiotemporal graph neural networks. Compared with existing technologies, its advantages are as follows:

[0032] This application obtained meteorological observation datasets and gridded datasets from meteorological stations within the study area. The gridded datasets include low-resolution meteorological grid data, high-resolution meteorological grid data, and high-resolution static land grid data. A pre-defined correlation analysis algorithm was used to determine the corresponding feature variables, enhancing the diversity and richness of the data sources and covering different resolutions and data types, thus improving the prediction accuracy of the neural network model. Simultaneously, a trained first spatiotemporal graph neural network was used to downscale the low-resolution meteorological grid data, enabling faster generation of predicted high-resolution meteorological grid data and significantly improving efficiency. Furthermore, the first spatiotemporal graph neural network provides higher spatial and temporal resolution, enhancing the detail and accuracy of meteorological data, thereby more accurately capturing the details of local meteorological phenomena and improving the resolution and accuracy of the downscaled meteorological data. In addition, the first spatiotemporal graph neural network also has higher generalization ability and can be used to process different local areas, different time scales, or different meteorological types. Attached Figure Description

[0033] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following is a brief introduction to the drawings required for use in the embodiments of the present application. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other relevant drawings can be obtained based on these drawings without creative work.

[0034] Figure 1This is a flowchart illustrating the meteorological data downscaling method based on a spatiotemporal graph neural network provided in this application embodiment;

[0035] Figure 2 This is a schematic diagram of the structure of the first spatiotemporal graph neural network provided in the embodiments of this application;

[0036] Figure 3 This is a schematic diagram of the downscaling results and prediction errors of the air temperature at a height of 2m above the ground provided in the embodiments of this application;

[0037] Figure 4 This is a schematic diagram of the structure of the second spatiotemporal graph neural network provided in the embodiments of this application;

[0038] Figure 5 This is a structural block diagram of the meteorological data downscaling device based on a spatiotemporal graph neural network provided in this application embodiment;

[0039] Figure 6 This is a schematic diagram of the internal structure of the computer device provided in the embodiments of this application. Detailed Implementation

[0040] The specific embodiments of this application will be described in further detail below with reference to the accompanying drawings and examples. The following examples are used to illustrate this application, but are not intended to limit the scope of this application.

[0041] To address the problems in the prior art, this application provides a meteorological data downscaling method, apparatus, electronic device, and storage medium based on a spatiotemporal graph neural network. This method can provide higher spatial and temporal resolution, enhance the detail and accuracy of meteorological data, and thus more accurately capture the details of local meteorological phenomena, thereby improving the resolution and accuracy of downscaled meteorological data.

[0042] See Figure 1 , Figure 1 This is a flowchart illustrating the meteorological data downscaling method based on a spatiotemporal graph neural network provided in this application embodiment.

[0043] In this embodiment of the application, the meteorological data downscaling method based on spatiotemporal graph neural networks described below can be applied to computer equipment such as servers.

[0044] In one embodiment, this application provides a meteorological data downscaling method based on a spatiotemporal graph neural network, comprising the following steps:

[0045] Step S110: Obtain meteorological observation datasets and gridded datasets from meteorological stations within the study area. The gridded datasets include low-resolution meteorological grid data, high-resolution meteorological grid data, and high-resolution static land grid data.

[0046] In one embodiment, there are usually multiple meteorological stations in the study area, and the meteorological observation dataset consists of meteorological observation data observed by the meteorological stations; optionally, the meteorological observation data may include data such as air temperature at a height of 2m above the ground, relative humidity at a height of 2m above the ground, wind speed and direction at a height of 10m above the ground, and cumulative rainfall.

[0047] Optionally, low-resolution meteorological grid data and high-resolution meteorological grid data may include data such as air temperature at a height of 2m above the ground, relative humidity at a height of 2m above the ground, wind speed and direction at a height of 10m above the ground, cumulative rainfall, surface temperature, soil moisture, and shortwave radiation; high-resolution land static grid data may include data such as building vectors, green space vectors, water body vectors, and land cover types.

[0048] In one embodiment, after acquiring high-resolution static land grid data, the building density, building height, impermeability, green space ratio, water body ratio, terrain height, and land use type within each grid can also be calculated based on the high-resolution static land grid data.

[0049] Step S120: Perform data preprocessing on the meteorological observation dataset and the gridded dataset, and use a preset correlation analysis algorithm to determine the corresponding feature variables.

[0050] In one embodiment, data preprocessing can be performed on meteorological observation datasets and gridded datasets by removing invalid data and filling in missing data. Removing invalid data can involve deleting data records containing obvious errors or outliers, for example, by identifying and removing data points that exceed a reasonable range through statistical analysis or by setting thresholds. Filling in missing data can be done by using appropriate methods to handle missing values, such as using the mean, median, nearest neighbor data points, or interpolation to fill in missing values, ensuring the integrity of the data in terms of time series and spatial distribution.

[0051] In one embodiment, after removing invalid data and filling in missing data, continuous data can be standardized to transform data of different scales to a uniform scale and eliminate the influence between different units; optionally, the data standardization process can adopt Min-Max standardization or Z-Score standardization.

[0052] By performing rigorous preprocessing on the data in the above manner, data noise and missing data can be significantly reduced, thereby improving data quality and reliability.

[0053] In one embodiment, the corresponding feature variables determined by the preset correlation analysis algorithm are used as input data of the trained first spatiotemporal graph neural network. Therefore, the corresponding feature variables can be the variables with the highest correlation to the high-resolution meteorological grid data, including air temperature at a height of 2m above the ground, surface temperature, shortwave radiation, and high-resolution land static grid data in the low-resolution meteorological grid data.

[0054] Step S130: Based on the meteorological observation dataset, the gridded dataset, and the corresponding feature variables, a dataset for downscaling is constructed.

[0055] Step S140: Input the dataset for downscaling into the trained first spatiotemporal graph neural network, and use the first spatiotemporal graph neural network to downscale the low-resolution meteorological grid data to obtain the predicted high-resolution meteorological grid data.

[0056] Specifically, the dataset used for downscaling is input into the first spatiotemporal graph neural network that has been trained, and the gridded dataset and corresponding feature variables in the dataset used for downscaling are used as the dataset to be downscaled and predicted and input into the first spatiotemporal graph neural network that has been trained.

[0057] In one embodiment, the first spatiotemporal graph neural network may include an embedding layer, a graph structure building module, a spatiotemporal convolution module, an attention module, and an output layer, as detailed below. Figure 2 , Figure 2 The structure and processing of the first spatiotemporal graph neural network are illustrated. The first spatiotemporal graph neural network can be obtained by training the spatiotemporal graph neural network to be trained using a constructed historical meteorological dataset. When training the spatiotemporal graph neural network to be trained, the historical meteorological dataset can be divided into a training set, a validation set, and a test set for training.

[0058] The first spatiotemporal graph neural network's embedding layer is used to map the input meteorological data and related features into a high-dimensional space to form embedding vectors, so that subsequent network modules can efficiently process different types of data. The embedding layer includes three functions: spatiotemporal encoding, average encoding, and feature embedding.

[0059] For the spatiotemporal coding portion, to enhance the spatiotemporal characteristics of meteorological data, spatiotemporal coding can be explicitly introduced as an auxiliary feature into the input features of each latitude and longitude grid. Specifically, the added auxiliary feature f st as follows:

[0060]

[0061] s encode =[sinh,cos h,sinw,cos w,sinh cos w,sinh sinw]

[0062] f st =[t encode ,s encode ]

[0063] Where t represents the current time, s represents the current position, h represents the latitude of the grid, and w represents the longitude of the grid;

[0064] For the average coding part, since the land use types calculated using high-resolution static land grid data are represented by different values, in order to eliminate the influence of the numerical value of the categorical variable on the model, the average coding method is used to embed the land use type features, and then the result is used as part of the input features.

[0065] For the feature embedding part, the initial features of a data sample, after undergoing spatiotemporal coding feature enhancement and average coding feature transformation, become the input features of the model. A 1×1 two-dimensional convolution can be used to perform preliminary feature extraction and transformation on the input features to obtain a high-dimensional embedding representation X. emb ∈R T,N,d This can enhance the model's expressive power and robustness. Here, N represents the number of latitude and longitude grids, T represents the size of the time window used during training, d0 represents the number of input features, and d represents the number of features in the high-dimensional embedding representation. This embedding process can be represented as:

[0066] X emb =BatchNorm2d(Conv2d(X))

[0067] Furthermore, the graph structure building module of the first spatiotemporal graph neural network is used to construct the node embedding representation X obtained from the embedding layer. emb Construct a graph structure and calculate the similarity between nodes based on the time-series embedding representation of each node. For this purpose, the K-Nearest Neighbors (KNN) method can be used to construct edges between pairs of nodes with high similarity. Finally, during the model optimization process, the corresponding graph structure is dynamically generated for use by the spatiotemporal graph neural network to better capture the high-order spatiotemporal dependency information between nodes.

[0068] Specifically, node embedding represents X emb After flattening, the node time series embedding representation X is obtained. emb ′∈R N,T×d Then, the adjacency matrix A is constructed using the K-nearest neighbor method; after each model iteration and optimization, a more expressive and robust embedding representation X is obtained. emb Therefore, the graph structure A constructed accordingly undergoes dynamic changes during training, as detailed below:

[0069] (1) Node embedding representation flattening:

[0070] X emb ′=reshape(X emb ,(N,T×d))

[0071] (2) Calculate the similarity between nodes:

[0072] The similarity matrix S between nodes can be calculated using Euclidean distance or cosine similarity, as shown in the following expression:

[0073] S ij =cosine_similarity(X ′ emb [i],X ′ emb [j])

[0074] Where S ij This represents the similarity between node i and node j;

[0075] (3) Construct the adjacency matrix:

[0076] An adjacency matrix A is constructed based on the similarity matrix S and the K-nearest neighbor method. For each node i, the edges between it and the K most similar nodes are retained. The elements A of the adjacency matrix are... ij Defined as:

[0077]

[0078] (4) Dynamically update the graph structure:

[0079] After each model iteration optimization, based on the new embedding representation X emb The adjacency matrix A is dynamically updated. This process is repeated during training to gradually optimize the graph structure and better capture the higher-order spatiotemporal dependencies between nodes.

[0080] Furthermore, the spatiotemporal convolution module of the first spatiotemporal graph neural network is a spatiotemporal graph neural network module used to process spatiotemporal graph data with dynamic graph structures. It combines temporal convolution and graph convolution, extracting and processing spatiotemporal features through multiple steps, thereby ensuring that the model can simultaneously handle the temporal dynamic changes and spatial distribution features of meteorological data. In specific implementation, a combination of temporal convolutional network (TCN) and graph convolutional network (GCN) is used to extract spatiotemporal features to obtain the spatiotemporal embedding representation H of nodes. The specific process is as follows:

[0081] (1) Temporal Convolutional Network (TCN):

[0082] This module uses two TCNs, one for capturing features in the temporal dimension of the embedded representations before and after the graph convolution operation; assuming the TCN kernel size is k, the temporal convolution operation can be represented as:

[0083] X TCN =Conv1D(X emb ,k)

[0084] (2) Graph Convolutional Network (GCN):

[0085] Graph convolution is performed on the features output by the first-layer TCN to capture features in the spatial dimension. Assuming the weight matrix of the graph convolution is W and the adjacency matrix is ​​A, the graph convolution operation can be represented as:

[0086] X GCN =σ(AX) TCN W)

[0087] Where σ represents the activation function, such as ReLU;

[0088] (3) Spatiotemporal graph neural network:

[0089] Ultimately, the entire spatiotemporal feature learning process can be represented as:

[0090] H = TCN2(GCN(TCN1(X) emb )))

[0091] Furthermore, the attention module of the first spatiotemporal graph neural network is used to learn the contextual information of nodes in the spatial dimension and the correlation information of the same node in the temporal dimension. This module introduces a cross-attention mechanism (Criss-CrossAttention), which calculates the relationship between target feature pixels and other pixels, weights the features of target pixels, and captures dense contextual information while maintaining long-distance spatial dependence.

[0092] Specifically, the output H of the spatiotemporal convolution module is used as the feature map input to the attention module. Then, cross-attention is used to capture the spatial context information of nodes and the temporal association information of the same node. Finally, the aggregated node embedding representation H′ is generated.

[0093]

[0094] A = softmax(f affinity (QK T ))

[0095] H′=A⊙V

[0096] Among them, Q and K are feature maps generated by H through a 1×1 convolutional layer. To reduce the computational complexity, let C' < C. V is a feature map with the same dimension as H and is used for feature adaptation. The affinity operation aims to calculate the relationships between each node in the graph and other nodes in the same row and column, and generate a new feature map. Then, softmax normalization is adopted to obtain a new attention map A. The ⊙ operation represents element-wise multiplication of matrices. After multiplying the attention map and V element-wise, the node embedding representation H' is obtained.

[0097] Furthermore, the output layer of the first spatio-temporal graph neural network is used to receive the node embedding representation output from the attention module and generate the predicted value of the initial high-resolution meteorological data. Specifically, the output layer may include a two-dimensional convolutional layer and a fully connected layer in the time and space dimensions, map the embedding representation with the time window length to the target time step, and map the high-dimensional feature back to the original space to generate the downscaled meteorological data.

[0098] In one embodiment, after the first spatio-temporal graph neural network performs downscaling on the low-resolution meteorological grid data to generate the downscaled meteorological data, the initial value of the generated high-resolution meteorological data can be post-processed to obtain the predicted high-resolution meteorological grid data. The data post-processing methods may include outlier detection and removal, interpolation in space and time, data smoothing and filtering in space, and data smoothing and filtering in time. Data post-processing can effectively improve the quality of the downscaled meteorological data and ensure the physical consistency, spatial continuity, and temporal continuity of the data.

[0099] During the downscaling process of the above-mentioned structure of the first spatio-temporal graph neural network, effective data processing and feature enhancement are carried out, thereby enhancing the detail performance and accuracy of meteorological data, capturing the details of local meteorological phenomena, and more accurately predicting high-resolution meteorological data. Specifically, through the above method, the embedding layer of the first spatio-temporal graph neural network regards meteorological data and its geographical information as graph node data, uses the graph neural network to process complex spatio-temporal relationships, improves the efficiency and accuracy of data processing, and enhances the feature representation ability through spatio-temporal coding and average coding, ensuring the robustness and stability of the model. The graph structure construction module dynamically updates the graph structure after each model iteration, captures the high-order spatio-temporal dependence information between nodes, improves the model performance. Compared with the traditional method using convolutional networks and static graph structures, it can better capture global information in space rather than being limited to local information, and adaptively adjusts the graph structure according to the optimization objective, which is beneficial to improving the model effect. The spatio-temporal convolutional module can simultaneously learn features in both the time and space dimensions, process the spatio-temporal dynamic changes of meteorological data, and improve the model prediction accuracy. The attention module improves the accuracy of meteorological data prediction by simultaneously capturing the correlation information in both the space and time dimensions.

[0100] The aforementioned meteorological data downscaling method based on spatiotemporal graph neural networks acquires meteorological observation datasets and gridded datasets from meteorological stations within the study area. The gridded datasets include low-resolution meteorological grid data, high-resolution meteorological grid data, and high-resolution static land grid data. A pre-defined correlation analysis algorithm is used to determine the corresponding feature variables, enhancing the diversity and richness of the data sources and covering different resolutions and data types, thereby improving the prediction accuracy of the neural network model. Simultaneously, a trained first spatiotemporal graph neural network is used to downscale the low-resolution meteorological grid data, enabling faster acquisition of predicted high-resolution meteorological grid data and significantly improving efficiency. Furthermore, the first spatiotemporal graph neural network provides higher spatial and temporal resolution, enhancing the detail and accuracy of meteorological data, thus more accurately capturing the details of local meteorological phenomena and improving the resolution and accuracy of the downscaled meteorological data. In addition, the first spatiotemporal graph neural network also has higher generalization ability and can be used to process different regions, different time periods, or different meteorological types.

[0101] For this, please refer to Figure 3 , Figure 3 This diagram illustrates the downscaling results and prediction errors of air temperature at a height of 2m above ground, obtained using the meteorological data downscaling method based on a spatiotemporal graph neural network according to embodiments of this application. The high-resolution meteorological grid data is simulated using a pre-set meteorological research and forecasting modeling system, a simulation process that requires considerable time. The predicted high-resolution meteorological grid data is obtained using a trained first spatiotemporal graph neural network, requiring extremely little time, significantly reducing processing time and improving efficiency. Furthermore, comparison with the high-resolution meteorological grid data shows that the predicted high-resolution meteorological grid data effectively reproduces the details of meteorological phenomena. The prediction error is the difference between the high-resolution meteorological grid data and the predicted high-resolution meteorological grid data; the prediction error is low and meets accuracy requirements.

[0102] In one embodiment, when acquiring meteorological observation datasets and gridded datasets from meteorological stations within the study area, the following can be done:

[0103] Acquire meteorological observation datasets from meteorological stations within the study area, as well as low-resolution meteorological data and high-resolution land static data;

[0104] The study area is divided into grids to obtain the corresponding grid.

[0105] Based on the grid division, low-resolution meteorological data, and high-resolution land static data, low-resolution meteorological grid data, high-resolution land static grid data, and high-resolution meteorological grid data are obtained. These three data are used as a gridded dataset.

[0106] In this embodiment, low-resolution meteorological data can be derived from numerical weather prediction models or remote sensing observations, while high-resolution static land data can be derived from resource satellite remote sensing inversion and geographic information system data.

[0107] In this embodiment, when dividing the study area into grids, the spatial resolution of the grid can be determined based on the fitting results of the semivariograms of different meteorological elements in the meteorological station observation data. When the meteorological station data is missing or insufficient for fitting, the spatial resolution of the grid can be directly set to 0.5km x 0.5km, and then the corresponding grid can be obtained by dividing the grid through the spatial resolution.

[0108] In this embodiment, when obtaining low-resolution meteorological grid data, high-resolution land static grid data, and high-resolution meteorological grid data based on grid division, low-resolution meteorological data, and high-resolution land static data, the low-resolution meteorological data can be interpolated to the grid division to obtain low-resolution meteorological grid data.

[0109] High-resolution static land grid data is calculated based on the grid division and high-resolution static land data.

[0110] High-resolution static land grid data and low-resolution meteorological grid data are input into a pre-set meteorological research and forecasting modeling system to obtain high-resolution meteorological grid data.

[0111] Specifically, low-resolution meteorological data can be interpolated to a grid using bilinear interpolation or cubic convolution interpolation methods to obtain low-resolution meteorological grid data. Meanwhile, high-resolution static land grid data and low-resolution meteorological grid data are input into a pre-set meteorological research and forecasting modeling system (such as WRF). During simulation, the urban canopy model is activated, and dynamic downscaling is performed using the WRF model to obtain high-resolution meteorological grid data.

[0112] The above methods can effectively and accurately collect various meteorological data, including station meteorological data and meteorological gridded data, ensuring the diversity of data sources and covering different resolutions and data types, thereby improving the prediction accuracy of the model.

[0113] In one embodiment, when constructing a dataset for downscaling based on meteorological observation datasets, gridded datasets, and corresponding feature variables, the following can be done:

[0114] Based on the determined number of lag steps, obtain the time lag variable corresponding to the current time;

[0115] Based on the meteorological observation dataset, the gridded dataset, the corresponding feature variables, and the corresponding time lag variables, a lag time feature dataset is constructed as a dataset for downscaling.

[0116] In this embodiment, the values ​​of the previous n time points (lag steps) can be introduced as features of the current time point to form time lag variables. The introduction of time lag variables helps the model capture the autocorrelation of data and understand how past states affect the current state, thus enabling the model to have better predictive performance.

[0117] The lag step number n can be determined in the following ways: (1) Autocorrelation analysis and partial autocorrelation analysis: Identify significant lag steps n by plotting and analyzing autocorrelation function and partial autocorrelation function graphs; (2) Domain knowledge: Combine professional knowledge in the field of meteorology to select time lag steps with practical physical significance. For example, the daily periodic variation in meteorological data may suggest using 24 or 12 hours as the lag step number n; (3) Cross-validation: Test the model performance of different n values ​​through cross-validation method and select the lag step number n that makes the model prediction effect the best.

[0118] In one embodiment, when constructing a dataset for downscaling, a moving average feature can also be introduced. This moving average feature is formed by calculating the average value of the data within a certain window size m. The moving average feature smooths out short-term fluctuations in the data and emphasizes long-term trends, enabling the model to better capture the overall trend of the data.

[0119] For a given window size m, it can be determined in the following ways: (1) Rolling window analysis: Try rolling windows with different window sizes m, and determine the best window size by verifying the model performance; (2) Rule of thumb: Select the window size m that is usually effective based on domain experience and historical data characteristics. For example, for daily periodic changes in meteorological data, it may be recommended to use 6 or 3 hours as the window size m.

[0120] See Figure 1 In one embodiment, in step S140, the dataset for downscaling is input into a trained first spatiotemporal graph neural network. After the first spatiotemporal graph neural network is used to downscale the low-resolution meteorological grid data to obtain the predicted high-resolution meteorological grid data, the meteorological data downscaling method based on spatiotemporal graph neural network of this application further includes:

[0121] Step S150: Input the dataset for downscaling and the predicted high-resolution meteorological grid data into the trained second spatiotemporal graph neural network, and use the second spatiotemporal graph neural network to correct the bias of the predicted high-resolution meteorological grid data to obtain the corrected high-resolution meteorological grid data.

[0122] Specifically, the dataset for downscaling is input into the trained second spatiotemporal graph neural network. The meteorological observation dataset, high-resolution land static grid data, corresponding feature variables, and predicted high-resolution meteorological grid data in the dataset for downscaling are used as the dataset to be biased and input into the trained second spatiotemporal graph neural network.

[0123] By correcting the biases of the predicted high-resolution meteorological grid data, the trends and detailed characteristics of meteorological changes can be reflected more accurately, further improving the resolution and accuracy of the downscaled meteorological data, making it more effective for various practical applications such as weather forecasting, disaster early warning, and environmental monitoring.

[0124] In one embodiment, in step S120, when preprocessing the meteorological observation dataset and the gridded dataset, and determining the corresponding feature variables using a preset correlation analysis algorithm, the determined corresponding feature variables include not only the variables with the highest correlation to the high-resolution meteorological grid data, but also the variables with the highest correlation to the meteorological observation data of the meteorological stations, including high-resolution meteorological data prediction values, land static data, and time information, etc., and the feature variables are used to construct a dataset for downscaling as input data for the second spatiotemporal graph neural network.

[0125] In one embodiment, the second spatiotemporal graph neural network may include a data processing module, a spatiotemporal representation learning module, an uncertainty modeling module, and an extrapolation generation module, as detailed in [reference needed]. Figure 4 , Figure 4 The structure and processing of the second spatiotemporal graph neural network are illustrated. The second spatiotemporal graph neural network can be obtained by training the spatiotemporal graph neural network to be trained using a constructed historical meteorological dataset and the corresponding predicted high-resolution meteorological grid data. When training the spatiotemporal graph neural network to be trained, the historical meteorological dataset and the corresponding predicted high-resolution meteorological grid data can be divided into training set, validation set and test set for training.

[0126] The data processing module of the second spatiotemporal graph neural network is used to construct the dataset and adjacency matrix based on the grid where the meteorological stations are located, specifically including:

[0127] (1) The input meteorological station data is regarded as nodes in the graph. The nodes at each time step are divided into context nodes and target nodes. Each node contains exogenous variables and predictive variables. The exogenous variables include the predicted high-resolution meteorological grid data obtained by the first spatiotemporal graph neural network and the corresponding land use static variables. The predictive variables are the meteorological observation data corresponding to the predicted high-resolution meteorological grid data.

[0128] (2) For missing values ​​of the target node prediction variables, learnable embedding vectors are used to fill them in;

[0129] (3) For all input nodes, add time features to their exogenous variables to achieve feature enhancement. The added time features are:

[0130]

[0131] Where t represents the current time;

[0132] (4) For all input nodes, calculate the adjacency matrix of the graph based on their latitude and longitude information, calculate the Euclidean distance between two nodes as the weight of the edge, and normalize the obtained adjacency matrix.

[0133] Furthermore, the spatiotemporal representation learning module is used to model the spatiotemporal dependencies of node features. Combining dilated noncausal convolution and graph convolution, it extracts and processes spatiotemporal features through multiple steps, thereby ensuring that the model can simultaneously handle the temporal variations and spatial distribution features of node data. Specifically, this includes:

[0134] (1) Dilated Noncausal Convolution (DNCConv):

[0135] To capture the long-term dependencies of node features, expand the receptive field of the convolutional neural network, and enable data correction to utilize both historical and future information, the spatiotemporal representation learning module uses two dilated noncausal convolutions to capture the temporal features of the embedded representations before and after the graph convolution operation. Assuming the kernel size of DNCConv is k and the dilation factor is d, the temporal convolution operation can be represented as:

[0136] X DNCConv =Conv1D(X emb ,k,d)

[0137] Among them, X emb This represents the embedding of nodes;

[0138] (2) Graph Convolutional Network (GCN):

[0139] Graph convolution is performed on the features output by the first-layer TCN to capture features in the spatial dimension. Assuming the weight matrix of the graph convolution is W and the adjacency matrix is ​​A, the graph convolution operation can be represented as:

[0140] X GCN =σ(AX) DNCConv W)

[0141] Where σ represents the activation function, such as ReLU;

[0142] (3) Spatiotemporal graph convolution:

[0143] Ultimately, the entire spatiotemporal representation learning process can be represented as:

[0144] H = DNCConv2(GCN(DNCConv1(X) emb )))

[0145] Furthermore, the uncertainty modeling module is used to improve model robustness and stability, and enhance model generalization ability. It introduces uncertainty modeling of node embedding representations through the use of graph Bayesian aggregation and latent variable state transitions with gating mechanisms, specifically including:

[0146] (1) Graph Bayesian aggregation: Aggregate the information of context nodes through Bayesian methods, taking into account the uncertainty of each context node, so that the model can better handle the uncertainty of different nodes, thereby improving the reliability and accuracy of extrapolation results;

[0147] (2) Latent variable state transition with gating: Based on the spatiotemporal representation obtained in the deterministic representation learning stage, latent variables are passed to the next layer through state transition with gating mechanism to capture deeper spatiotemporal dependencies and pass and aggregate uncertain information layer by layer. By introducing gating mechanism, the effective information of latent variables can be retained in the state transition process while removing redundant information, thereby improving the stability of the model.

[0148] Furthermore, the extrapolation generation module is used to extrapolate the predictor variables of the target node, that is, to predict the distribution of the predictor variables of the target node. Specifically, by combining the latent variable representations of all layers and the exogenous variables of the target node, a likelihood function is used to generate the distribution of the target node. This process is essentially a conditional generation process, in which the prediction of the target node is based on the information of its surrounding context nodes and the uncertainty learned in the model. The final extrapolation result can be obtained by sampling from the distribution, or the mean of the distribution can be directly used as the predicted value, which is the corrected high-resolution meteorological grid data.

[0149] The aforementioned second spatiotemporal graph neural network structure effectively captures the long-term dependence and spatial distribution characteristics of node features during the bias correction process, and handles the uncertainty of different nodes, thereby achieving the purpose of data bias correction and further improving the resolution and accuracy of downscaled meteorological data. Specifically, through the above methods, the spatiotemporal representation learning module combines dilated noncausal convolution (DNCConv) and graph convolution (GCN), which can more effectively capture the long-term dependence and spatial distribution characteristics of node features in the meteorological field, improving the model's representation ability on spatiotemporal data. The uncertainty modeling module introduces graph Bayesian aggregation and latent variable state transition, enabling the model to better handle the uncertainty of different nodes, and improves the robustness and stability of the model through gating mechanisms. The extrapolation generation module generates the distribution of the target node's predictive variables by combining the latent variable representations of all layers and the exogenous variables of the target node. It can sample from the predictive distribution or directly use the mean of the distribution to obtain the predicted value, improving the flexibility and accuracy of the prediction results.

[0150] In order to implement the methods corresponding to the above embodiments and achieve the corresponding functions and technical effects, a meteorological data downscaling device based on spatiotemporal graph neural network is provided below.

[0151] See Figure 5 , Figure 5 This is a structural block diagram of a meteorological data downscaling device based on a spatiotemporal graph neural network provided in an embodiment of this application.

[0152] In one embodiment, the meteorological data downscaling device based on spatiotemporal graph neural networks of this application includes:

[0153] The data acquisition unit 510 is used to acquire meteorological observation datasets and gridded datasets from meteorological stations within the study area. The gridded datasets include low-resolution meteorological grid data, high-resolution meteorological grid data, and high-resolution static land grid data. The data preprocessing unit 520 is used to preprocess the meteorological observation datasets and gridded datasets and determine the corresponding feature variables using a preset correlation analysis algorithm.

[0154] Dataset construction unit 530 is used to construct a dataset for downscaling based on meteorological observation datasets, gridded datasets, and corresponding feature variables;

[0155] The data downscaling unit 540 is used to input the dataset for downscaling into the trained first spatiotemporal graph neural network, and use the first spatiotemporal graph neural network to downscale the low-resolution meteorological grid data to obtain the predicted high-resolution meteorological grid data.

[0156] In one embodiment, the data acquisition unit 510 may be specifically used for:

[0157] Acquire meteorological observation datasets from meteorological stations within the study area, as well as low-resolution meteorological data and high-resolution land static data;

[0158] The study area is divided into grids to obtain the corresponding grid.

[0159] Based on the grid division, low-resolution meteorological data, and high-resolution land static data, low-resolution meteorological grid data, high-resolution land static grid data, and high-resolution meteorological grid data are obtained. These three data are used as a gridded dataset.

[0160] As an optional implementation, when the data acquisition unit 510 obtains low-resolution meteorological grid data, high-resolution land static grid data, and high-resolution meteorological grid data based on the grid division, low-resolution meteorological data, and high-resolution land static data, it can:

[0161] Low-resolution meteorological data is interpolated onto a grid to obtain low-resolution meteorological grid data;

[0162] High-resolution static land grid data is calculated based on the grid division and high-resolution static land data.

[0163] High-resolution static land grid data and low-resolution meteorological grid data are input into a pre-set meteorological research and forecasting modeling system to obtain high-resolution meteorological grid data.

[0164] In one embodiment, the dataset construction unit 530 may be specifically used for:

[0165] Based on the determined number of lag steps, obtain the time lag variable corresponding to the current time;

[0166] Based on the meteorological observation dataset, the gridded dataset, the corresponding feature variables, and the corresponding time lag variables, a lag time feature dataset is constructed as a dataset for downscaling.

[0167] In one embodiment, the meteorological data downscaling device based on spatiotemporal graph neural network of this application further includes: a data deviation correction unit 550, which is used to input the dataset for downscaling and the predicted high-resolution meteorological grid data into a trained second spatiotemporal graph neural network, and use the second spatiotemporal graph neural network to correct the deviation of the predicted high-resolution meteorological grid data to obtain corrected high-resolution meteorological grid data.

[0168] The aforementioned meteorological data downscaling device based on spatiotemporal graph neural networks can implement the meteorological data downscaling method based on spatiotemporal graph neural networks described above. For specific limitations and other details of the above-described embodiment of the meteorological data downscaling device based on spatiotemporal graph neural networks, please refer to the content of the meteorological data downscaling method based on spatiotemporal graph neural networks described above; further details will not be repeated in the embodiments.

[0169] In one embodiment, this application provides an electronic device including a memory and a processor. The memory stores a computer program, and the processor runs the computer program to enable the electronic device to perform the above-described meteorological data downscaling method based on a spatiotemporal graph neural network.

[0170] Alternatively, the aforementioned electronic device may be a computer device, such as a server.

[0171] In one embodiment, the internal structure of the computer device of this application can be as follows: Figure 6 As shown.

[0172] In one embodiment, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described meteorological data downscaling method based on a spatiotemporal graph neural network.

[0173] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can also be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0174] In addition, the functional modules in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0175] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0176] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application. It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0177] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0178] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Similarly, in the description of this application, the terms "first," "second," etc., are used only for descriptive distinction and should not be construed as indicating or implying relative importance. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

Claims

1. A meteorological data downscaling method based on spatiotemporal graph neural networks, characterized in that, include: Acquire meteorological observation datasets and gridded datasets from meteorological stations within the study area. The gridded datasets include low-resolution meteorological grid data, high-resolution meteorological grid data, and high-resolution static land grid data. The process includes: acquiring meteorological observation datasets from meteorological stations within the study area, as well as low-resolution meteorological data and high-resolution static land data; dividing the study area into 500m×500m grids to obtain corresponding grids; interpolating the low-resolution meteorological data into the grids to obtain low-resolution meteorological grid data; calculating high-resolution static land grid data based on the grids and the high-resolution static land data; calculating building density, building height, impermeability, green space ratio, water body ratio, terrain height, and land use type within each grid based on the high-resolution static land grid data; inputting the high-resolution static land grid data and the low-resolution meteorological grid data into a preset meteorological research and forecasting modeling system; and enabling the urban canopy model during simulation to obtain high-resolution meteorological grid data with a grid spatial resolution of 500m×500m. The meteorological observation dataset and the gridded dataset are preprocessed, and the corresponding feature variables are determined using a preset correlation analysis algorithm. This includes: after removing invalid data and filling in missing data, the continuous data is standardized to convert data of different scales to a uniform scale and eliminate the influence between different units. The data standardization method adopts Min-Max standardization or Z-Score standardization. The corresponding feature variables determined by the preset correlation analysis algorithm are used as input data for the trained first spatiotemporal graph neural network. The corresponding feature variables are the variables with the highest correlation to the high-resolution meteorological grid data, including air temperature at 2m above the ground, surface temperature, shortwave radiation and high-resolution land static grid data in the low-resolution meteorological grid data. The low-resolution meteorological data is interpolated to the grid to obtain low-resolution meteorological grid data. The high-resolution land static grid data and the low-resolution meteorological grid data are input into a preset meteorological research and forecasting modeling system for simulation. During the simulation, the urban canopy model is enabled and dynamic downscaling is performed through WRF mode to obtain high-resolution meteorological grid data. Based on the meteorological observation dataset, the gridded dataset, and the corresponding feature variables, a dataset for downscaling is constructed, with a grid spatial resolution of 500m×500m. The dataset used for downscaling is input into a trained first spatiotemporal graph neural network. The first spatiotemporal graph neural network is used to downscale the low-resolution meteorological grid data to obtain predicted high-resolution meteorological grid data with a grid spatial resolution of 500m×500m. The first spatiotemporal graph neural network includes an embedding layer, a graph structure construction module, a spatiotemporal convolution module, an attention module, and an output layer. The embedding layer of the first spatiotemporal graph neural network has the functions of spatiotemporal encoding, average encoding, and feature embedding. The graph structure construction module of the first spatiotemporal graph neural network dynamically constructs an adjacency matrix based on the K-nearest neighbor method according to the geographic spatial relationship. The output layer of the first spatiotemporal graph neural network includes a two-dimensional convolutional layer and a fully connected layer in the temporal and spatial dimensions.

2. The meteorological data downscaling method based on spatiotemporal graph neural networks according to claim 1, characterized in that, The step of constructing a dataset for downscaling based on the meteorological observation dataset, the gridded dataset, and the corresponding feature variables includes: Based on the determined number of lag steps, obtain the time lag variable corresponding to the current time; Based on the meteorological observation dataset, the gridded dataset, the corresponding feature variables, and the corresponding time lag variables, a lag time feature dataset is constructed as a dataset for downscaling.

3. The meteorological data downscaling method based on spatiotemporal graph neural networks according to claim 1, characterized in that, After inputting the dataset used for downscaling into a trained first spatiotemporal graph neural network, and using the first spatiotemporal graph neural network to downscale the low-resolution meteorological grid data to obtain predicted high-resolution meteorological grid data, the method further includes: The dataset used for downscaling and the predicted high-resolution meteorological grid data are input into a trained second spatiotemporal graph neural network. The second spatiotemporal graph neural network is then used to correct the bias in the predicted high-resolution meteorological grid data to obtain the corrected high-resolution meteorological grid data.

4. The meteorological data downscaling method based on spatiotemporal graph neural networks according to claim 3, characterized in that, The second spatiotemporal graph neural network includes a data processing module, a spatiotemporal representation learning module, an uncertainty modeling module, and an extrapolation generation module.

5. A meteorological data downscaling device based on a spatiotemporal graph neural network, characterized in that, include: The data acquisition unit is used to acquire meteorological observation datasets and gridded datasets from meteorological stations within the study area. The gridded datasets include low-resolution meteorological grid data, high-resolution meteorological grid data, and high-resolution static land grid data. The data preprocessing unit is used to preprocess the meteorological observation dataset and the gridded dataset, and to determine the corresponding feature variables using a preset correlation analysis algorithm. The dataset construction unit is used to construct a dataset for downscaling based on the meteorological observation dataset, the gridded dataset, and the corresponding feature variables. The data downscaling unit is used to input the dataset for downscaling into a trained first spatiotemporal graph neural network, and use the first spatiotemporal graph neural network to downscale the low-resolution meteorological grid data to obtain predicted high-resolution meteorological grid data.

6. An electronic device, characterized in that, The device includes a memory and a processor, wherein the memory stores a computer program and the processor runs the computer program to enable the electronic device to perform the meteorological data downscaling method based on a spatiotemporal graph neural network according to any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the meteorological data downscaling method based on a spatiotemporal graph neural network as described in any one of claims 1 to 4.

Images