A micro-regional meteorological data encoding method based on graph neural network

Through the micro-region meteorological data encoding method based on graph neural network, a-Grid and a-Mesh structures are constructed, combined with spatiotemporal eigenvectors, and information propagation and fusion are used by ST-GNN model, which solves the problem of comprehensive capture of meteorological systems in the existing technology. The problem of time and space changes has achieved deeper understanding and simulation of meteorological systems.

CN118627548BActive Publication Date: 2024-11-08SINE SPACE (ANHUI) TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411105727.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-13
Publication Date
2024-11-08
Estimated Expiration
2044-08-13

Smart Images

  • Figure CN118627548B_ABST
    Figure CN118627548B_ABST
Patent Text Reader

Abstract

The present invention relates to the field of meteorological technology, and discloses a method for encoding micro-regional meteorological data based on a graph neural network, comprising the following steps: in a micro-region a, constructing an a-Grid structure and a nested a-Mesh structure, the a-Grid structure records geographic and meteorological information as a starting point, and the a-Mesh structure is densely arranged with small regular triangles. The present invention incorporates time series processing capabilities based on GNN, effectively propagates information in a time-space hybrid graph, so that the model can not only capture the local spatial characteristics of the current moment, but also take into account the impact of historical moments, and how these characteristics change over time, thereby promoting the effective fusion and propagation of information in the time-space dimension, so that the model can more comprehensively and deeply understand and simulate the behavior of meteorological systems in the real world.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of meteorology, and more specifically, to a micro-region meteorological data encoding method based on graph neural network. Background Art

[0002] The meteorological system is a typical spatiotemporal coupled system, in which weather phenomena and meteorological conditions vary with time and space, and there are complex interactions between the two.

[0003] The occurrence of a rainfall event in one area may affect meteorological parameters such as humidity and temperature in neighboring areas. At the same time, the meteorological conditions at different time points in the same area are also highly correlated. Therefore, considering only spatial relationships or analyzing time series alone cannot fully capture the dynamic characteristics of the meteorological system. Summary of the invention

[0004] The present invention provides a micro-region meteorological data encoding method based on graph neural network, which solves the technical problem in the related technology that only considering spatial relationships or analyzing time series alone cannot fully capture the dynamic characteristics of the meteorological system.

[0005] The present invention provides a micro-region meteorological data encoding method based on graph neural network, comprising the following steps:

[0006] S100: In micro-area a, an a-Grid structure with a maximum resolution of 50 m and a nested a-Mesh structure were constructed. The a-Grid structure records geographic and meteorological information as a starting point, and the a-Mesh structure is densely arranged with small regular triangles;

[0007] S200: Convert 69 meteorological variables into vector form and compile a time series containing historical meteorological data for each grid node to create a spatiotemporal feature vector;

[0008] S300: Initialize the graph structure between a-Grid and a-Mesh, combine the spatiotemporal feature vectors to create a spatiotemporal hybrid graph, where a-Grid nodes are connected to adjacent a-Mesh nodes through edges to form a preliminary interactive network, and build an interactive graph model between a-Grid and a-Mesh. Aggregate information by initializing the code on a-Grid and propagating it to a-Mesh through a specially designed edge structure, and then transmit the information back to a-Grid.

[0009] S400: Constructing the ST-GNN model by incorporating spatiotemporal convolution, graph attention, and recurrent neural network components during initialization, transforming the data through MLP, and training using historical data;

[0010] S500: Deploy the optimized ST-GNN model to a cloud server or high-performance cluster, and process meteorological data through an API interface.

[0011] Furthermore, in step S100, a-Grid and a-Mesh structures are included:

[0012] a-Grid construction, for micro-area a, creates a grid structure with a maximum spatial resolution of 50m according to its spatial range; each grid node records the geographical location and the initial meteorological variable value as the starting point of the spatiotemporal sequence;

[0013] a-Mesh construction, using small regular triangles with a side length of no more than 50m to build a-Mesh, which is tightly embedded inside the a-Grid to achieve seamless coverage;

[0014] The triangle vertices serve as a-Mesh nodes and are connected by edges to form a highly interconnected network.

[0015] Furthermore, in the a-Grid structure, micro-areas are set The boundary is , the spatial resolution is ;

[0016] The a-Grid construction process involves Divide into A rectangular grid, where is the number of grids along the width and length of the micro-area, which meets the total area coverage requirement and the side length of each grid is close to a-Grid;

[0017] Grid node representation: the center or corner of each grid is marked as a node , ,in The geographic location of Indicates that the initial meteorological variable value set is ,in Indicates The grid starts at The value of a meteorological variable.

[0018] Furthermore, the a-Mesh structure also includes the following:

[0019] The definition of small equilateral triangles is as follows. The side length of each small equilateral triangle is set to , compatible with the spatial resolution of a-Grid;

[0020] The nodes and edges are as follows: In a-Mesh, the vertices of each small equilateral triangle are As a network node, is the total number of nodes;

[0021] If the nodes are connected by the edges of triangles, they are defined as edge-connected, denoted by connect and , forming a graph ;

[0022] Each a-Mesh node Aligning the grid nodes of a-Grid , and the small regular triangles are arranged continuously, without overlap and covering a micro area The border every part of.

[0023] Furthermore, the steps to create the spatiotemporal feature vector are as follows:

[0024] S210: For the 69 meteorological variables in micro-area a, denoted as , and normalize it:

[0025] Normalization:

[0026]

[0027] in, is the normalized Meteorological variable values, and Respectively represent the minimum and maximum values ​​of the variable in the entire data set;

[0028] S220: For each grid node To construct spatiotemporal series data, we first need to integrate historical meteorological observations;

[0029] Assume that each node has The observation records at each time point record the status of 69 meteorological variables, and the data at each time point are organized into a feature vector:

[0030]

[0031] in, Represents at a point in time On the grid nodes No. The observed values ​​of meteorological variables are is the corresponding timestamp, used to mark the position of the time series;

[0032] For each node, we get a time series feature matrix .

[0033] Furthermore, the number of meteorological variables is 69, of which 4 are surface variables and 65 are atmospheric variables with different atmospheric pressures; the number of atmospheric pressures is 13 and the number of atmospheric variables is 5;

[0034] The surface variables include the temperature at 2 meters, the u component of the wind at 10 meters, the v component of the wind at 10 meters, and the mean sea level pressure;

[0035] Among them are atmospheric variables in different atmospheric pressures, where atmospheric pressures include 1000hPa, 925hPa, 850hPa, 700hPa, 600hPa, 500hPa, 400hPa, 300hPa, 250hPa, 200hPa, 150hPa, 100hPa and 50hPa.

[0036] The atmospheric variables include potential, specific humidity, temperature, u component of wind and v component of wind.

[0037] Furthermore, in step S300, the following contents are included:

[0038] The steps for initializing the graph structure between a-Grid and a-Mesh are as follows:

[0039] Construction process: define two node sets and Represent the nodes in a-Grid and a-Mesh respectively, where ,and , each node Contains spatial coordinates ;

[0040] Edges of a graph It consists of two parts: Indicates the connection from a-Grid to a-Mesh, Represents the connections within a-Mesh;

[0041] Initial Image It forms the basic framework of the space-time hybrid graph;

[0042] The specific steps of information aggregation by initializing the code on a-Grid and propagating it to a-Mesh through a specially designed edge structure are as follows:

[0043] a-Grid initial encoding: For each a-Grid node , the meteorological variable vector it carries , use multi-layer perceptron to encode and obtain high-dimensional feature vector , where the parameter set of MLP is ;

[0044] Design connection method: for each Select the most recent indivual , establish a connection , ensuring that the encoded information can be propagated from a-Grid to the corresponding a-Mesh nodes;

[0045] a-Mesh information aggregation: define information aggregation function , using graph convolution , Graph Attention Mechanism ;

[0046] in is the adjacency matrix inside a-Mesh, Represents the feature matrix of the a-Mesh node in the previous layer;

[0047] New feature vector pass renew;

[0048] Backward Propagation: Design the reverse connection from a-Mesh to a-Grid to allow information flow Update, which can be expressed as ,in is the fusion function of back propagation, Representation and The set of all connected a-Mesh nodes;

[0049] aGM-GNN graph definition: The whole process forms a mechanism for spatiotemporal information propagation and fusion, recorded as a micro-regional meteorological coding graph neural network, which is represented as an iterative process;

[0050] in Indicates The node feature matrix of the first step is transformed into a node feature matrix, and through multiple iterative propagation and information fusion, the deep encoding and fusion of spatiotemporal features are finally achieved.

[0051] Furthermore, the steps of constructing the ST-GNN model are as follows:

[0052] Model initialization: Design the ST-GNN model, combining spatiotemporal convolutional layers, graph attention mechanisms, and recurrent neural networks or LSTM units designed specifically for processing time series to capture spatiotemporal dynamic features;

[0053] Spatial-temporal feature encoding and propagation: Initially, MLP is used to perform feature mapping on the meteorological data in a-Grid;

[0054] Through the ST-GNN model, feature propagation is performed in the spatiotemporal mixed graph, graph convolution is used to capture spatial dependencies, and RNN / LSTM units capture time series features to achieve deep fusion of spatiotemporal features;

[0055] Training and optimization: The historical spatiotemporal series meteorological data set is used for training, and the mean square error is used as the loss function to evaluate the difference between the predicted value and the actual value.

[0056] A micro-regional meteorological data coding system based on graph neural network, comprising:

[0057] Space-time structure building module: responsible for creating the space-time infrastructure within the micro-region;

[0058] Data preprocessing and feature engineering module: standardize the raw meteorological data and convert it into a format suitable for model input;

[0059] Spatiotemporal hybrid graph construction and coding propagation module: Initialize the graph structure between a-Grid and a-Mesh, define the connection method between nodes, and realize the two-way propagation of information in the grid and mesh structure;

[0060] ST-GNN model design and training module: Design and implement the ST-GNN model, integrate spatiotemporal convolutional layers, graph attention mechanisms, and recurrent neural network components, and perform deep learning training on preprocessed data;

[0061] Model deployment and service interface module: Deploy the trained ST-GNN model to a cloud server or high-performance computing cluster to ensure that the model runs efficiently in the actual environment.

[0062] A storage medium storing non-temporary computer-readable instructions for executing one or more steps in the above-mentioned micro-regional meteorological data encoding method based on graph neural network.

[0063] The beneficial effects of the present invention are:

[0064] Incorporating time series processing capabilities on the basis of GNN can more effectively propagate information in mixed space-time graphs, so that the model can not only capture the local spatial characteristics of the current moment, but also take into account the impact of historical moments and how these characteristics change over time, thereby promoting the effective fusion and dissemination of information in the space-time dimensions, enabling the model to more comprehensively and deeply understand and simulate the behavior of meteorological systems in the real world. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 It is a flow chart of a micro-region meteorological data encoding method based on graph neural network of the present invention;

[0066] Figure 2 It is a structural block diagram of a micro-regional meteorological data encoding system based on graph neural network of the present invention.

[0067] In the figure: 101, spatiotemporal structure construction module; 102, data preprocessing and feature engineering module; 103, spatiotemporal hybrid graph construction and encoding propagation module; 104, ST-GNN model design and training module; 105, model deployment and service interface module. DETAILED DESCRIPTION

[0068] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that the discussion of these embodiments is only to enable those skilled in the art to better understand and implement the subject matter described herein, and the functions and arrangements of the elements discussed may be changed without departing from the scope of protection of the contents of this specification. Each example may omit, replace or add various processes or components as needed. In addition, the features described relative to some examples may also be combined in other examples.

[0069] See also Figure 1-Figure 2 As shown, a micro-regional meteorological data encoding method based on graph neural network includes the following steps:

[0070] S100: Micro-regional space-time structure design

[0071] a-Grid and a-Mesh Construction

[0072] a-Grid construction: For micro-area a, a grid structure with a maximum spatial resolution of 50m is created according to its spatial scope (such as urban blocks, road layout);

[0073] Each grid node not only records the geographical location, but also stores the initial meteorological variable values ​​as the starting point of the spatiotemporal series;

[0074] a-Mesh construction: a-Mesh is constructed using small regular triangles with a side length of no more than 50m, which are tightly embedded inside the a-Grid to ensure seamless coverage of the space;

[0075] The triangle vertices serve as a-Mesh nodes, connected to each other through edges, forming a highly interconnected network that promotes efficient transmission of spatiotemporal information;

[0076] It should be noted that the definition of micro-area a is as follows:

[0077] The area range of micro-area a may be the space between buildings in a city, the space formed by roads, etc., and its spatial resolution does not exceed 50 m. It is necessary to encode the meteorological data in such micro-area a.

[0078] These variables cover most of the common meteorological elements and are sufficient to play an important role in the weather forecast of micro-region a.

[0079] The whole process of encoding first requires designing the grid structure and mesh structure of the micro-area;

[0080] Then, a multi-layer perceptron neural network (MLP) is used to vector map the initial historical meteorological variables. It is necessary to design the connection method between the grid and the mesh so that the feature representation of the meteorological data after the multi-layer perceptron neural network mapping can be propagated to the mesh structure, so as to be processed in the mesh structure and the processed results can be propagated to the grid structure again.

[0081] In one embodiment of the present invention, step S100 also includes the following contents:

[0082] a-Grid Construction

[0083] Definition: Micro-area The boundary is , the spatial resolution is ;

[0084] The a-Grid construction process involves Divide into A rectangular grid, where is the number of grids along the width and length of the micro-area, which meets the total area coverage requirement and the side length of each grid is close to a-Grid;

[0085] Grid node representation: The center or corner (depending on the design) of each grid is marked as a node , ,in The geographic location of Indicates that the initial meteorological variable value set is ,in Indicates The grid starts at The value of a meteorological variable.

[0086] a-Mesh construction

[0087] Definition of small equilateral triangle: The side length of each small equilateral triangle is set to , ensuring compatibility with the spatial resolution of a-Grid.

[0088] Nodes and edges: In a-Mesh, the vertices of each triangle As a network node, is the total number of nodes. If the nodes are connected by triangle edges, they are defined as edge-connected, denoted as connect and , forming a graph .

[0089] Embedding and seamless coverage: To ensure seamless integration of a-Mesh and a-Grid, the algorithm must ensure that each a-Mesh node Grid nodes located at or very close to a-Grid , and all small triangles are closely arranged, without overlap and covering Every part of the system can maximize the use of space and continuously transmit information.

[0090] S200: Spatiotemporal Data Preprocessing and Feature Engineering

[0091] Standardize or normalize 69 meteorological variables and convert them into vector form suitable for model input;

[0092] Spatiotemporal series construction: Generate time series data for each grid node, integrate historical meteorological observations, and construct a spatiotemporal feature vector containing timestamps;

[0093] S210: Standardization / normalization of meteorological variables

[0094] For the 69 meteorological variables in micro-region a (denoted as ), and normalize it:

[0095] Min-max scaling (normalized):

[0096]

[0097] in, is the normalized Meteorological variable values, and They represent the minimum and maximum values ​​of the variable in the entire data set, respectively.

[0098] S220: Spatiotemporal sequence construction

[0099] For each grid node To construct spatiotemporal series data, we first need to integrate historical meteorological observations. Assume that each node has The observation records at each time point record the status of 69 meteorological variables. The data at each time point can be organized into a feature vector:

[0100]

[0101] in, Represents at a point in time On the grid nodes No. The observed values ​​of meteorological variables are It is the corresponding timestamp, which is used to mark the position of the time series.

[0102] In this way, for each node, a time series feature matrix can be obtained .

[0103] Through the above steps, the original meteorological data is converted into a series of spatiotemporal feature vectors suitable for the input of the machine learning model, laying the foundation for subsequent spatiotemporal prediction or analysis.

[0104] In one embodiment of the present invention, the number of meteorological variables is 69, including 4 surface variables, 5 atmospheric variables, and 13 different atmospheric pressures. The description of atmospheric variables can be represented by the variable name set pressure layer, for example, z1000 represents the value of the z meteorological element at an altitude of 1000hPa.

[0105] There are 4 surface variables, including temperature at 2 meters, u component of wind at 10 meters, v component of wind at 10 meters and mean sea level pressure; there are 5 atmospheric variables, including potential, specific humidity, temperature, u component of wind and v component of wind; there are 13 different atmospheric pressure levels, including 1000hPa, 925hPa, 850hPa, 700hPa, 600hPa, 500hPa, 400hPa, 300hPa, 250hPa, 200hPa, 150hPa, 100hPa and 50hPa;

[0106] Specifically, the number of meteorological variables is 69, of which there are 4 surface variables and 65 atmospheric variables with different atmospheric pressures; the number of atmospheric pressures is 13 and the number of atmospheric variables is 5;

[0107] The surface variables include the temperature at 2 meters, the u component of the wind at 10 meters, the v component of the wind at 10 meters, and the mean sea level pressure;

[0108] Among them are atmospheric variables in different atmospheric pressures, where atmospheric pressures include 1000hPa, 925hPa, 850hPa, 700hPa, 600hPa, 500hPa, 400hPa, 300hPa, 250hPa, 200hPa, 150hPa, 100hPa and 50hPa.

[0109] The atmospheric variables include potential, specific humidity, temperature, u component of wind and v component of wind.

[0110] S300: Spatiotemporal hybrid graph structure design and coding propagation process

[0111] Graph structure initialization: Build the preliminary layout of a-Grid and a-Mesh, ensure that each a-Grid node can be connected to the adjacent a-Mesh node through edges, and form a preliminary graph structure.

[0112] Initial coding and information dissemination:

[0113] a-Grid initial encoding: First, on a-Grid, the meteorological data of each grid node are initially encoded using a multi-layer perceptron (MLP) or other suitable encoding techniques to convert the original meteorological variables into high-dimensional feature vectors in preparation for information dissemination.

[0114] Design connection mode: clarify the edge connection mode from a-Grid to a-Mesh to ensure that the encoded meteorological characteristics can be effectively transmitted from a-Grid nodes to corresponding a-Mesh nodes.

[0115] a-Mesh information aggregation: In the a-Mesh structure, information aggregation operations are performed on the meteorological features received from the a-Grid (for example, through graph convolution or graph attention mechanism). This step realizes the integration of local information and enrichment of features.

[0116] Back propagation: Design the back propagation path from a-Mesh to a-Grid, and transmit the processed information back to the a-Grid node, so that the feature representation of the a-Grid node can integrate the advanced features from a-Mesh, and finally form a more comprehensive and in-depth meteorological element variable encoding result.

[0117] aGM-GNN graph definition: The interactive network between a-Grid and a-Mesh constructed in this process forms a micro-area meteorological coding graph neural network (aGM-GNN), which emphasizes the two-way propagation and fusion of information between grid and mesh structures.

[0118] In one embodiment of the present invention, the specific contents of the above are as follows:

[0119] S310: Graph structure initialization

[0120] Construction process: define two node sets and Represent the nodes in a-Grid and a-Mesh respectively, where ,and , each node Contains spatial coordinates .

[0121] Edges of a graph It consists of two parts:

[0122] Indicates the connection from a-Grid to a-Mesh, Represents the connections within a-Mesh. Initial diagram It constitutes the basic framework of the space-time hybrid graph.

[0123] S320: Initial coding and information dissemination. The specific steps are as follows:

[0124] S321: a-Grid initial encoding: For each a-Grid node , the meteorological variable vector it carries , use multi-layer perceptron (MLP) to encode and obtain high-dimensional feature vector , where the parameter set of MLP is ;

[0125] S322: Design connection method: for each Select the most recent indivual , establish a connection , ensuring that the encoded information can be propagated from a-Grid to the corresponding a-Mesh nodes.

[0126] S323: a-Mesh information aggregation: define information aggregation function , for example using graph convolution Or graph attention mechanism ,in is the adjacency matrix inside a-Mesh, Represents the feature matrix of the a-Mesh node in the previous layer.

[0127] New feature vector pass renew.

[0128] S324: Backward Propagation: Design the reverse connection from a-Mesh to a-Grid to allow information flow Update, which can be expressed as ,in is the fusion function of back propagation, Representation and The collection of all connected a-Mesh nodes.

[0129] S325: aGM-GNN graph definition: The whole process forms a mechanism for spatiotemporal information propagation and fusion, denoted as micro-regional meteorological coding graph neural network (aGM-GNN), which can be mathematically represented as an iterative process, where Indicates The node feature matrix of the first step is transformed into a node feature matrix, and through multiple iterative propagation and information fusion, the deep encoding and fusion of spatiotemporal features are finally achieved.

[0130] This process not only considers the feature transfer in the spatial dimension, but also incorporates the dynamic characteristics of time series data, reflecting the joint processing capabilities of spatiotemporal data.

[0131] S400: Spatiotemporal Graph Neural Network (ST-GNN) Model Design and Training

[0132] Model initialization: Design the ST-GNN model, combining spatiotemporal convolutional layers, graph attention mechanisms, and recurrent neural networks (RNNs) or LSTM units designed specifically for processing time series to capture spatiotemporal dynamic features.

[0133] Spatial-temporal feature encoding and propagation: Initially, MLP is used to perform feature mapping on the meteorological data in a-Grid;

[0134] Through the ST-GNN model, feature propagation is performed in the spatiotemporal mixed graph, graph convolution is used to capture spatial dependencies, and RNN / LSTM units capture time series features to achieve deep fusion of spatiotemporal features;

[0135] Training and optimization: The historical spatiotemporal series meteorological data set is used for training, and the mean square error (MSE) is used as the loss function to evaluate the difference between the predicted value and the actual value;

[0136] Apply strategies such as early stopping and dynamic learning rate to optimize the training process and improve the model convergence speed and prediction performance.

[0137] In one embodiment of the present invention, step S400 includes the following steps:

[0138] S410: Model initialization

[0139] The bidirectional information propagation and fusion process between a-Grid and a-Mesh completed in step 3 produces a series of node representations with rich spatiotemporal features;

[0140] These encoded and aggregated feature vectors (i.e. and ) as the input feature of the ST-GNN model, where Is an a-Grid node The feature representation after propagation and feedback, Is a-Mesh node The feature representation of

[0141] When building the ST-GNN model, multiple components are integrated to process spatiotemporal data:

[0142] Graph convolution layer: For a node , its feature update can be expressed as:

[0143]

[0144] in, and is the weight matrix and bias term of graph convolution, is a non-linear activation function (such as ReLU), Representation Node The set of neighbor nodes of Representation Node The neighbor nodes of is an aggregate function, such as sum or average, Represents the time step Time Node The hidden state of Represents the time step Time Node The update status of .

[0145] Graph attention layer: for nodes With its neighbors The feature fusion of ;

[0146]

[0147] in, is the attention vector, represents transpose, is the weight matrix, represents vector concatenation, is the activation function, Represents the time step Time Node The hidden state of

[0148] Then normalize to get the attention weight , and update the node features accordingly;

[0149]

[0150] Recurrent Neural Network (RNN) / Long Short-Term Memory (LSTM) Unit: For time series features, for each node Apply a recurrent neural network or LSTM unit to the feature sequence:

[0151]

[0152] in, is the previous time step The hidden state of is the updated hidden state.

[0153] S420: Spatiotemporal feature encoding and propagation

[0154] MLP feature mapping: Use a multi-layer perceptron to preliminarily encode the raw meteorological data, which can be formalized as follows:

[0155]

[0156] in, is the feature vector after MLP transformation, is the original meteorological feature vector.

[0157] Spatiotemporal feature propagation: Combining the above mechanisms, ST-GNN iteratively updates node features in the spatiotemporal hybrid graph until the preset number of layers or convergence conditions are reached.

[0158] S430: Training and Optimization

[0159] Loss function: Mean Squared Error (MSE) is used as the loss function to measure the predicted value With the true value The gap:

[0160]

[0161] in, is the actual observed value, is the model prediction value, is the total number of samples;

[0162] Optimization strategy: Use the back-propagation algorithm combined with an optimizer (such as Adam) to minimize the loss function. Implement the Early Stopping strategy to prevent overfitting, that is, stop training when the performance of the validation set no longer improves.

[0163] Optimization algorithm: Adopt Adam optimizer, learning rate Can be adjusted dynamically, for example by batch or training round decay:

[0164]

[0165] in, is the exponential moving average of the square of the historical gradient, is a very small positive value added for numerical stability, For the time step The model parameters when For the time step The model parameters at the time of , where η is the learning rate, which determines the step size of parameter update, is the loss function About Model Parameters At the current parameter value The gradient of , indicating the direction in which the loss decreases fastest;

[0166] Dynamically adjust the learning rate (Learning Rate Schedule), such as learning rate decay, to accelerate convergence and improve model performance.

[0167] S500: Spatiotemporal model validation and application deployment

[0168] Performance evaluation: On an independent spatiotemporal validation set, the generalization ability of the model is evaluated through indicators such as accuracy, recall, and F1 score to ensure the stable performance of the model on unknown data;

[0169] Deployment and service: Deploy the optimized ST-GNN model to a cloud server or high-performance computing cluster;

[0170] Develop an API interface to support real-time or batch uploading of meteorological data, return the encoded meteorological characteristics based on the input time and space coordinates, or directly provide future meteorological forecast results to meet the needs of various application scenarios such as meteorological analysis, urban planning, and disaster warning.

[0171] In one embodiment of the present invention, step S500 includes the following steps:

[0172] S510: Performance Evaluation

[0173] Evaluation Metrics: On independent spatiotemporal validation sets The following classic evaluation indicators are used to measure the generalization ability of the model:

[0174] Accuracy: The ratio of correctly predicted samples to the total number of samples.

[0175]

[0176] Among them, TP (True Positive), TN (True Negative), FP (False Positive), and FN (False Negative) represent the number of true positive examples, true negative examples, false positive examples, and false negative examples, respectively.

[0177] Recall: The ratio of true positive examples to actual positive examples.

[0178]

[0179] F1 Score: The harmonic mean of precision and recall, which comprehensively considers the accuracy and completeness of the prediction.

[0180]

[0181] Among them, precision is defined as TP / (TP+FP).

[0182] Through these indicators, the prediction performance of the model on unseen data can be comprehensively evaluated to ensure the stability and reliability of the model.

[0183] S520: Deployment and Services

[0184] Model deployment: Deploy the trained ST-GNN model to a cloud server or high-performance computing cluster. During the deployment process, it is necessary to consider model version control, resource allocation, load balancing, etc. to ensure the stable operation of the model under high concurrent access.

[0185] API interface development: Build an API service layer to provide unified meteorological data processing and forecasting services. The main interface functions include:

[0186] Data upload interface: allows users to upload meteorological observation data in real time or batches. The data format must follow pre-defined standards.

[0187] Feature encoding query interface: receiving spatiotemporal coordinates As input, returns the encoded meteorological characteristics of that space-time point.

[0188] Weather forecast interface: Based on the input time and space coordinates, it provides weather forecast results for a certain period of time in the future.

[0189] The entire encoding process first requires designing the grid structure and mesh structure of the micro-area, and then using the multi-layer perceptron neural network (MLP) to perform vector mapping on the initial historical meteorological variables. It is necessary to design the connection method between the grid and the mesh so that the feature representation of the meteorological data after mapping by the multi-layer perceptron neural network can be propagated to the mesh structure, thereby processing it in the mesh structure and propagating the processed results to the grid structure again.

[0190] There are many ways to implement the present invention. It can be implemented in the form of an API interface to provide users with meteorological data encoding services, or it can be embedded in other weather forecast products in the form of a background algorithm. The following describes the specific operations of implementing the present invention in the form of an API interface.

[0191] First, the service provider needs to use historical meteorological data to train the aGM-GNN model, then obtain the weight coefficient of the aGM-GNN inference model, and then deploy it on the server side and open the interface;

[0192] The client uses the http request interface service and can obtain the encoded data of meteorological variables from the interface by passing in the longitude and latitude range.

[0193] like Figure 2 As shown in the figure, a micro-regional meteorological data coding system based on graph neural network is also proposed, including:

[0194] Spacetime Structure Building Blocks 101

[0195] Mission: Responsible for creating the space-time infrastructure within the micro-region.

[0196] First, a high-resolution a-Grid structure is constructed to record geographic location and initial meteorological data. Second, an a-Mesh structure is designed and nested to fill the interior of the a-Grid with small equilateral triangles to achieve fine division of space and seamless coverage of information.

[0197] Data Preprocessing and Feature Engineering Module 102

[0198] Task: Standardize the raw meteorological data and convert it into a format suitable for model input.

[0199] By generating a spatiotemporal feature vector containing time series of historical meteorological data for each grid node, preparation is provided for subsequent model training.

[0200] Spatiotemporal hybrid graph construction and coding propagation module 103

[0201] Task: Initialize the graph structure between a-Grid and a-Mesh, define the connection method between nodes, and realize the two-way propagation of information between grids and grid structures.

[0202] This module uses a multi-layer perceptron (MLP) for initial encoding, and promotes the propagation and return of features from a-Grid to a-Mesh through a graph structure, completing the deep encoding and fusion of spatiotemporal features.

[0203] ST-GNN model design and training module 104

[0204] Task: Design and implement the ST-GNN model, integrate spatiotemporal convolutional layers, graph attention mechanisms, and recurrent neural network components, and perform deep learning training on preprocessed data.

[0205] The historical spatiotemporal series dataset is used to preprocess the data through MLP, and MSE is used as the loss function for model optimization training.

[0206] Model deployment and service interface module 105

[0207] Task: Deploy the trained ST-GNN model to a cloud server or high-performance computing cluster to ensure that the model can run efficiently in the actual environment.

[0208] Develop API interfaces to support real-time and batch meteorological data upload, return encoded meteorological characteristics or future forecast results according to specified time and space coordinates, and meet the needs of diverse application scenarios such as meteorological analysis, urban planning, disaster warning, etc.

[0209] At least one embodiment of the present disclosure provides a storage medium storing non-temporary computer-readable instructions for executing one or more steps in the aforementioned AI image super-resolution detail enhancement method based on film and television production.

[0210] The computer program may be stored / distributed on suitable media, such as optical storage media or solid-state media supplied together with or as part of other hardware, but may also be distributed in other forms, such as via the Internet or other wired or wireless telecommunication systems.

[0211] The above describes an embodiment of the present invention, but this embodiment is not limited to the above-mentioned specific implementation mode. The above-mentioned specific implementation mode is merely illustrative and not restrictive. Under the guidance of this embodiment, ordinary technicians in this field can also make more forms of equivalent embodiments, all of which are within the protection of this embodiment.

Claims

1. A micro-regional meteorological data encoding method based on graph neural network, characterized in that: The following steps are involved: S100: In micro-area a, construct an a-Grid structure and a nested a-Mesh structure, the a-Grid structure records geographic and meteorological information as a starting point, and the a-Mesh structure is densely arranged with small regular triangles; S200: converting several meteorological variables into vector form, and compiling a time series containing historical meteorological data for each grid node to create a spatiotemporal feature vector; The steps to create a spatiotemporal feature vector are as follows: S210: For the 69 meteorological variables in micro-area a, denoted as , and normalize it: The normalized expression is as follows: ; in, is the normalized Meteorological variable values, and Respectively represent the minimum and maximum values ​​of the variable in the entire data set; S220: For each grid node To construct spatiotemporal series data, we first need to integrate historical meteorological observations; Assume that each node has The observation records at each time point record the status of 69 meteorological variables, and the data at each time point are organized into a feature vector: ; in, Represents at a point in time On the grid nodes No. The observed values ​​of meteorological variables are is the corresponding timestamp, used to mark the position of the time series; For each node, we get a time series feature matrix ; S300: Initialize the graph structure between a-Grid and a-Mesh, combine the spatiotemporal feature vectors to create a spatiotemporal hybrid graph, where a-Grid nodes are connected to adjacent a-Mesh nodes through edges to form a preliminary interactive network, build an interactive graph model between a-Grid and a-Mesh, aggregate information by initializing the code on a-Grid and propagating it to a-Mesh through the edge structure, and then transmit the information back to a-Grid; The steps for initializing the graph structure between a-Grid and a-Mesh are as follows: Construction process: define two node sets and Represent the nodes in a-Grid and a-Mesh respectively, where ,and , each node Contains spatial coordinates ; Edges of a graph It consists of two parts: Indicates the connection from a-Grid to a-Mesh, Represents the connections within a-Mesh; Two sets of nodes in the initial graph Two sets of nodes in the initial graph ; The specific steps of initializing the code on a-Grid and propagating it to a-Mesh through the edge structure for information aggregation are as follows: a-Grid initial encoding: For each a-Grid node , the meteorological variable vector it carries , use multi-layer perceptron to encode and obtain high-dimensional feature vector , where the parameter set of MLP is ; Design connection method: for each Select the most recent indivual , establish a connection , ensuring that the encoded information is propagated from a-Grid to the corresponding a-Mesh nodes; Mesh information aggregation: define information aggregation function , using graph convolution , Graph Attention Mechanism ; in is the adjacency matrix inside a-Mesh, Represents the feature matrix of the a-Mesh node in the previous layer; New feature vector pass renew; Backward Propagation: Design the reverse connection from a-Mesh to a-Grid to allow information flow Update, expressed as ,in is the fusion function of back propagation, Representation and The set of all connected a-Mesh nodes; aGM-GNN graph definition: The step process of edge structure propagation to a-Mesh for information aggregation forms a mechanism for spatiotemporal information propagation and fusion, recorded as micro-regional meteorological coding graph neural network, represented as an iterative process; in Indicates The node feature matrix of the step is propagated and fused multiple times through iterations, and finally the deep encoding and fusion of spatiotemporal features are achieved; S400: Constructing the ST-GNN model by incorporating spatiotemporal convolution, graph attention, and recurrent neural network components during initialization, transforming the data through MLP, and training using historical data; The steps to build the ST-GNN model are as follows: Model initialization: Design the ST-GNN model, combining spatiotemporal convolutional layers, graph attention mechanisms, and RNN / LSTM units designed specifically for processing time series to capture spatiotemporal dynamic features; The bidirectional information propagation and fusion process between a-Grid and a-Mesh generates several node representations of spatiotemporal features; The encoded and aggregated feature vector, and As the input features of the ST-GNN model, Is an a-Grid node The feature representation after propagation and feedback, Is a-Mesh node The feature representation of When building the ST-GNN model, multiple components are integrated to process spatiotemporal data: Graph convolution layer: For a node , its feature update can be expressed as: ; in, and is the weight matrix and bias term of graph convolution, is a nonlinear activation function, Representation Node The set of neighbor nodes of Representation Node The neighbor nodes of is an aggregate function, Represents the time step Time Node The hidden state of Represents the time step Time Node Update status of Graph attention layer: for nodes With its neighbors The feature fusion of ; ; in, is the attention vector, represents transpose, is the weight matrix, represents vector concatenation, is the activation function, Represents the time step Time Node The hidden state of Then normalize to get the attention weight , and update the node features accordingly; ; Recurrent Neural Network / Long Short-Term Memory Unit: For time series features, for each node Apply a recurrent neural network or LSTM unit to the feature sequence: ; in, is the previous time step The hidden state of is the updated hidden state; Spatial-temporal feature encoding and propagation: Initially, MLP is used to perform feature mapping on the meteorological data in a-Grid; Through the ST-GNN model, feature propagation is performed in the spatiotemporal mixed graph, graph convolution is used to capture spatial dependencies, and RNN / LSTM units capture time series features, so that spatiotemporal features can be deeply integrated; Training and optimization: Use historical spatiotemporal series meteorological data sets for training, use mean square error as the loss function, and evaluate the difference between the predicted value and the actual value; S500: Deploy the optimized ST-GNN model to a cloud server or high-performance cluster, and process meteorological data through an API interface.

2. According to the micro-regional meteorological data encoding method based on graph neural network in claim 1, it is characterized in that: In step S100, a-Grid and a-Mesh are constructed: a-Grid construction, for micro-area a, creates a grid structure with a maximum spatial resolution of 50m according to its spatial range; each grid node records the geographical location and the initial meteorological variable value as the starting point of the spatiotemporal sequence; a-Mesh construction, using small regular triangles with a side length of no more than 50m to build a-Mesh, which is tightly embedded inside the a-Grid to achieve seamless coverage; The triangle vertices serve as a-Mesh nodes and are connected by edges to form a highly interconnected network.

3. A micro-regional meteorological data encoding method based on graph neural network according to claim 2, characterized in that: In the a-Grid structure, micro-areas are set The boundary is , the spatial resolution is ; The a-Grid construction process involves Divide into A rectangular grid, where is the number of grids along the width and length of the micro-region, which meets the total area coverage requirement and the side length of each grid is close to a-Grid; is the number of grids along the width and length of the micro-region, which meets the total area coverage requirement and the side length of each grid is close to a-Grid; Grid node representation: the center or corner of each grid is marked as a node , ,in The geographic location of Indicates that the initial meteorological variable value set is ,in Indicates The grid starts at The value of a meteorological variable.

4. The micro-regional meteorological data encoding method based on graph neural network according to claim 3 is characterized in that: In the a-Mesh construction, the following are also included: The definition of small equilateral triangles is as follows. The side length of each small equilateral triangle is set to , compatible with the spatial resolution of a-Grid; The nodes and edges are as follows: In a-Mesh, the vertices of each small equilateral triangle are As a network node, is the total number of nodes; If the nodes are connected by the edges of triangles, they are defined as edge-connected, denoted by connect and , forming a graph ; Each a-Mesh node Aligning the grid nodes of a-Grid , and the small regular triangles are arranged continuously, without overlap and covering a micro area The border every part of.

5. A micro-regional meteorological data encoding method based on graph neural network according to claim 4, characterized in that: The number of meteorological variables is 69, of which 4 are surface variables and 65 are atmospheric variables with different atmospheric pressures; the number of atmospheric pressures is 13 and the number of atmospheric variables is 5; The surface variables include the temperature at 2 meters, the u component of the wind at 10 meters, the v component of the wind at 10 meters, and the mean sea level pressure; The atmospheric variables in different atmospheric pressures include 1000hPa, 925hPa, 850hPa, 700hPa, 600hPa, 500hPa, 400hPa, 300hPa, 250hPa, 200hPa, 150hPa, 100hPa and 50hPa; The atmospheric variables include potential, specific humidity, temperature, u component of wind and v component of wind.

6. A micro-region meteorological data encoding system based on graph neural network, encoding is achieved by the micro-region meteorological data encoding method based on graph neural network as described in any one of claims 1 to 5, characterized in that: include: Space-time structure building module: responsible for creating the space-time infrastructure within the micro-region; Data preprocessing and feature engineering module: standardize the raw meteorological data and convert it into a format suitable for model input; Spatiotemporal hybrid graph construction and coding propagation module: initialize the graph structure between a-Grid and a-Mesh, define the connection method between nodes, and realize the two-way propagation of information in the grid and mesh structure; ST-GNN model design and training module: Design and implement the ST-GNN model, integrate spatiotemporal convolutional layers, graph attention mechanisms, and recurrent neural network components, and perform deep learning training on preprocessed data; Model deployment and service interface module: Deploy the trained ST-GNN model to a cloud server or high-performance computing cluster to ensure that the model runs efficiently in the actual environment.

7. A storage medium, characterized in that: Non-temporary computer-readable instructions are stored for executing one or more steps in the micro-regional meteorological data encoding method based on graph neural network as described in any one of claims 1-5.