Multi-element meteorological prediction method and device based on space-time dynamic graph convolution

Abstract

The invention provides a multivariate meteorological prediction method and device based on space-time dynamic graph convolution, and belongs to the field of meteorological prediction, and the method comprises the steps: selecting a plurality of meteorological factors most related to a specific meteorological element to form a feature factor subset; constructing a circular neighborhood based on a target point based on a spatial Euclidean distance, and randomly sampling a plurality of sampling points in the neighborhood; for each sampling point and each target point, constructing a feature vector based on a feature factor, and calculating a mahalanobis distance between the target point and each sampling point; for each time point in the time sequence, a graph structure is constructed, and the weight value of the edge of the graph structure is equal to the reciprocal of the mahalanobis distance; and executing graph convolution operation on the graph structure of each time point, and obtaining a predicted target meteorological element value based on a time sequence after graph convolution. According to the method, the limitation of a static correlation structure is overcome, the model can adapt to the change rule of meteorological elements at different time points by designing a dynamic graph structure evolved along with time, and the prediction accuracy is improved.

Description

Technical Field

[0001] This invention belongs to the field of time series meteorological forecasting, and particularly relates to a multivariate meteorological forecasting method and apparatus based on spatiotemporal dynamic graph convolution. Background Technology

[0002] Weather forecasting is crucial for transportation, logistics, agriculture, and energy production. Modern numerical weather prediction (NWP) has been widely applied across various sectors, and the quality of weather forecasts continues to improve steadily.

[0003] In recent years, data-driven deep learning (DL) models have been developed for weather forecasting, reducing computational costs by several orders of magnitude compared to state-of-the-art NWP models. Weather forecasting is achieved by building data-driven models for predicting large-scale atmospheric circulation. These models can be trained on a variety of data sources, including climate model outputs, global circulation model (GCM) data, reanalysis products, and combinations of climate model outputs and reanalysis products.

[0004] In reality, meteorological factors interact to form complex systems, and meteorological conditions in geographically adjacent regions often exhibit strong correlations, primarily due to the continuity and propagation characteristics of meteorological systems. Local weather systems (such as clouds and pressure fields) follow specific movement trajectories and evolutionary patterns, causing spatiotemporal correlations in meteorological changes in adjacent regions. Existing technologies have begun to address this characteristic, such as introducing methods based on spatial heterogeneity to adaptively capture region-specific meteorological dependencies, and employing fast hierarchical graph neural networks to effectively capture spatial dependencies based on Tobler's law and the second geographical law. These attempts have been effective in correcting biases in local numerical weather prediction, confirming the importance of considering spatial dependencies.

[0005] However, current weather forecasting technologies still predominantly employ methods that predict individual regions independently, rarely taking into full account the complex inter-regional dependencies. This approach, which ignores spatial correlations between regions, may significantly limit the accuracy and reliability of weather forecasts.

[0006] In existing technologies, meteorological forecasting models generally employ methods that predict each region independently, rarely fully considering the complex dependencies between regions. Although current technologies have begun to address spatial heterogeneity and introduce graph neural networks to capture spatial dependencies, these models typically use static graph structures built based on fixed geographical distances or preset rules, failing to reflect the dynamic characteristics of meteorological systems over time. Especially when weather systems evolve rapidly, static graph structures struggle to accurately represent the actual correlations of meteorological elements between regions at different points in time, thus limiting the accuracy of forecasts.

[0007] In existing technologies, the methods for measuring inter-regional similarity are too simplistic, often using a single distance metric (such as Euclidean distance or correlation coefficient), which makes it difficult to fully capture the complex interactions between multi-dimensional meteorological factors, resulting in insufficient accuracy when constructing inter-regional associations.

[0008] In existing technologies, most meteorological forecasting models are designed individually for specific meteorological elements, lacking versatility and flexibility, and making it difficult to adapt to the forecasting needs of different meteorological elements. This dedicated design approach not only increases the complexity of model development and maintenance, but also makes it difficult to fully utilize the inherent correlations between feature factors, thus limiting the overall performance of the forecasting system. Summary of the Invention

[0009] To address the above technical problems, this invention provides a multivariate meteorological forecasting method and apparatus based on spatiotemporal dynamic graph convolution. The technical solution of this invention is as follows:

[0010] A multivariate meteorological forecasting method based on spatiotemporal dynamic graph convolution, the method comprising:

[0011] Step 1: Select multiple meteorological factors that are most relevant to a specific meteorological element to form a subset of characteristic factors;

[0012] Step 2: Construct a circular neighborhood based on the target point using spatial Euclidean distance, and randomly sample several sampling points within this neighborhood; for each sampling point and the target point, construct a feature vector based on feature factors and calculate the Mahalanobis distance between the target point and each sampling point;

[0013] Step 3: For each time point in the time series, construct a graph structure, where the weights of the edges are equal to the reciprocal of the Mahalanobis distance. Perform a graph convolution operation on the graph structure at each time point, and obtain the predicted target meteorological element values ​​based on the time series after graph convolution.

[0014] A multivariate weather forecasting device based on spatiotemporal dynamic graph convolution, comprising:

[0015] The feature factor subset construction module is used to select multiple meteorological factors that are most relevant to a specific meteorological element to form a feature factor subset.

[0016] The Mahalanobis distance calculation module constructs a circular neighborhood of the target point based on spatial Euclidean distance, and randomly samples several sampling points within this neighborhood; for each sampling point and the target point, a feature vector is constructed based on feature factors and the Mahalanobis distance between the target point and each sampling point is calculated.

[0017] The meteorological element value prediction module constructs a graph structure for each time point in the time series, with the weights of the graph structure edges equal to the reciprocal of the Mahalanobis distance. A graph convolution operation is performed on the graph structure at each time point, and the predicted target meteorological element value is obtained based on the time series after graph convolution.

[0018] An electronic device includes: one or more processors; and a memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement a multivariate weather forecasting method based on spatiotemporal dynamic graph convolution.

[0019] A computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, cause the processor to implement a multivariate weather forecasting method based on spatiotemporal dynamic graph convolution.

[0020] The present invention has the following beneficial effects:

[0021] (1) This invention overcomes the limitations of static correlation structure by designing a dynamic graph structure that evolves over time, enabling the model to adapt to the changing patterns of meteorological elements at different time points and improving prediction accuracy.

[0022] (2) This invention proposes a similarity measurement method that combines spatial Euclidean distance and Mahalanobis distance. Based on the feature factor selection using random forest, this method first constructs circular neighborhood relationships based on spatial Euclidean distance, and then quantifies the feature similarity between the neighborhood and the target region through Mahalanobis distance. It comprehensively considers the influence of spatial location and multidimensional feature factors, overcomes the limitations of a single distance measurement, and can more accurately identify and quantify the meteorological correlation between regions.

[0023] (3) This invention constructs a multivariate meteorological forecasting framework applicable to the forecasting of multiple meteorological elements by selectively choosing feature factors for specific meteorological elements and generating a dynamic graph structure that evolves over time, thereby improving the versatility, flexibility, and accuracy of the forecasting model. This framework extends similarity calculation to each moment in the time series, enabling the graph structure to evolve dynamically over time, adapting to the changing patterns of meteorological elements at different time points, effectively overcoming the technical problem that static graph structures cannot reflect the time-varying characteristics of the meteorological system. By performing graph convolution operations at each time point and combining them with a time series forecasting model, a forecasting model capable of capturing the spatial characteristics and temporal evolution of meteorological elements between regions is constructed, forming a general framework applicable to the forecasting of multiple meteorological elements. Attached Figure Description

[0024] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other. To achieve the above objectives, this invention adopts the following technical solution.

[0026] This invention provides a multivariate meteorological forecasting method based on spatiotemporal dynamic graph convolution, the flowchart of which is shown below. Figure 1 As shown, it includes:

[0027] Step 1: Select multiple meteorological factors that are most relevant to a specific meteorological element to form a subset of characteristic factors;

[0028] Step 2: Construct a circular neighborhood based on the target point using spatial Euclidean distance, and randomly sample several sampling points within this neighborhood; for each sampling point and the target point, construct a feature vector based on feature factors and calculate the Mahalanobis distance between the target point and each sampling point;

[0029] Step 3: For each time point in the time series, construct a graph structure, where the weights of the edges are equal to the reciprocal of the Mahalanobis distance. Perform a graph convolution operation on the graph structure at each time point, and obtain the predicted target meteorological element values ​​based on the time series after graph convolution.

[0030] Step 1: Select multiple meteorological factors most relevant to a specific meteorological element to form a subset of characteristic factors; specifically:

[0031] Suppose that the target meteorological elements are to be predicted (That is, the meteorological variables that need to be predicted, such as temperature, humidity, air pressure, etc.) A set of possible related meteorological factors (of which, each) (This refers to a meteorological factor, such as wind speed, cloud cover, precipitation, etc.). The objective of this invention is to select from these factors... species and The most relevant meteorological factors constitute a subset of characteristic factors. ,in, ( (Number of features selected). Includes:

[0032] Step 1.1: Construct the training dataset ,in, Indicates the first one sample 3D feature vector (each sample contains all) (Observed values ​​of various meteorological factors). Indicates the corresponding meteorological elements The observed values ​​(actual observed values ​​of the target variable); This represents the total number of samples (the number of observation records contained in the dataset).

[0033] Step 1.2: Initialize the random forest model, including Decision Tree (in (This refers to the number of decision trees).

[0034] Step 1.3: For each decision tree ( ), by sampling with replacement from the training set Extraction 1 sample, forming a sub-training set .

[0035] Step 1.4: Set the size of the random feature subset (each node of the tree) Number of features considered).

[0036] Step 1.5: Construct a decision tree For each node of the tree ,from Randomly selected from the features Each feature constitutes a subset of feature factors. .

[0037] Step 1.6: For each feature Determine the set of possible split points , It represents the number of possible split points.

[0038] Step 1.7: For each split point Calculate information gain:

[0039] ;

[0040] in, It is a node Entropy;

[0041] Is using features and split point Conditional entropy after partitioning.

[0042] Step 1.8: Select the optimal split point .

[0043] Step 1.9: Select the optimal feature and the corresponding split point .

[0044] Step 1.10: Based on features and split point Node The samples are divided into left and right child nodes, meaning the samples are divided into two groups: those with feature values ​​less than or equal to the threshold and those with feature values ​​greater than the threshold.

[0045] Step 1.11: Recursively construct the left and right subtrees until the maximum depth is reached.

[0046] Step 1.12: Initialize importance scores for all features .

[0047] Step 1.13: For each decision tree Determine its outside-bag sample set That is, the set of samples that were not selected to build the current tree.

[0048] Step 1.14: Calculate the original bag outside error:

[0049] ;

[0050] in, It is a loss function; This refers to the number of samples outside the bag; It is a decision tree for samples The predicted value.

[0051] Step 1.15: For each feature By randomly shuffling Chinese characteristics The value of generates perturbed out-of-bag samples. .

[0052] Step 1.16: Calculate the out-of-bag error after perturbation:

[0053] ;

[0054] Step 1.17: Calculate features exist The importance of the middle: .

[0055] Step 1.18: Update Features Overall importance score: .

[0056] Step 1.19: Score based on feature importance Sort all features in descending order: .

[0057] Step 1.20: Select the one with the highest importance score. These features are used as the final feature subset: .

[0058] Step 2: Construct a circular neighborhood based on the target point using spatial Euclidean distance, and randomly sample several points within this neighborhood; for each sample point and the target point, construct a feature vector based on feature factors and calculate the Mahalanobis distance between the target point and each sample point; specifically:

[0059] Step 2.1: Construct a circular neighborhood; including:

[0060] To predict target points Centered on, based on spatial Euclidean distance Generate circular neighborhood :

[0061] ;

[0062] in, It predicts the coordinates of the target point; It represents the coordinates of any point within the neighborhood; It is the radius of the circular neighborhood; Therefore The circular neighborhood centered on the center.

[0063] Step 2.2: Random sampling;

[0064] In the constructed circular neighborhood Inside, random sampling The set consists of 1 sampling point :

[0065] ;

[0066] ;

[0067] ;

[0068] in, It is the first The coordinates of each sampling point; This is the total number of sampling points; The radius is randomly generated and satisfies ; It is a randomly generated angle that satisfies ; It is the set of all sampling points.

[0069] Step 2.3: Construct feature vectors;

[0070] For each sampling point and target point Select Construct a feature vector from 1 feature factor:

[0071] ;

[0072] ;

[0073] in, The target point eigenvectors; Sampling points eigenvectors; It is the first Each characteristic factor at point The value; It is the number of selected feature factors.

[0074] Step 2.4: Calculate the Mahalanobis distance;

[0075] Calculate the Mahalanobis distance between the target point and each sampling point:

[0076] ;

[0077] in, The target point With sampling points Mahalanobis distance between them; It is the difference in eigenvectors; It is the covariance matrix of the feature factors, with dimension . ; It is the inverse of the covariance matrix.

[0078] covariance matrix The calculation formula is:

[0079] ;

[0080] in, It is the number of samples used to calculate the covariance matrix; It is the average of the feature vectors of all samples; It is the first The feature vector of each sample. The samples are defined by time intervals. Inside, The length of the time series, for each time point. All sampling points and predicted target points It consists of 3D feature vectors.

[0081] Step 3: For each time point in the time series, construct a graph structure, where the weights of the edges are equal to the reciprocal of the Mahalanobis distance. Perform a graph convolution operation on the graph structure at each time point, and obtain the predicted target meteorological element values ​​based on the time series after graph convolution.

[0082] Step 3.1 Construction of the spatiotemporal dynamic graph structure;

[0083] Considering a time series length of ,in This represents the total number of discrete time points. For each time point... Constructing a graph structure ,in, This represents the set of nodes in the graph, including the target point. and sampling point set , Indicates the total number of sampling points; Indicates a point in time The set of edges containing the target point With each sampling point , The connection between them; Indicates a point in time The weight matrix, where the edges weight The value is equal to the reciprocal of the Mahalanobis distance, that is... .

[0084] Step 3.2 Spatiotemporal dynamic graph convolution operation;

[0085] At each time point graph structure Execution graph convolution:

[0086] ;

[0087] in, Indicates a point in time The node feature matrix after graph convolution; It is a diagonal matrix, and its diagonal elements ; For time points The node feature matrix, with dimension is Each row represents the value of a target meteorological element at a node; For time points The learnable parameter matrix; It is a non-linear activation function.

[0088] For all time points After performing graph convolution, the time series of graph convolution is obtained. .

[0089] Step 3.3 Time series forecasting;

[0090] Input the time series data after graph convolution into the time series prediction model:

[0091] ;

[0092] in, This indicates the predicted target meteorological element values; Represents a time series forecasting function; It is a time series obtained after graph convolution.

[0093] Step 3.4 Model optimization;

[0094] Define loss function ,in, Indicates the total loss; Represents the loss calculation function; This represents the meteorological element values ​​predicted by the model; This represents the actual values ​​of meteorological elements.

[0095] The model parameters, including the graph convolutional layer parameters, are optimized using the backpropagation algorithm. And the parameters of the time series prediction model.

[0096] like Figure 1 As shown, the input is historical meteorological observation data, including the target point. and sampling point set In time series The characteristic factors and target meteorological elements.

[0097] For each time point : Indicates the current processing point in time. This represents an incrementing operation at a specific point in time. This indicates that the index of the current time point is less than the total length of the time series. Conditional judgment.

[0098] 1. Calculate the target point With each sampling point , Mahalanobis distance between ;

[0099] 2. Constructing the graph ,side The weight is the reciprocal of the Mahalanobis distance, i.e. ;

[0100] 3. Perform graph convolution operations using the optimized model to obtain the feature matrix. ;

[0101] Feature time series based on graph convolution Using the optimized time series forecasting model Predict target meteorological elements.

[0102] Output: Predicted meteorological element values .

[0103] Another aspect of the present invention provides a multivariate weather forecasting device based on spatiotemporal dynamic graph convolution, comprising:

[0104] The feature factor subset construction module is used to select multiple meteorological factors that are most relevant to a specific meteorological element to form a feature factor subset.

[0105] The Mahalanobis distance calculation module constructs a circular neighborhood of the target point based on spatial Euclidean distance, and randomly samples several sampling points within this neighborhood; for each sampling point and the target point, a feature vector is constructed based on feature factors and the Mahalanobis distance between the target point and each sampling point is calculated.

[0106] The meteorological element value prediction module constructs a graph structure for each time point in the time series, with the weights of the graph structure edges equal to the reciprocal of the Mahalanobis distance. A graph convolution operation is performed on the graph structure at each time point, and the predicted target meteorological element value is obtained based on the time series after graph convolution.

[0107] Another aspect of the present invention provides an electronic device, comprising: one or more processors; and a memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement a multivariate weather forecasting method based on spatiotemporal dynamic graph convolution.

[0108] Another aspect of the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enable the processor to implement a multivariate weather forecasting method based on spatiotemporal dynamic graph convolution.

Claims

1. A multivariate meteorological forecasting method based on spatiotemporal dynamic graph convolution, characterized in that, The method includes: Step 1: Select multiple meteorological factors that are most relevant to a specific meteorological element to form a subset of characteristic factors; Step 2: Construct a circular neighborhood based on the target point using spatial Euclidean distance, and randomly sample several sampling points within this neighborhood; for each sampling point and the target point, construct a feature vector based on feature factors and calculate the Mahalanobis distance between the target point and each sampling point; Step 3: For each time point in the time series, construct a graph structure, where the weights of the edges are equal to the reciprocal of the Mahalanobis distance. Perform a graph convolution operation on the graph structure at each time point, and obtain the predicted target meteorological element values ​​based on the time series after graph convolution.

2. The multivariate meteorological forecasting method based on spatiotemporal dynamic graph convolution according to claim 1, characterized in that, In step 1: For the predicted target meteorological elements ,have A set of related meteorological factors , of which each To represent a meteorological factor, from the set of meteorological factors Select Species and meteorological elements The most relevant meteorological factors constitute a subset of characteristic factors. ,in .

3. The multivariate meteorological forecasting method based on spatiotemporal dynamic graph convolution according to claim 2, characterized in that, meteorological elements These are the predicted meteorological variables, including temperature, humidity, and air pressure.

4. The multivariate meteorological forecasting method based on spatiotemporal dynamic graph convolution according to claim 2, characterized in that, Meteorological factors include wind speed, cloud cover, and precipitation.

5. The multivariate meteorological forecasting method based on spatiotemporal dynamic graph convolution according to claim 1, characterized in that, Step 1 includes: Step 1.1: Construct the training dataset ,in, Indicates the first one sample Dimensional features; Indicates the corresponding meteorological elements Observed values; Indicates the total number of samples; Step 1.2: Initialize the random forest model, including Decision Tree ; Step 1.3: For each decision tree ,in, From the training set Extraction 1 sample, forming a sub-training set ; Step 1.4: Set the size of the random feature factor subset , For each node of the decision tree Number of features to consider; Step 1.5: Construct each decision tree For each node of the tree ,from Randomly selected from the features Each feature constitutes a subset of feature factors. ; Step 1.6: For each feature Determine the set of possible split points , It represents the number of possible split points; Step 1.7: For each split point Calculate information gain: ， in, It is a node Entropy; Is using features and split point Conditional entropy after partitioning; Step 1.8: Select the optimal split point ; Step 1.9: Select the optimal feature and the corresponding split point ; Step 1.10: Based on features and split point Node The samples are divided into left and right child nodes, meaning the samples are divided into two groups: those with feature values ​​less than or equal to the threshold and those with feature values ​​greater than the threshold. Step 1.11: Recursively construct the left and right subtrees until the maximum depth is reached; Step 1.12: Initialize importance scores for all features ; Step 1.13: For each decision tree Determine its outside-bag sample set That is, the set of samples that were not selected to build the current tree; Step 1.14: Calculate the original bag outside error: ； in, It is a loss function; This refers to the number of samples outside the bag; It is a decision tree for samples The predicted value; Step 1.15: For each feature By randomly shuffling Chinese characteristics The value of generates perturbed out-of-bag samples. ; Step 1.16: Calculate the out-of-bag error after perturbation: ； Step 1.17: Calculate features exist The importance of the middle: ; Step 1.18: Update Features Overall importance score: ; Step 1.19: Score based on feature importance Sort all features in descending order: ; Step 1.20: Select the one with the highest importance score. These features are used as the final feature subset: .

6. The multivariate meteorological forecasting method based on spatiotemporal dynamic graph convolution according to claim 5, characterized in that, Step 2 includes: Step 2.1: Using the predicted target point Centered on, based on spatial Euclidean distance Generate circular neighborhood : ; in, It predicts the coordinates of the target point; It represents the coordinates of any point within the neighborhood; It is the radius of the circular neighborhood; Therefore A circular neighborhood centered on the center; Step 2.2: In the constructed circular neighborhood Inside, random sampling The set consists of 1 sampling point : ； ； ； in, It is the first The coordinates of each sampling point; This is the total number of sampling points; The radius is randomly generated and satisfies ; It is a randomly generated angle that satisfies ; It is the set of all sampling points; Step 2.3: For each sampling point and target point Select Construct a feature vector from 1 feature factor: ； ； in, The target point eigenvectors; Sampling points eigenvectors; It is the first Each characteristic factor at point The value; It is the number of selected feature factors; Step 2.4: Calculate the Mahalanobis distance between the target point and each sampling point: ； in, The target point With sampling points Mahalanobis distance between them; It is the difference in eigenvectors; It is the covariance matrix of the feature factors, with dimension . ; It is the inverse of the covariance matrix; covariance matrix The calculation formula is: ； in, It is the average of the feature vectors of all samples; It is the first The feature vector of each sample; the sample is composed of time intervals. Inside, The length of the time series, for each time point. All sampling points and predicted target points It consists of 3D feature vectors.

7. The multivariate meteorological forecasting method based on spatiotemporal dynamic graph convolution according to claim 6, characterized in that, Step 3 includes: Step 3.1, Considering the time series length is... ,in, This represents the total number of discrete time points, for each time point Constructing a graph structure ,in, This represents the set of nodes in the graph, including the target point. and sampling point set , Indicates the total number of sampling points; Indicates a point in time The set of edges containing the target point With each sampling point , The connection between them; Indicates a point in time The weight matrix, where the edges weight The value is equal to the reciprocal of the Mahalanobis distance, that is... ; Step 3.2, at each time point graph structure Execution graph convolution: ； in, Indicates a point in time The node feature matrix after graph convolution; It is a diagonal matrix, and its diagonal elements ; For time points The node feature matrix, with dimension is Each row represents the value of a target meteorological element at a node; For time points The learnable parameter matrix; It is a non-linear activation function; For all time points After performing graph convolution, the time series of graph convolution is obtained. ; Step 3.3: Input the time series data after graph convolution into the time series prediction model: ； in, This indicates the predicted target meteorological element values; Represents a time series forecasting function; It is a time series obtained after graph convolution.

8. The multivariate meteorological forecasting method based on spatiotemporal dynamic graph convolution according to claim 7, characterized in that, In time series forecasting models, a loss function is defined. ,in, Indicates the total loss; Represents the loss calculation function; This represents the meteorological element values ​​predicted by the model; Representing true meteorological element values; optimizing model parameters, including graph convolutional layer parameters, using a backpropagation algorithm. And the parameters of the time series prediction model.

9. A multivariate meteorological forecasting device based on spatiotemporal dynamic graph convolution, characterized in that, include: The feature factor subset construction module is used to select multiple meteorological factors that are most relevant to a specific meteorological element to form a feature factor subset. The Mahalanobis distance calculation module constructs a circular neighborhood of the target point based on spatial Euclidean distance, and randomly samples several sampling points within this neighborhood; for each sampling point and the target point, a feature vector is constructed based on feature factors and the Mahalanobis distance between the target point and each sampling point is calculated. The meteorological element value prediction module constructs a graph structure for each time point in the time series, with the weights of the graph structure edges equal to the reciprocal of the Mahalanobis distance. A graph convolution operation is performed on the graph structure at each time point, and the predicted target meteorological element value is obtained based on the time series after graph convolution.

10. An electronic device, characterized in that, include: One or more processors; A memory for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the multivariate weather forecasting method based on spatiotemporal dynamic graph convolution as described in any one of claims 1 to 8.

11. A computer-readable storage medium, characterized in that, It stores executable instructions that, when executed by a processor, cause the processor to implement the multivariate weather forecasting method based on spatiotemporal dynamic graph convolution as described in any one of claims 1 to 8.

Images