AI model meteorological element prediction method based on high-resolution visible light cloud image

Abstract

The present invention belongs to the field of AI meteorological forecasting technology. In order to solve the technical problems that the existing AI model training cost is high, the model is complex, and the accuracy of local meteorological element prediction is still low, an "AI model meteorological element prediction method based on high-resolution visible light cloud map" is proposed. The initial meteorological element data is obtained for data preprocessing and then divided into training data and test data; the RegionCast model is constructed in combination with the RegionMesh multi-grid map, and the AutoEncoder is used as the basic framework. The embedding layer, Grid2Mesh layer, multi-grid map processor, Mesh2Grid layer and restoration layer are added between the encoder and the decoder in sequence; the training data is input to iteratively train the RegionCast model to obtain a regional meteorological element prediction model; the test data is input into the trained regional meteorological element prediction model to obtain the final meteorological element prediction result, and the time series information of continuous time is captured by autoregression, and features such as spherical coordinate system are added to supplement the spatial features to enhance the prediction accuracy.

Description

Technical Field

[0001] The invention belongs to the technical field of AI weather forecasting. Background Art

[0002] With the rapid development of social economy, the impact of weather changes on agriculture, transportation, energy and other fields is becoming increasingly significant. Accurate weather forecasts are directly related to production safety and economic benefits. However, traditional weather forecasts mainly rely on physical models and numerical methods to simulate atmospheric physical processes for predictions, but the calculations are complex and require the use of large computers for calculations. At the same time, the prediction accuracy is low when dealing with complex climates. It relies on artificially set assumptions and is difficult to fully capture local meteorological phenomena, complex terrain effects and the interaction between multiple meteorological elements. It has limitations.

[0003] Meteorological element prediction methods based on AI models provide new solutions to the above problems. These models have powerful parallel computing capabilities, can respond to rapidly changing weather conditions in a timely manner, and can take advantage of big data. Through deep learning technology, they can flexibly adapt to different meteorological conditions, reduce the impact of artificially set assumptions, and relatively improve the accuracy and reliability of predictions.

[0004] However, the existing meteorological element prediction methods based on AI models are large in size and have high training costs. They mainly rely on global or regional low-resolution data. High-resolution visible light cloud image data can not only clearly depict the boundaries and layers of clouds, but also track the time series changes of clouds, providing high-quality data support for AI models to capture local climate characteristics. These high-resolution data have not been effectively applied, and lack a comprehensive analysis of the interaction between multiple meteorological elements, resulting in the low accuracy of local meteorological element predictions. Summary of the invention

[0005] In order to solve the technical problems of high training cost and complex model of existing AI models and low accuracy of local meteorological element prediction, the present invention proposes an "AI model meteorological element prediction method based on high-resolution visible light cloud images".

[0006] AI model meteorological element prediction method based on high-resolution visible light cloud image, such as Figure 1As shown in the figure, the ERA5 reanalysis data of the target area and the satellite medium-high resolution visible light cloud image data are obtained as the initial meteorological element data; the initial meteorological element data are preprocessed to obtain the target meteorological element tensor, and the target meteorological element tensor is divided into training data and test data; the training data is regarded as the grid node features evenly distributed in space, and then the corresponding regional multi-grid map is constructed according to the geographical area where the data is located, which is recorded as Mesh node. After the regional multi-grid map is constructed, the Grid node and the corresponding Mesh node in space are connected to obtain the RegionMesh multi-grid map; the RegionCast model is constructed in combination with the RegionMesh multi-grid map, and the training data is input to iteratively train the RegionCast model to obtain the regional meteorological element prediction model; the test data is input into the trained regional meteorological element prediction model to obtain the final meteorological element prediction result;

[0007] The structure of the RegionCast model is as follows Figure 3 As shown in the figure, AutoEncoder is used as the basic framework, and an embedding layer, a Grid2Mesh layer, a multi-grid graph processor, a Mesh2Grid layer, and a reduction layer are sequentially added between the encoder and the decoder;

[0008] The embedding layer is used to map the training data and the RegionMesh multi-grid map into a hidden embedding space;

[0009] The Grid2Mesh layer uses the GNN message passing mechanism to enable the Grid nodes and Mesh nodes to update the embedding of unidirectional edges, and then updates the embedding of Mesh nodes through unidirectional edges;

[0010] The multi-grid graph processor is capable of performing message passing between nodes on the RegionMesh multi-grid graph;

[0011] The Mesh2Grid layer updates the embedding of the input latitude and longitude grid through the one-way edge from Mesh node to Grid node in the reverse direction;

[0012] The restoration layer can map the processed embedding back to the original feature space, and predict the meteorological elements of each longitude and latitude grid point at the next moment by performing a reverse transformation on the embedding of the Grid node.

[0013] Technical effects:

[0014] The present invention aims to solve the problem that the existing AI model meteorological element prediction method does not fully utilize the information of multiple elements, adjacent spaces and high-resolution local data. By combining high-resolution visible light cloud image data to extract relevant features, the prediction accuracy is enhanced. In order to solve the problem that the existing prediction algorithm model is huge and the training cost is high, a regionalized, multi-meteorological element multi-grid RegionCast model is constructed, and the time series information of continuous time is captured by autoregression. The spatial features are supplemented by adding features such as spherical coordinate system, which can not only integrate high-resolution visible light cloud image data, but also reduce unnecessary computing consumption. By aggregating the meteorological element features of adjacent locations, local information can be fully utilized, thereby significantly improving the prediction accuracy.

[0015] In order to further verify the accuracy and precision of the prediction results of the present invention, the regional meteorological element prediction model is used to predict the meteorological elements of the target area (38.61°-53.51°N, 118.83°-135.13°E) for 8 consecutive days and 6 hours. The target area has a large dimensional span, with many terrains such as oceans, mountains, and plains, and the climate has obvious differences. The output of the final meteorological element prediction results is compared with the real data of ERA5 reanalysis data and high-resolution visible light cloud map data. The root mean square error (RMSE) and mean absolute error (MAE) are used as evaluation indicators. The comparison results of the evaluation indicators at each moment are shown in the following table:

[0016]

[0017]

[0018] It can be seen from the data in the table that the present invention can achieve continuous 186 hours of meteorological element prediction, and the root mean square error between the final meteorological element prediction result and the real data is 1.9487, and the mean absolute error is 1.0222. The root mean square error of the prediction result in the first 48 hours is less than 1, and the average error compared with the real data is only 0.8996, which improves the highest accuracy of the local meteorological element prediction algorithm in this field. The above experiments show that the regional meteorological element prediction method based on the RegionCast model proposed by the present invention can fully extract the long-term correlation characteristics between multiple elements of adjacent locations, and achieve high-accuracy prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a flowchart of the overall process of the present invention.

[0020] Figure 2 This is a schematic diagram of the RegionMesh multi-grid graph structure.

[0021] Figure 3This is a schematic diagram of the overall structure of the RegionCast model. DETAILED DESCRIPTION

[0022] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0023] An embodiment of the present application discloses an “AI model meteorological element prediction method based on high-resolution visible light cloud images”, which includes: obtaining ERA5 reanalysis data of the target area and satellite medium- and high-resolution visible light cloud image data as initial meteorological element data.

[0024] In this embodiment, it is first determined that the target area to be predicted is 38.61°-53.51°N, 118.83°-135.13°E, and the initial meteorological element data are specifically the fifth-generation global reanalysis data product ERA5 reanalysis data provided by ECMWF (European Centerfor Mediumrange Weather Forecasts) and the high-resolution visible light cloud image data of the sunflower-8 satellite, wherein multiple meteorological elements in the ERA5 reanalysis data and 6 visible light bands in the high-resolution visible light cloud image data of the sunflower-8 satellite, totaling 77 meteorological elements, are shown in the following table.

[0025]

[0026] Optionally, in order to save computing resources, part of the atmospheric variables of the ERA5 reanalysis data is extracted from the pressure layer, and the pressure layers are selected as 50, 100, 150, 250, 300, 400, 500, 600, 700, 750, 850, 925, and 1000, with the unit being MPa.

[0027] Preprocessing the initial meteorological element data to obtain a target meteorological element tensor, and dividing the target meteorological element tensor into training data and test data;

[0028] In this embodiment, the data preprocessing may specifically include: uniformly adjusting the spatial resolution to 0.25° and the temporal resolution to 6 hours to ensure consistency when compared and analyzed with real data; in order to prevent gradient explosion and stabilize the training process, the initial meteorological element data is mean normalized; in order to prevent certain elements from dominating the model training process, the Z-Score standardization method is used for standardization.

[0029] The training data is regarded as the grid node features evenly distributed in space, and then the corresponding regional multi-grid map is constructed according to the geographical area where the training data is located, which is recorded as Mesh node. After the regional multi-grid map is constructed, the Grid node and the corresponding Mesh node in space are connected to obtain the RegionMesh multi-grid map. The construction process of RegionMesh multi-grid map is as follows Figure 2 As shown, specifically, the longitude and latitude coordinates of the input Grid nodes are used in a pooling manner to construct spatially uniformly distributed Mesh nodes. The characteristics of the Mesh nodes are the three coordinates of the spherical coordinate system where the node is located. Then, the links (edges) between the Mesh nodes are constructed, that is, adjacent nodes and high-order neighbor nodes. The characteristics of the edges between Mesh nodes are defined as the distance between the nodes at both ends of the edge in the spherical coordinate system and the difference between the three coordinates of the spherical coordinate system. The definition of the edge features between the Grid nodes and the Mesh nodes is the same as that of the edge features between the Mesh nodes.

[0030] In this embodiment, the convolution kernel size of the RegionMesh multi-grid map is set to 3, the step size is 3, and the number of layers is set to 4.

[0031] The RegionCast model is constructed by combining the RegionMesh multi-grid graph. The structure of the RegionCast model is as follows: Figure 3 As shown in the figure, AutoEncoder is used as the basic framework to obtain the accurate position of the grid points in the feature space, and an embedding layer, a Grid2Mesh layer, a multi-grid graph processor, a Mesh2Grid layer and a reduction layer are added between the encoder and the decoder in sequence.

[0032] The embedding layer is used to map the training data and the RegionMesh multi-grid graph into a hidden embedding space. Specifically, the embedding layer can be a multi-layer perceptron structure. Through the embedding layer, the edges from the Grid node to the Mesh node, the edges from the Mesh node to the Grid node, and the edges of the RegionMesh multi-grid graph can be mapped from low-dimensional attributes to a higher-dimensional hidden embedding space so as to match the input training data, which is more conducive to the accurate position of the learning node in the hidden embedding space.

[0033] The Grid2Mesh layer uses the GNN message passing mechanism to enable the Grid nodes and Mesh nodes to update the embedding of unidirectional edges, and then updates the embedding of Mesh nodes through unidirectional edges. Specifically, the Grid2Mesh layer uses GNN to extract relevant features and effectively captures the spatial relationship of meteorological elements. Through the GNN message passing mechanism, the information of the Grid node is first transmitted to the edge from the Grid node to the Mesh node, and then integrated and transmitted to the corresponding Mesh node through these unidirectional edges, realizing efficient information flow and feature sharing, so that the RegionCast model can learn the mutual influence of meteorological elements in adjacent spaces and provide rich feature information for the Mesh nodes.

[0034] The multi-grid graph processor can perform message transmission between nodes on the RegionMesh multi-grid graph. The multi-grid graph processor uses the advantages of GNN to aggregate information of adjacent nodes and dynamically update the feature representation of each node. This aggregation method based on neighborhood information can effectively capture the complex relationship between meteorological elements in the local area, and enhances the RegionCast model's ability to understand multiple meteorological elements as a whole.

[0035] The Mesh2Grid layer updates the embedding of the input longitude and latitude grid through the unidirectional edge from Mesh node to Grid node. Through the unidirectional edge from Mesh node to Grid node, the Mesh2Grid layer ensures that each Grid node receives the latest information from the surrounding Mesh nodes. This process not only enhances the adaptability of the RegionCast model to local meteorological characteristics, but also improves the accuracy of the final prediction results.

[0036] The restoration layer can map the processed embedding back to the original feature space, and predict the meteorological elements of each longitude and latitude grid point at the next moment by reversely transforming the embedding of the Grid node. As the last module of the RegionCast model, the restoration layer ensures the interpretability of the output of the prediction results, so that it can intuitively reflect the actual meteorological elements.

[0037] After the above RegionCast model is constructed, the training data is input to iteratively train the RegionCast model to obtain the regional meteorological element prediction model, and the test data is input into the trained regional meteorological element prediction model to obtain the final meteorological element prediction result.

[0038] In this embodiment, the back propagation algorithm used in the iterative training is Adam, which requires less memory and has higher computational efficiency. The initial learning rate r of the network is set to 5e-4, and the learning rate decay factor is set to 0.05. The total number of training iterations is 100 times, and the validation set score is calculated once for each iteration. Finally, the model with the highest validation set score is selected and migrated to the test set for correction and score calculation. The above iterative training process is a conventional method and will not be described in detail.

[0039] In this embodiment, a graphics processing unit (GPU) server is used, the chip is NVIDIA GeForce RTX 3060, and the memory capacity is 12GB.

[0040] The preferred embodiments of the present invention are described in detail above in conjunction with the accompanying drawings. However, the present invention is not limited to the specific details in the above embodiments. Within the technical concept of the present invention, a variety of simple modifications can be made to the technical solution of the present invention, and these simple modifications all belong to the protection scope of the present invention.

Claims

1. The AI ​​model meteorological element prediction method based on high-resolution visible light cloud images is characterized by: Obtain the ERA5 reanalysis data of the target area and the satellite medium- and high-resolution visible light cloud image data as the initial meteorological element data; perform data preprocessing on the initial meteorological element data to obtain the target meteorological element tensor, and divide the target meteorological element tensor into training data and test data; The training data is regarded as the grid node features evenly distributed in space, and then the corresponding regional multi-grid map is constructed according to the geographical area where the data is located, which is recorded as Mesh node. After the regional multi-grid map is constructed, the Grid node and the corresponding Mesh node in space are connected to obtain the RegionMesh multi-grid map; the RegionCast model is constructed in combination with the RegionMesh multi-grid map, and the regional meteorological element prediction model is obtained after iterative training of the RegionCast model by inputting the training data; Input the test data into the trained regional meteorological element prediction model to obtain the final meteorological element prediction result; The structure of the RegionCast model uses AutoEncoder as the basic framework, and adds an embedding layer, a Grid2Mesh layer, a multi-grid graph processor, a Mesh2Grid layer, and a reduction layer between the encoder and the decoder in sequence; The embedding layer is used to map the training data and the RegionMesh multi-grid map into a hidden embedding space; The Grid2Mesh layer uses the GNN message passing mechanism to enable the Grid nodes and Mesh nodes to update the embedding of unidirectional edges, and then updates the embedding of Mesh nodes through unidirectional edges; The multi-grid graph processor is capable of performing message passing between nodes on the RegionMesh multi-grid graph; The Mesh2Grid layer updates the embedding of the input latitude and longitude grid through the one-way edge from Mesh node to Grid node in the reverse direction; The restoration layer can map the processed embedding back to the original feature space, and predict the meteorological elements of each longitude and latitude grid point at the next moment by performing a reverse transformation on the embedding of the Grid node.

2. The AI ​​model meteorological element prediction method based on high-resolution visible light cloud image according to claim 1 is characterized in that: The meteorological elements in the ERA5 reanalysis data are atmospheric variables of temperature, westerly wind, northerly wind, potential, and humidity, and surface variables of two-meter temperature, two-meter humidity, 10-meter westerly wind, 10-meter northerly wind, total precipitation, and mean sea level pressure. The high-resolution visible light cloud map data are six visible light band cloud map data of three visible light bands, one near infrared band, and two infrared bands of the Sunflower-8 satellite.

3. The AI ​​model meteorological element prediction method based on high-resolution visible light cloud image according to claim 1 is characterized in that: The data preprocessing includes: uniformly adjusting the spatial resolution to 0.25° and the temporal resolution to 6 hours; performing mean normalization processing on the initial meteorological element data; and performing standardization processing using the Z-Score standardization method.

4. The AI ​​model meteorological element prediction method based on high-resolution visible light cloud image according to claim 1 is characterized in that: The RegionMesh multi-grid graph construction process is as follows: the longitude and latitude coordinates of the input Grid nodes are used to construct spatially uniformly distributed Mesh nodes by pooling. The characteristics of the Mesh nodes are the three coordinates of the spherical coordinate system where the nodes are located. Then, the links (edges) between the Mesh nodes are constructed, i.e., adjacent nodes and high-order neighbor nodes. The convolution kernel size of the RegionMesh multi-grid graph is set to 3, the step size is 3, and the number of layers is set to 4.

5. The AI ​​model meteorological element prediction method based on high-resolution visible light cloud image according to claim 1 is characterized in that: The embedding layer is a multi-layer perceptron structure, which maps the edges from Grid nodes to Mesh nodes, the edges from Mesh nodes to Grid nodes, and the edges of RegionMesh multi-grid graphs from low-dimensional attributes to a hidden embedding space of higher latitude to match the input training data; The Grid2Mesh layer uses GNN to extract relevant features and effectively capture the spatial relationship of meteorological elements. Through the message passing mechanism of GNN, the information of the Grid node is first transmitted to the edge from the Grid node to the Mesh node, and then integrated and transmitted to the corresponding Mesh node through these unidirectional edges; The multi-grid graph processor aggregates information of adjacent nodes and dynamically updates the feature representation of each node; The Mesh2Grid layer updates the embedding of the input latitude and longitude grid through the unidirectional edges from Mesh nodes to Grid nodes in the reverse direction. Through the unidirectional edges from Mesh nodes to Grid nodes, the Mesh2Grid layer ensures that each Grid node receives the latest information from the surrounding Mesh nodes.

6. The AI ​​model meteorological element prediction method based on high-resolution visible light cloud image according to claim 1 is characterized in that: The back propagation algorithm used for iterative training of the RegionCast model is Adam, the initial learning rate r of the network is set to 5e-4, and the learning rate decay factor is set to 0.05; The total number of training iterations is 100.

Images