Medium-term weather prediction method based on graph contrast learning
By constructing a graph structure with two layers of nodes, Grid and Mesh, and combining graph comparative learning, the problems of information loss and error accumulation at the regional scale in existing AI meteorological models are solved, achieving higher accuracy and more stable medium-term meteorological forecasts.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-25
- Publication Date
- 2026-04-07
AI Technical Summary
Existing AI meteorological models struggle to balance spatial resolution and computational complexity at the regional scale, leading to information loss and error accumulation. Furthermore, they neglect spatial consistency and temporal smoothness among meteorological elements, affecting the stability and accuracy of long-term time series forecasts.
A graph-based contrastive learning approach is adopted to construct a multi-layer graph structure with two layers of nodes: Grid and Mesh. The efficient propagation and multi-scale fusion of meteorological elements are achieved through a message passing mechanism, and a Mesh contrastive module is introduced to enhance the model's self-constraint capability.
It significantly improves the accuracy and stability of medium-term weather forecasts, enhances the robustness and generalization ability of the model, enables stable operation in complex climate scenarios, and provides higher-precision regional meteorological operational forecasts.
Smart Images
Figure CN121806157A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of artificial intelligence weather prediction, in particular to a medium-term weather prediction method based on graph contrastive learning. BACKGROUND
[0002] With the rapid development of artificial intelligence technology in the field of meteorology, deep learning-based weather prediction models have become an important supplement to numerical weather prediction (NWP). In particular, the GraphCast model proposed by Google DeepMind, by establishing a multi-layer message passing network on the spherical graph structure, realizes the end-to-end prediction of medium-term weather, and performs outstanding performance in global-scale temperature, wind field and potential height prediction. However, the existing AI weather model still has some shortcomings, which is difficult to meet the demand of high-precision prediction in regional scale or specific scene.
[0003] Firstly, the traditional AI weather model mainly adopts a global unified graph structure (such as the global spherical grid in GraphCast), which is difficult to balance spatial resolution and model calculation complexity in local or regional scale, resulting in information loss and error accumulation in regional weather evolution in medium scale (3-10 days). Secondly, the existing model mainly relies on supervised learning, ignoring the spatial consistency and temporal smoothness constraints between meteorological elements, and the model is prone to overfitting to local disturbances or abnormal noise, affecting the generalization ability and making it difficult to guarantee the stability and physical consistency of the model in long-term time series prediction. SUMMARY
[0004] In view of the problem that the traditional algorithm is insufficient in describing the regional spatial relationship and difficult to guarantee the stability and accuracy of medium-term prediction, the present application proposes a medium-term weather prediction method based on graph contrastive learning. A meteorological field modeling method based on multi-layer graph structure is proposed, which realizes efficient propagation and multi-scale fusion of meteorological elements in spatial topology by introducing Grid and Mesh double-layer nodes and their message passing mechanism, thereby improving the expression ability of regional weather process.
[0005] The method comprises the following steps: S1, obtaining a meteorological data set and preprocessing the meteorological data; S2, constructing a Mesh graph structure for training; S3, constructing a GCLCast model: The input data of the GCLCast model is first processed by a feature embedding module to generate a high-dimensional feature representation of the initial Grid node. The high-dimensional feature representation of the initial Grid node is divided into two branches. One branch is processed by a perturbation enhancement module to generate a perturbation-enhanced Grid node. The other branch of the initial Grid node high-dimensional feature representation is input into an encoder module, a graph message passing module, a decoder module, and a prediction output module in sequence together with the perturbation-enhanced Grid node. The GCLCast model further includes a Mesh comparison module that receives the output of the graph message passing module and generates a comparison constraint feedback to the encoder module and the graph message passing module. S4, training the GCLCast model according to the preprocessed data set and the Mesh graph structure; S5, inputting any continuous 2 time point meteorological field data into the trained GCLCast model to predict the distribution of meteorological elements in the future 72 to 192 hours.
[0006] Further, the feature embedding module maps the input data of the GCLCast model from the physical space to the high-dimensional space through an MLP unit to obtain a high-dimensional embedding representation. The time features corresponding to the input data of the GCLCast model are further encoded into an additional channel, which is spliced with the high-dimensional embedding representation to obtain the initial Grid node high-dimensional feature representation.
[0007] Further, the perturbation enhancement module receives the initial Grid node high-dimensional feature representation, and after random perturbation and Gaussian smoothing processing in sequence, generates a perturbation-enhanced Grid node.
[0008] Further, the encoder module receives the initial Grid node high-dimensional feature representation and the perturbation-enhanced Grid node, and then processes them through a Grid2Mesh graph neural network processing unit to generate the first updated Mesh node features and Mesh edge features of the initial Grid node high-dimensional feature representation and the perturbation-enhanced Grid node, respectively.
[0009] Further, the graph message passing module includes L layers of message passing units stacked in sequence, and the message passing units in each layer are connected through a residual connection. The graph message passing module receives the output of the encoder module to generate the second updated Mesh node features and Mesh edge features of the initial Grid node high-dimensional feature representation and the perturbation-enhanced Grid node, respectively.
[0010] Further, the decoder module includes a Mesh feature mapping unit, a Mesh2Grid reverse mapping unit, and a Grid feature integration unit. The decoder module receives the initial Grid node high-dimensional feature representation and the Mesh node features after the second update of the perturbation enhanced Grid node, and generates a final Grid embedding representation in combination with the Mesh graph structure.
[0011] Further, the prediction output module comprises a feature compression unit, a linear mapping unit and a multi-time output unit. The prediction output module receives the final Grid embedding representation and generates a final intermediate weather prediction result.
[0012] Further, the Mesh contrast module comprises an extraction and alignment unit, a neighborhood construction unit, a positive sample construction unit, a negative sample screening unit and a contrast loss calculation unit.
[0013] Further, the preprocessing comprises mean normalization and variance standardization.
[0014] The method has the following advantages: (1) The method takes the graph neural network as the core framework, combines feature embedding, graph message passing and Mesh layer contrast mechanism, realizes multi-scale fusion and spatial consistency modeling of meteorological elements, establishes an information flow path between Grid and Mesh nodes, and introduces a contrast learning enhanced model to improve the accuracy and stability of intermediate weather prediction.
[0015] (2) The GCLCast model proposed in the application realizes higher precision, stronger robustness and better generalization ability by introducing a graph contrast learning module. Compared with traditional deep learning models, the application has significant improvements in physical consistency constraints, spatio-temporal feature fusion and adaptability to observation noise, and can stably operate in complex climate scenarios, providing a new, reliable and scalable technical solution for regional weather business prediction based on AI. BRIEF DESCRIPTION OF DRAWINGS
[0016] Figure 1 The structure diagram of the GCLCast model of the application; Figure 2 The RMSE column contrast graph of key meteorological elements of the GCLCast model of the application on the test set, wherein the abscissa represents the surface variable and the ordinate represents the value of RMSE; Figure 3 The MAE column contrast graph of key meteorological elements of the GCLCast model of the application on the test set, wherein the abscissa represents the surface variable and the ordinate represents the value of MAE; Figure 4Figures of RMSE curves of key meteorological elements of the GCLCast model on the test set, in which the abscissa represents the surface variable, and the ordinate represents the value of RMSE. DETAILED DESCRIPTION
[0017] The technical solutions of the present application will be described clearly and completely below in conjunction with the drawings. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.
[0018] Embodiment 1, The present embodiment provides a mid-term weather prediction method based on graph contrastive learning.
[0019] S1, obtain a meteorological data set, and preprocess the meteorological data; The present application uses the 5th generation reanalysis data product ERA5 of the European Centre for Medium-Range Weather Forecasts (ECMWF) as the main training data (meteorological data set), and combines satellite observation data (such as visible and infrared bands) to form a multi-meteorological element input set. Each meteorological element includes temperature, humidity, wind field, potential height, cumulative precipitation, etc., with a spatial resolution of 0.25° and a time resolution of 6 hours. The experimental area is selected as 60°-0°N, 80°-140°E. To improve the stability of the model, all input elements (meteorological data set) are preprocessed: mean normalization and variance standardization processing, and a standardized file is generated. At the same time, according to the pressure layer selection strategy, the three-dimensional field data of the key pressure layer (such as 1000 hPa, 850 hPa, 500 hPa and 250 hPa, etc.) is extracted to construct a high-dimensional input tensor, which provides a consistent spatiotemporal feature basis for graph structure modeling. The meteorological parameters of ERA5 are shown in Table 1 below: Table 1
[0020] S2, construct a Mesh graph structure for training; On the basis of the aligned meteorological element data, the present application regards the input tensor as a spatial grid (Grid node), and constructs a regionalized multi-layer graph structure according to the geographical coordinates and latitude and longitude range. By aggregating the geographical location features of the Grid nodes, the corresponding Mesh nodes are generated, forming a Grid-Mesh double-layer structure. In this structure, the features of the Mesh nodes are defined by the three components of the spherical coordinates, and one-way edges are established between the Grid nodes and the Mesh nodes to transfer local meteorological information, and bidirectional edges are established between the Mesh nodes to capture regional climate dynamic processes. The features of each edge include the distance and coordinate difference between the nodes in the spherical coordinate system, thereby realizing the spatial topological representation of meteorological elements.
[0021] S3. Construct the GCLCast model: like Figure 1 As shown, the input data of the GCLCast model first passes through the feature embedding module to generate the initial high-dimensional feature representation of the Grid node. The initial high-dimensional feature representation of the Grid node is divided into two branches. One branch is processed by the perturbation enhancement module to generate the perturbation enhanced Grid node. The initial high-dimensional feature representation of the Grid node and the perturbation enhanced Grid node of the other branch are input into the encoder module, graph message passing module, decoder module and prediction output module in sequence. The GCLCast model also includes a Mesh contrast module, which receives the output of the graph messaging module and generates contrast constraints that are fed back to the encoder module and the graph messaging module.
[0022] The Feature Embedding module first performs embedding mapping on the preprocessed meteorological elements and static geographic features (such as elevation, land-sea distribution, etc.) of the Grid nodes. This involves using a Multilayer Linear Mapping Network (MLP) to map the original input (data from two consecutive time points) from the physical space to a high-dimensional embedding space, and standardizing the meteorological variables to enable the model to uniformly handle input features of different dimensions. Secondly, this module encodes temporal features as an additional channel to enhance the model's adaptability to different time periods. The data flow is as follows: Original Grid node meteorological elements and static geographic features → MLP mapping and standardization → High-dimensional embedding representation → Concatenated temporal feature encoding → Output initial high-dimensional feature representation of the Grid nodes for subsequent perturbation enhancement and encoder modules.
[0023] The Spatial Augmentation module introduces controlled perturbations into the original Grid node embedding representation at the feature level to simulate observation noise and spatial correlation uncertainty, thereby improving the model's robustness to anomalous perturbations in medium-term weather forecasts.
[0024] Specifically, this module applies random perturbations to the initial high-dimensional feature representation of the Grid nodes and constrains the spatial continuity of the perturbations through a two-dimensional Gaussian smoothing operation, ensuring that the enhanced features conform to the statistical characteristics of real meteorological fields in both amplitude and spatial structure. The perturbation enhancement module only operates on the feature layer of the Grid nodes and does not introduce additional structural parameters.
[0025] The data flow is as follows: initial high-dimensional feature representation of Grid nodes → random perturbation injection → Gaussian smoothing → output perturbation to enhance Grid node features, which are used to input the encoder module and subsequent network structure in parallel with the original Grid node features.
[0026] The encoder module (Encoder) is composed of a Grid2Mesh graph neural network unit, which is used to map the meteorological features on the Grid nodes to the Mesh nodes, realizing the feature conversion from the regular grid space to the irregular graph structure space.
[0027] In this module, the original Grid node features and the perturbation enhanced Grid node features are respectively taken as inputs, and processed through the Grid2Mesh mapping unit with shared parameters. According to the predefined Grid-Mesh connection relationship, the Grid node features are aggregated to the corresponding Mesh nodes according to the neighborhood, and the related Mesh edge features are updated synchronously to depict the local spatial relationship.
[0028] The encoder module does not perform write-back update on the Grid node features, and its output is the corresponding Mesh node features and Mesh edge features under two views, providing consistent intermediate representations for the subsequent graph message passing module and Mesh contrastive module.
[0029] The core part of the graph message passing module (Graph Processor) model is composed of L (L can be set by the person skilled in the art according to actual needs) consecutive stacked message passing units, including a Mesh edge feature update unit, a Mesh node feature update unit and a cross-layer residual connection structure, which is used to realize multi-step feature propagation and spatial information fusion within the Mesh graph. First, the Mesh edge feature update unit performs nonlinear transformation on the edge features according to the current features of the adjacent nodes, thereby extracting the local spatial relationship between nodes; then, the updated edge features are input into the Mesh node feature update unit, and the information of multiple adjacent edges is aggregated to the target Mesh node through neighborhood aggregation operation, realizing the representation enhancement of regional dynamic process. The above two units are iteratively executed in a stacked manner, and a residual path is introduced between each layer to maintain gradient stability and feature consistency. The data flow of the whole module is: Mesh node features and Mesh edge features → edge feature update unit → node feature update unit → iterative multi-layer → maintain gradient and feature stability between each layer through residual path → output updated Mesh node features and edge features (initial Grid node high-dimensional feature representation and perturbation enhanced Grid node second updated Mesh node features and Mesh edge features), which are used for subsequent Mesh contrastive module and decoder module.
[0030] The Mesh contrastive module (Mesh Contrastive) is used to introduce self-supervised contrastive constraints at the Mesh node layer to enhance the modeling ability of the model for spatial consistency and perturbation invariance.
[0031] The module receives the Mesh node features from the original view and the perturbed enhanced view. After feature normalization and dimension alignment, the node neighborhood relationship is constructed based on the Mesh graph structure. The positive samples are composed of the features of the same Mesh node in different views and their neighborhood aggregated features. The negative samples are selected from non-neighbor Mesh nodes.
[0032] The contrastive loss serves as an auxiliary training target and participates in the model optimization process together with the prediction error loss. Through backpropagation, it indirectly constrains the parameter updates in the encoder module and the graph message passing module, thereby improving the robustness and generalization ability of the overall feature representation.
[0033] The decoder module (Decoder) is composed of a Mesh feature mapping unit, a Mesh2Grid reverse mapping unit, and a Grid feature integration unit. It is responsible for restoring the high-dimensional representation of Mesh nodes to the original grid structure. First, the updated features of Mesh nodes enter the Mesh feature mapping unit, which performs nonlinear mapping through a multi-layer perceptron to extract intermediate features that can be transmitted downward. Then, the mapped features are input into the Mesh2Grid reverse mapping unit, which transfers and aggregates the features of each Mesh node to its corresponding multiple Grid nodes according to the pre-set Mesh-Grid connection relationship. Finally, the Grid feature integration unit updates the aggregated features from different Mesh nodes and the Grid node's own features, forming the final Grid embedding representation with spatial continuity and physical consistency. The data flow is: Mesh node features → Mesh feature mapping → Mesh2Grid reverse mapping → Grid feature integration → output final Grid embedding representation → input prediction output module.
[0034] Prediction output module: The prediction output module (Linear Predict) is used to map the high-dimensional embedding representation of Grid nodes output by the decoder back to the physical meteorological element space, generating multi-variable and multi-time step intermediate meteorological prediction results.
[0035] First, the feature compression unit normalizes the dimension of the Grid embedding representation. Then, the linear mapping unit converts the features into corresponding meteorological element prediction values, including temperature, wind speed, humidity, potential height, and cumulative precipitation. The multi-time step prediction results are output in a unified mapping structure at once and are aligned with the ERA5 reanalysis data for calculating the supervised learning loss.
[0036] The prediction output module does not participate in graph structure calculation and only undertakes the function of mapping features to physical quantities, maintaining a clear division of labor with the decoder module in terms of function.
[0037] S4, training the GCLCast model according to the preprocessed data set and the Mesh graph structure; S5, inputting meteorological field data of any continuous two time points into the trained GCLCast model to predict the distribution of meteorological elements in the future 72 to 192 hours; By using the trained GCLCast model, meteorological field data at any time can be inputted to perform continuous time series reasoning, and the distribution of meteorological elements at multiple time steps in the future can be outputted. The GCLCast model can realize stable prediction for up to 8 days or more, and maintain high spatial consistency and physical constraint characteristics. The evaluation results of the ERA5 reanalysis data show that the present application significantly reduces the prediction error in medium-term weather prediction and improves the ability to describe complex weather processes.
[0038] Embodiment 2, This embodiment is a further limitation of embodiment 1. The model trained based on the ERA5 meteorological data set from 2016 to 2023 is used for comparative evaluation, and the test set is the first quarter data of 2025. Each prediction starts from a certain time point and continuously predicts for 8 days with a time interval of 6 hours. The ERA5 reanalysis data and high-precision cloud image data are compared with the true value, and RMSE and MAE are used as evaluation indicators. The comparison results are shown in Figure 2 、 3 and 4. From the comparison results, under the same data set and computing resource conditions, the prediction accuracy of the GCLCast model of the present application on the main meteorological variables is better than that of the FourCastNet and GraphCast models. Taking the surface temperature (t2m) as an example, the RMSE of GCLCast is 3.638, which is decreased by about 11.3% compared with FourCastNet (4.104) and decreased by about 4.2% compared with GraphCast (3.798), and the prediction accuracy rate of temperature error within ±2°C reaches 66.34%, which is significantly better than the comparison models. At the same time, in the prediction of precipitation elements, the RMSE of GCLCast is 2.119, which is lower than that of FourCastNet (2.212) and GraphCast (2.128), and the TS score reaches 0.586, which shows significant improvement in the identification of mesoscale precipitation and the capture of severe convective weather. In addition, in the prediction of upper air elements such as geopotential height (z) and wind field (u, v), GCLCast also shows lower bias and smoother time series evolution trend, proving its advantages in spatial continuity and physical consistency.
[0039] The comprehensive evaluation result shows that the GCLCast model proposed in the application realizes higher precision, stronger robustness and better generalization ability by introducing a graph contrast learning module. Compared with the traditional deep learning model, the application has significant improvement in physical consistency constraint, spatiotemporal feature fusion and adaptability to observation noise, and can stably run under complex climate scenarios, providing a new, reliable and scalable technical solution for regional meteorological business prediction based on AI.
Claims
1. A medium-term weather forecasting method based on graph comparison learning, characterized in that, The method includes the following steps: S1. Obtain the meteorological dataset and preprocess the meteorological data; S2. Construct the Mesh graph structure for training; S3. Construct the GCLCast model: The input data of the GCLCast model first passes through the feature embedding module to generate the high-dimensional feature representation of the initial Grid node. The high-dimensional feature representation of the initial Grid node is divided into two branches. One branch is processed by the perturbation enhancement module to generate the perturbation enhanced Grid node. The high-dimensional feature representation of the initial Grid node and the perturbation enhanced Grid node of the other branch are input into the encoder module, graph message passing module, decoder module and prediction output module in sequence. The GCLCast model also includes a Mesh contrast module, which receives the output of the graph message passing module and generates contrast constraints that are fed back to the encoder module and the graph message passing module. S4. Train the GCLCast model based on the preprocessed dataset and Mesh graph structure; S5. Input meteorological field data for any two consecutive moments into the trained GCLCast model to predict the distribution of meteorological elements for the next 72 to 192 hours.
2. The medium-term weather forecasting method based on graph contrastive learning according to claim 1, characterized in that, The feature embedding module maps the input data of the GCLCast model from the physical space to a high-dimensional space through the MLP unit to obtain a high-dimensional embedding representation; The temporal features corresponding to the input data of the GCLCast model are then encoded into additional channels and concatenated with the high-dimensional embedding representation to obtain the initial high-dimensional feature representation of the Grid nodes.
3. The medium-term weather forecasting method based on graph contrastive learning according to claim 2, characterized in that, The perturbation enhancement module receives the initial high-dimensional feature representation of the Grid node, performs random perturbation and Gaussian smoothing processing in sequence, and then generates perturbation-enhanced Grid nodes.
4. The medium-term weather forecasting method based on graph comparison learning according to claim 3, characterized in that, After receiving the initial high-dimensional feature representation of the Grid node and the perturbation-enhanced Grid node, the encoder module processes them through the Grid2Mesh graph neural network processing unit to generate the initial high-dimensional feature representation of the Grid node and the Mesh node features and Mesh edge features after the first update of the perturbation-enhanced Grid node.
5. The medium-term weather forecasting method based on graph comparison learning according to claim 4, characterized in that, The graph message passing module contains L layers of continuously stacked message passing units, with each layer of message passing units connected by residuals. The graph message passing module receives the output of the encoder module and generates the initial high-dimensional feature representation of the Grid node and the Mesh node features and Mesh edge features of the perturbation-enhanced Grid node after the second update.
6. The medium-term weather forecasting method based on graph contrast learning according to claim 5, characterized in that, The decoder module includes: a Mesh feature mapping unit, a Mesh2Grid reverse mapping unit, and a Grid feature integration unit; The decoder module receives the initial high-dimensional feature representation of the Grid node and the Mesh node features after the second update of the perturbation-enhanced Grid node, and then combines them with the Mesh graph structure to generate the final Grid embedding representation.
7. The medium-term weather forecasting method based on graph contrastive learning according to claim 6, characterized in that, The prediction output module includes: a feature compression unit, a linear mapping unit, and a multi-time-effect output unit; The forecast output module receives the final Grid embedding representation and generates the final medium-term weather forecast result.
8. The medium-term weather forecasting method based on graph contrastive learning according to claim 7, characterized in that, The Mesh contrast module includes: extraction and alignment unit, neighborhood construction unit, positive sample construction unit, negative sample selection unit, and contrast loss calculation unit.
9. The medium-term weather forecasting method based on graph contrastive learning according to claim 8, characterized in that, The preprocessing includes mean normalization and variance standardization.