A weather forecasting method based on hierarchical graph neural network and latent variable injection

Abstract

The application discloses a meteorological prediction method based on hierarchical graph neural network and hidden variable injection, and belongs to the technical field of meteorological science and high-performance computing, which comprises the following steps: obtaining real-time global high-resolution meteorological reanalysis data, preprocessing to obtain a global meteorological state tensor, constructing a global meteorological probability prediction model, extracting and integrating the global meteorological state tensor step by step to obtain the coarsest level node features, performing feature extraction, distribution prediction and reparameterization sampling on the coarsest level node to obtain hidden variables, injecting the hidden variables into the coarsest level node features in a non-intrusive residual superposition mode to obtain fused large-scale circulation features, and then performing step-by-step optimization transmission on the fused large-scale circulation features to obtain meteorological prediction of the next moment. The application can more efficiently explore multi-modal distribution, optimize the probability distribution, greatly reduce the calculation cost, and improve the clarity, uncertainty quantification capability and reasoning efficiency of global meteorological prediction.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence, deep learning, meteorological science and high-performance computing, specifically to a weather forecasting method and system based on hierarchical graph neural networks and latent variable injection. Background Technology

[0002] The key to meteorological probabilistic forecasting technology lies in redesigning deterministic forecasting models using latent variable generation methods. This enables the models to quantify uncertainties and fit multimodal distributions, thereby improving the clarity and reliability of long-term forecasts. Under complex conditions such as long-term evolution, sudden extreme weather events, and significant atmospheric chaos, the conditional distributions predicted by deterministic models often undergo mean regression. Therefore, traditional deep learning meteorological forecasting algorithms are susceptible to oversmoothing effects, leading to blurred forecast fields, loss of high-frequency textures, or missed extreme events.

[0003] Existing graph neural network models perform well in short-term forecasts, but medium- and long-term forecasts are still affected by the statistical characteristics of the mean squared error loss function, causing the forecast results to tend to be climatically average. Summary of the Invention

[0004] To address these challenges, a weather forecasting method based on hierarchical graph neural networks and latent variable injection is proposed. This method combines deterministic large-scale circulation characteristics with stochastic latent variable perturbations to achieve targeted probabilistic feature learning. This improvement enhances the model's ability to capture the possibility of chaotic atmospheric evolution and improves the forecast's dispersion and error consistency.

[0005] To achieve the above objectives, this invention presents a weather forecasting method based on hierarchical graph neural networks and latent variable injection, characterized in that the method includes the following steps:

[0006] S1: Acquire real-time global high-resolution meteorological reanalysis data and preprocess it to obtain the global meteorological state tensor; the global meteorological state tensor includes geopotential height, temperature, and wind speed variables;

[0007] S2: Based on a hierarchical graph neural network, a global meteorological probability forecast model is constructed to obtain the meteorological forecast field at the next moment. The global meteorological probability forecast model is constructed with a hierarchical graph communication topology as its core. The hierarchical graph communication topology includes latitude and longitude grids and multi-level icosahedral grids. The core components of the multi-level icosahedral grids include a latent variable injection module, a multi-level graph processor evolution module, and a decoding module.

[0008] S21: Input the global meteorological state tensor into the hierarchical graph communication topology to obtain the coarsest level node features;

[0009] S22: Input the features of the coarsest-level nodes into the latent variable injection module to obtain latent variables; the latent variable injection module includes a feature extraction unit, a distribution prediction unit, and a reparameterized sampling unit; the latent variable injection module embeds the coarsest level of the multi-level icosahedral mesh;

[0010] S23: The latent variables and the coarsest-level node features are fused using a non-intrusive residual stacking method to obtain fused large-scale circulation features;

[0011] S24: The fused large-scale circulation features are input into the multi-layer graph processor evolution module to obtain the finest-level features; the multi-layer graph processor evolution module is constructed with stacked graph neural network blocks as the core; the multi-layer graph processor evolution module runs through the entire multi-level icosahedral mesh;

[0012] S25: Input the finest level features into the decoding module to obtain the weather forecast field for the next moment; the decoding module is embedded at the junction of the finest level of the icosahedral grid and the latitude and longitude grid; the decoding module is built with the decoder as the core;

[0013] S26: Input the global meteorological state tensor into the global meteorological probability forecast model and train and optimize the parameters through dual-path forward propagation, loss calculation, and back propagation to obtain the optimized global meteorological probability forecast model;

[0014] S3: Collect real-time global high-resolution meteorological reanalysis data, preprocess it, and input it into the optimized global meteorological probability forecast model to obtain the meteorological forecast field for the next moment.

[0015] Preferably, in step S21, the global meteorological state tensor is input into the hierarchical graph communication topology to obtain the coarsest-level node features. The specific process is as follows:

[0016] The global meteorological state tensor is input into the latitude and longitude grid for mapping to obtain grid node features. The grid node features are then transmitted to the finest level of the multi-level icosahedral grid using a bipartite graph message passing method to obtain fine-level meteorological information. The fine-level meteorological information is then extracted and integrated level by level using a hierarchical coding mechanism and then transmitted to the coarsest level of the icosahedral grid to obtain the coarsest level node features.

[0017] Preferably, in step S22, the features of the coarsest-level nodes are input into the latent variable injection module to obtain latent variables. The specific process is as follows:

[0018] The features of the coarsest-level nodes are input into the feature extraction unit for feature enhancement to obtain enhanced feature vectors. The enhanced feature vectors are then input into the distribution prediction unit to obtain the mean and variance of the latent variables. Finally, the mean and variance of the latent variables are input into the reparameterized sampling unit for random sampling to obtain the latent variables.

[0019] Preferably, in step S23, the fused large-scale circulation features are obtained, and the specific process is as follows:

[0020] The latent variables and the coarsest-level node features are adjusted to ensure consistency in feature dimension and spatial scale. The latent variables are then superimposed as random perturbation terms onto the coarsest-level node features to obtain fused large-scale circulation features. The calculation formula is as follows:

[0021]

[0022] in, Represents latent variables; This represents the feature of the coarsest level node; This indicates the integration of large-scale circulation characteristics.

[0023] Preferably, in step S24, the fused large-scale circulation features are input into the multi-layer graph processor evolution module to obtain the finest-level features. The specific process is as follows:

[0024] The large-scale circulation features are input into stacked graph neural network blocks. Each graph neural network block performs three steps in sequence: edge update, node aggregation, and state update. Through these three steps, the large-scale circulation features are passed from the coarsest level to the finest level to obtain the finest level features.

[0025] Preferably, in step S25, the finest-level features are input into the decoding module to obtain the meteorological forecast field for the next moment. The specific process is as follows:

[0026] The finest-level features are input into the decoder, and the decoding mechanism maps the finest-level features back to the latitude and longitude grid. The features are then processed through the projection layer to finally obtain the weather forecast field for the next moment.

[0027] Preferably, in step S26, the global meteorological state tensor is input into the global meteorological probability forecasting model and trained and optimized through dual-path forward propagation, loss calculation, and backpropagation to obtain an optimized global meteorological probability forecasting model. The specific process is as follows:

[0028] The CRPS approximate loss function based on dual-path forward propagation is used to calculate the total model loss by maximizing the difference between two samples in the global meteorological state tensor to approximate continuous hierarchical probability scoring. Backpropagation and weight optimization are performed based on the total loss, and training is continuously iterated until the total model loss value tends to stabilize and the forecast accuracy and ensemble dispersion reach the preset standards. Training is then stopped to obtain the optimized global meteorological probability forecast model. The CRPS approximate loss function based on dual-path forward propagation includes ensemble average reconstruction loss, divergence regularization term, and diversity loss term.

[0029] The formula for calculating the average reconstruction loss of the ensemble is as follows:

[0030]

[0031] in, This represents the average reconstruction loss of the set; Indicates mean square error; This represents the first predicted sample generated by the two-way forward propagation; This represents the second predicted sample generated by the two-path forward propagation; This represents the true values ​​of global high-resolution meteorological reanalysis data;

[0032] The formula for calculating the divergence regularization term is as follows:

[0033]

[0034] in, Indicates the KL divergence loss; Represents the squared term of the mean; Represents the variance term;

[0035] The formula for calculating the diversity loss term is as follows:

[0036]

[0037] in, Represents the diversity loss term; This represents the diversity loss weighting coefficient;

[0038] The formula for the total loss of the model is as follows:

[0039]

[0040] in, This represents the total loss of the model; This represents the weight coefficient of the KL divergence regularization term.

[0041] The advantages of this invention compared to existing technologies are as follows:

[0042] This invention utilizes a variational autoencoder to inject latent variables at the coarsest level of the hierarchical graph, simulating the initial uncertainty of large-scale circulation. It effectively solves the problem of excessive smoothing caused by mean regression in traditional deterministic models in long-term forecasts, and ensures that the generated forecast field retains clear high-frequency texture and physical structure.

[0043] This invention introduces a direct linear injection technique for latent variables, employing a parameter-free residual superposition method to directly integrate random perturbations into physical features. Compared to complex feature modulation networks, this method avoids the introduction of additional parameters and the potential gradient vanishing problem, ensuring direct coupling between latent variables and original features on the numerical scale, enabling the model to explore multimodal distributions more efficiently.

[0044] This invention optimizes the probability distribution without generating a large-scale ensemble during training by designing a dual-path forward propagation and CRPS approximation training strategy, which significantly reduces the computational cost of the probability model. The generated ensemble forecast has good dispersion and error consistency, and can accurately quantify forecast uncertainty. Attached Figure Description

[0045] Figure 1 This is a flowchart of a weather forecasting method based on hierarchical graph neural networks and latent variable injection according to the present invention.

[0046] Figure 2 This is a schematic diagram illustrating the training strategy of a weather forecasting method based on hierarchical graph neural networks and latent variable injection, as described in this invention.

[0047] Figure 3 This is a visualization comparison of the long-term forecast performance of the meteorological forecasting method based on hierarchical graph neural network and latent variable injection proposed in this invention with that of existing deterministic baseline models. Detailed Implementation

[0048] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments and accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] like Figure 1 As shown, the meteorological forecasting method based on hierarchical graph neural networks and latent variable injection proposed in this invention includes the following steps:

[0050] S1: Acquire real-time global high-resolution meteorological reanalysis data and preprocess it to obtain the global meteorological state tensor; the global meteorological state tensor includes geopotential height, temperature, and wind speed variables;

[0051] The dataset used is the 1.5° (121×240 global dataset) version of weatherbench2.

[0052] S2: Based on a hierarchical graph neural network, a global meteorological probability forecast model is constructed to obtain the meteorological forecast field at the next moment. The global meteorological probability forecast model is constructed with a hierarchical graph communication topology as its core. The hierarchical graph communication topology includes latitude and longitude grids and multi-level icosahedral grids. The core components of the multi-level icosahedral grids include a latent variable injection module, a multi-level graph processor evolution module, and a decoding module.

[0053] S21: Input the global meteorological state tensor into the hierarchical graph communication topology to obtain the coarsest level node features;

[0054] This involves constructing a hierarchical graph communication topology module that includes latitude and longitude grids and multi-level icosahedral grids. First, the global meteorological state tensor, containing variables such as geopotential height, temperature, and wind speed, is mapped to grid node features. Then, through Grid2Mesh bipartite graph message passing, the grid node information is aggregated to the finest level of the icosahedral grid. Finally, using a hierarchical encoding mechanism, through multi-level pooling or cross-layer connections, the information is further extracted to the coarsest level. The coarsest level nodes carry large-scale global circulation characteristics, resulting in the coarsest level node features.

[0055] S22: Input the features of the coarsest-level nodes into the latent variable injection module to obtain latent variables; the latent variable injection module includes a feature extraction unit, a distribution prediction unit, and a reparameterized sampling unit; the latent variable injection module embeds the coarsest level of the multi-level icosahedral mesh;

[0056] The workflow of the latent variable injection module consists of three parts: a feature extraction unit, a distribution prediction unit, and a reparameterized sampling unit.

[0057] The feature extraction unit refers to the unit that extracts the large-scale circulation features of the coarsest level (Level 0) grid nodes after the graph neural network encoder completes the Grid2Mesh bipartite graph aggregation and multi-level mesh upsampling operations. In this embodiment, the coarsest level is constructed based on the 12 base vertices of a regular icosahedron. This level of nodes has the largest receptive field and highly condenses long-wave circulation information at the global planetary wave scale and large-scale climate background features. The feature extraction process is implemented through a multilayer perceptron (MLP) and pooling operations to ensure that key high-dimensional semantic information transferred from the fine-grained grid is preserved.

[0058] The distributed prediction unit refers to using two parallel multilayer perceptrons (MLPs) as prediction heads, respectively inputting features from the coarsest level nodes. Output hidden variables Parameters of the Gaussian posterior distribution it follows: mean vector Sum of logarithmic variance vector Each MLP consists of an input layer, a SiLU activation function layer, and an output linear layer. To ensure the stability of numerical computation and prevent excessively large variance predictions from causing gradient explosion or excessively small variance predictions from causing numerical underflow, the log-variance vector is... Perform hard truncation (Clamp) to restrict its value range to [-10, 10];

[0059] The reparameterized sampling unit refers to the use of the reparameterization trick to introduce random noise that follows a standard normal distribution. The latent variables with gradient backpropagation capability are calculated. This step transforms the deterministic characteristic representation into random variables, thereby simulating the inherent randomness of the initial atmospheric field and physical processes in the latent space. The calculation formula is as follows:

[0060]

[0061] in, Represents latent variables; Indicates random noise; Indicates the mean of the distribution; Indicates the variance of the distribution; This represents a matrix with a mean of 0 and a covariance equal to the identity matrix. The standard normal distribution;

[0062] S23: The latent variables and the coarsest-level node features are fused using a non-intrusive residual stacking method to obtain fused large-scale circulation features;

[0063] The direct linear injection of latent variables employs a non-intrusive residual superposition method to directly inject the generated latent variables into the large-scale circulation features of the coarsest layer features, thereby achieving parameter-free introduction of uncertainty.

[0064] The direct linear injection specifically includes the following steps:

[0065] Direct residual stacking: This method directly stacks the generated latent variables without defining additional weight parameters. As a random perturbation term, it is superimposed on the features of the coarsest-level nodes. Above. The calculation formula is:

[0066]

[0067] in, Represents latent variables; This represents the feature of the coarsest level node; This indicates the fusion of large-scale circulation characteristics; This operation represents element-wise addition. While preserving the original physical feature boundaries and structure, it linearly injects random perturbations. This parameter-free direct injection method avoids the gradient vanishing problem caused by improper weight parameter training, ensuring direct coupling between latent variables and physical features on a numerical scale. This allows the model to effectively perceive multimodal distributions in the early stages of training, preventing the posterior distribution collapse problem common in variational autoencoders.

[0068] S24: The large-scale circulation features are input into the multi-layer graph processor evolution module to obtain the finest layer node features; the multi-layer graph processor evolution module is constructed with stacked graph neural network blocks as the core; the multi-layer graph processor evolution module runs through the entire multi-level icosahedral mesh.

[0069] Multilayer graph processor evolution module: will integrate large-scale circulation features Input includes A graph processor with 16 layers (e.g., 16 layers). The graph processor consists of stacked graph neural network blocks, each containing three sub-steps: edge update, node aggregation, and state update. During evolution, the random latent variables injected in the coarsest layer (Level 0) are gradually passed down through cross-layer and same-layer connections between grids, spreading like waves to the finest layer (Level 5), thereby influencing the global feature distribution and obtaining the finest layer node features;

[0070] S25: Input the finest layer node features into the decoding module to obtain the weather forecast field for the next moment; the decoding module is embedded at the junction of the finest layer of the icosahedral mesh and the latitude and longitude mesh; the decoding module is built with the decoder as the core;

[0071] Decoding Module: After evolution, the finest layer node features are aggregated to latitude and longitude nodes through bipartite graph edges. To recover high-resolution meteorological field details, the finest layer node features are projected back into the physical variable space by a decoder composed of a multilayer perceptron, outputting the prediction increment. The final weather forecast field value for the next moment. From the current state Adding this to the predicted increment yields the (residual connection):

[0072]

[0073] in, This indicates the weather forecast field for the next moment. This represents the current meteorological state field. Represents the predicted increment field;

[0074] S26: Input the global meteorological state tensor into the global meteorological probability forecast model and train and optimize the parameters through dual-path forward propagation, loss calculation, and back propagation to obtain the optimized global meteorological probability forecast model;

[0075] The paper proposes a CRPS approximate loss function based on dual-path forward propagation. This function maximizes the approximate continuous hierarchical probability score by maximizing the difference between two samples, thereby improving the dispersion and skill consistency of the ensemble predictions generated by the model. Figure 2 As shown;

[0076] The CRPS approximate loss function based on dual-path forward propagation includes ensemble average reconstruction loss, KL divergence regularization term, and diversity loss term. The formula for calculating the total loss function of the model is as follows:

[0077]

[0078] in, This represents the ensemble average reconstruction loss, used to ensure forecast accuracy. Unlike traditional single-sample loss, this invention calculates the mean squared error (MSE) between the means of two random samples and the true values, effectively filtering out random noise and focusing on deterministic forecasting techniques.

[0079]

[0080] in, This represents the KL divergence loss, used to constrain the latent space distribution. This term calculates the divergence between the predicted posterior distribution and the standard normal prior distribution, preventing the latent variable encoding from including too much non-random information or causing overfitting.

[0081]

[0082] in, This represents the diversity loss, based on the energy fraction principle, and is used to maximize the distance between two samples. This term forces the model to generate two samples... and While maintaining accuracy, the resulting set should be as different as possible to ensure sufficient dispersion to cover the possibilities of real atmospheric evolution.

[0083]

[0084] in, Usually set to Magnitude It is set to around 0.1 to balance accuracy and diversity.

[0085] The global meteorological state tensor is input into the model, and the total loss function value is calculated using a dual-path sampling strategy. Backpropagation is performed, and the connection weights are optimized using a selected optimizer and a cosine annealing learning rate scheduling strategy. Random perturbations are introduced using a latent variable direct injection mechanism to break the mean regression limitation of the deterministic model. Finally, a meteorological probability forecast model based on hierarchical graph neural network and latent variable direct linear injection is obtained.

[0086] The training process employs an end-to-end probabilistic optimization strategy: During training, a dual-path training strategy utilizes diversity loss (L1 distance) to drive the model to explore multimodal distributions in the latent space, while simultaneously constraining the accuracy of the ensemble mean through reconstruction loss (MSE). Due to the direct linear injection mechanism, the model does not require a complex weight warm-up phase, effectively balancing deterministic accuracy and probabilistic dispersion in the early stages of training, achieving rapid convergence and robust probabilistic prediction capabilities.

[0087] S3: Collect real-time global high-resolution meteorological reanalysis data, preprocess it, and input it into the optimized global meteorological probability forecast model to obtain the meteorological forecast field for the next moment.

[0088] A comparison of the visualization effects of this invention with existing deterministic baseline models in long-term forecasting, such as... Figure 3 As shown, this direct linear injection method can preserve the original structure of physical features to the greatest extent, solve the problem of gradual loss of details in deterministic models during long-term recursion, and effectively avoid the phenomenon of collapse of the posterior distribution of latent variables.

[0089] Although preferred embodiments of this application have been described, those skilled in the art, once they learn the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0090] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if these modifications and variations of this application fall within the scope of the claims of this application and their equivalents, this application also intends to include these modifications and variations.

Claims

1. A weather forecasting method based on hierarchical graph neural networks and latent variable injection, characterized in that, include: S1: Acquire real-time global high-resolution meteorological reanalysis data and preprocess it to obtain the global meteorological state tensor; the global meteorological state tensor includes geopotential height, temperature, and wind speed variables; S2: Based on a hierarchical graph neural network, a global meteorological probability forecast model is constructed to obtain the meteorological forecast field at the next moment. The global meteorological probability forecast model is constructed with a hierarchical graph communication topology as its core. The hierarchical graph communication topology includes latitude and longitude grids and multi-level icosahedral grids. The core components of the multi-level icosahedral grids include a latent variable injection module, a multi-level graph processor evolution module, and a decoding module. S21: Input the global meteorological state tensor into the hierarchical graph communication topology to obtain the coarsest level node features; S22: Input the features of the coarsest-level nodes into the latent variable injection module to obtain latent variables; The latent variable injection module includes a feature extraction unit, a distribution prediction unit, and a reparameterized sampling unit; The latent variable injection module is embedded in the coarsest level of a multi-level icosahedral mesh; The features of the coarsest-level nodes are input into the feature extraction unit for feature enhancement to obtain enhanced feature vectors. The enhanced feature vectors are then input into the distribution prediction unit to obtain the mean and variance of the latent variables. Finally, the mean and variance of the latent variables are input into the reparameterized sampling unit for random sampling to obtain the latent variables. The feature enhancement process is implemented through a multilayer perceptron (MLP) and pooling operations. S23: The latent variables and the coarsest-level node features are fused using a non-intrusive residual stacking method to obtain fused large-scale circulation features; S24: The fused large-scale circulation features are input into the multi-layer graph processor evolution module to obtain the finest-level features; the multi-layer graph processor evolution module is constructed with stacked graph neural network blocks as the core; the multi-layer graph processor evolution module runs through the entire multi-level icosahedral mesh; S25: Input the finest level features into the decoding module to obtain the weather forecast field for the next moment; the decoding module is embedded at the junction of the finest level of the icosahedral grid and the latitude and longitude grid; the decoding module is built with the decoder as the core; S26: Input the global meteorological state tensor into the global meteorological probability forecast model and train and optimize the parameters through dual-path forward propagation, loss calculation, and back propagation to obtain the optimized global meteorological probability forecast model; S3: Collect real-time global high-resolution meteorological reanalysis data, preprocess it, and input it into the optimized global meteorological probability forecast model to obtain the meteorological forecast field for the next moment.

2. The weather forecasting method based on hierarchical graph neural network and latent variable injection according to claim 1, characterized in that, In step S21, the global meteorological state tensor is input into the hierarchical graph communication topology to obtain the coarsest-level node features. The specific process is as follows: The global meteorological state tensor is input into the latitude and longitude grid for mapping to obtain grid node features. The grid node features are then transmitted to the finest level of the multi-level icosahedral grid using a bipartite graph message passing method to obtain fine-level meteorological information. The fine-level meteorological information is then extracted and integrated level by level using a hierarchical coding mechanism and then transmitted to the coarsest level of the icosahedral grid to obtain the coarsest level node features.

3. The weather forecasting method based on hierarchical graph neural network and latent variable injection according to claim 1, characterized in that, In step S23, the fused large-scale circulation features are obtained, and the specific process is as follows: The latent variables and the coarsest-level node features are adjusted to ensure consistency in feature dimension and spatial scale. The latent variables are then superimposed as random perturbation terms onto the coarsest-level node features to obtain fused large-scale circulation features. The calculation formula is as follows: ； in, Represents latent variables; This represents the feature of the coarsest level node; This indicates the integration of large-scale circulation characteristics.

4. The weather forecasting method based on hierarchical graph neural network and latent variable injection according to claim 1, characterized in that, In step S24, the fused large-scale circulation features are input into the multi-layer graph processor evolution module to obtain the finest-level features. The specific process is as follows: The large-scale circulation features are input into stacked graph neural network blocks. Each graph neural network block performs three steps in sequence: edge update, node aggregation, and state update. Through these three steps, the large-scale circulation features are passed from the coarsest level to the finest level to obtain the finest level features.

5. The weather forecasting method based on hierarchical graph neural network and latent variable injection according to claim 1, characterized in that, In step S25, the finest-level features are input into the decoding module to obtain the meteorological forecast field for the next moment. The specific process is as follows: The finest-level features are input into the decoder, and the decoding mechanism maps the finest-level features back to the latitude and longitude grid. The features are then processed through the input projection layer to finally obtain the weather forecast field for the next moment.

6. The weather forecasting method based on hierarchical graph neural network and latent variable injection according to claim 1, characterized in that, In step S26, the global meteorological state tensor is input into the global meteorological probability forecasting model, and the model is trained and optimized through dual-path forward propagation, loss calculation, and backpropagation to obtain an optimized global meteorological probability forecasting model. The specific process is as follows: The CRPS approximate loss function based on dual-path forward propagation is used to calculate the total model loss by maximizing the difference between two samples in the global meteorological state tensor to approximate continuous hierarchical probability scoring. Backpropagation and weight optimization are performed based on the total loss, and training is continuously iterated until the total model loss value tends to stabilize and the forecast accuracy and ensemble dispersion reach the preset standards. Training is then stopped to obtain the optimized global meteorological probability forecast model. The CRPS approximate loss function based on dual-path forward propagation includes ensemble average reconstruction loss, divergence regularization term, and diversity loss term. The formula for calculating the average reconstruction loss of the ensemble is as follows: ； in, This represents the average reconstruction loss of the set; Indicates mean square error; This represents the first predicted sample generated by the two-path forward propagation; This represents the second predicted sample generated by the two-path forward propagation; This represents the true value of global high-resolution meteorological reanalysis data; The formula for calculating the divergence regularization term is as follows: ； in, express Divergence loss; Represents the squared term of the mean; Represents the variance term; The formula for calculating the diversity loss term is as follows: ； in, Represents the diversity loss term; This represents the diversity loss weighting coefficient; The formula for the total loss of the model is as follows: ； in, This represents the total loss of the model; This represents the weight coefficient of the KL divergence regularization term.

7. A computer-readable storage medium, characterized in that, Used to store computer-readable programs or instructions, which, when executed by a processor, can implement the steps in the weather forecasting method based on hierarchical graph neural networks and latent variable injection as described in any one of claims 1 to 6.

Images