Method for global multi-output residual ensemble weather prediction

Abstract

This invention discloses a multi-output residual ensemble weather forecasting method for global applications, primarily addressing the problem that the accuracy of mainstream AI-based weather forecasting models declines significantly with increasing forecast time steps. The method includes the following steps: S1, establishing a multi-output residual ensemble weather forecasting model, which includes a naive method module, a visual converter, a linear decoder, m residual modules, an ensemble module, and an adder; S2, inputting weather images into the naive method module, the visual converter, and the residual modules; S3, integrating and constraining the residual blocks in the ensemble module, outputting the integrated residuals; S4, outputting the final weather forecast result. This invention better suits the characteristics of weather data and can better handle heterogeneous integrated data, thereby improving the accuracy of short- and medium-term weather forecasts.

Description

Technical Field

[0001] This invention belongs to the field of weather forecasting technology, specifically, it relates to a multi-output residual integrated weather forecasting method for global applications. Background Technology

[0002] There are two main types of existing weather forecasting models. One is the traditional numerical weather prediction model based on physical laws. This model models the atmospheric and oceanic states at the current moment and uses physical laws to predict future weather conditions. However, this model has drawbacks such as huge computational costs, the need to solve very difficult nonlinear physical processes, slow speed in generating weather forecast results, and significant inherent biases. Therefore, this model is difficult to meet the needs of weather forecasting under complex climatic conditions.

[0003] Another type is the AI-driven data-driven weather forecasting model. This model uses deep learning to train on observational and reanalysis data, thus avoiding the need to solve difficult partial differential equations and reducing errors. Furthermore, data-driven weather forecasting models require lower computational costs and can generate forecasts in a shorter time. Currently, mainstream AI-driven data-driven weather forecasting models require large amounts of reanalysis data as training sets, resulting in low accuracy. Moreover, the accuracy decreases significantly as the forecast time step increases. Summary of the Invention

[0004] The purpose of this invention is to provide a multi-output residual integrated weather forecasting method for global applications, which mainly addresses the problem that the prediction accuracy of mainstream AI-based weather forecasting models decreases significantly as the number of prediction time steps increases.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A multi-output residual integrated weather forecasting method for global application includes the following steps:

[0007] S1. Establish a multi-output residual integrated weather prediction model, which includes a naive method module, a visual converter, a linear decoder, m residual modules, an integration module, and an adder.

[0008] S2 inputs the weather image into the Naive Method module, the Visual Converter, and the Residual Module; where:

[0009] The weather images processed by the Naive Method module are saved as the base output for weather forecasting results.

[0010] After being trained by a visual converter, the weather image is output as a master residual block by a linear decoder.

[0011] The weather image is processed through m residual modules to output m residual blocks;

[0012] S3, take m residual blocks as input to the integration module, integrate and constrain the residual blocks in the integration module, and output the integrated residual;

[0013] S4, the residual output of the integration module and the basic output saved by the naive method mechanism of the naive method module are added together by an adder to obtain the final weather forecast result.

[0014] Furthermore, in this invention, the linear decoder consists of a multi-head self-attention mechanism and a multilayer perceptron; each multilayer perceptron contains two fully connected layers, which use ReLU as the activation function and perform layer normalization before each block; wherein, the composition expression of the linear decoder is:

[0015] ,

[0016] ;

[0017] In the formula, This is the output of MSA. This is the output of the MLP; LN is the normalization symbol.

[0018] Furthermore, in this invention, the residual module consists of three fully connected layers. ReLU is used as the activation function in the first two layers, and no activation function is used in the last fully connected layer; wherein, the definition is... If the input is the residual module, then:

[0019] ,

[0020] In the formula, Represents the output of each residual module. This represents the residual module; therefore, the process of outputting the residual block is as follows:

[0021] ,

[0022] ,

[0023] ,

[0024] ;

[0025] In the formula, The matrix representing the linear transformation. The input weather image is processed through m residual modules. Each residual module outputs a residual block, resulting in m residual blocks.

[0026] Furthermore, in step S2, the visual converter segments the input image into 2D patch blocks. The sequence, each patch block is [size missing]. Total number of segments Linear projection is used to flatten and compress the patch into a D-dimensional space; the expression for flattening and compressing the patch into a D-dimensional space is:

[0027]

[0028] Here, the input weather image is defined as a high-latitude tensor, i.e. C represents the quantity of atmospheric variables, and H and W represent height and width, respectively; Represents a linear projection vector; It is position embedding. It is the input of the linear encoder.

[0029] Further, in step S2, the expression for integrating and constraining the residual blocks in the integration module is:

[0030] ,

[0031] In the formula, This represents the residual output of the vision converter. Indicates the first Each residual module, Indicates the first The input of each residual module, Indicates the first The output of each residual module The output after integration is represented by the image. After integrating all the residuals, a trainable tensor of the same size as the input weather image is trained using a visual transformer to constrain the integrated residuals.

[0032] Furthermore, in step S4, the expression for the final weather forecast result is output as follows:

[0033] ,

[0034] In the formula, This represents the baseline value preserved by the naive method mechanism. The constraint process representing the residual. This represents the final prediction result.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] (1) The multi-output residual ensemble model provided by this invention fully utilizes the characteristics of the residual mechanism to simulate subtle changes in weather conditions, which is consistent with the characteristics of weather data itself. It can better handle heterogeneous ensemble data, thereby improving the accuracy of short- and medium-term weather forecasts. In addition, once trained, the model provided by this invention can quickly output weather forecast results in a short time without huge computational costs and avoids the difficulty of solving nonlinear physical processes, thus better handling complex and changeable weather and climate conditions.

[0037] (2) The multi-output residual ensemble model W-MRI for global weather forecasting provided by this invention is an improvement on the current mainstream data-driven weather forecasting model (NVIDIA's FourCastNet model), which improves the accuracy of weather forecasting. Attached Figure Description

[0038] Figure 1 This is a schematic diagram of the multi-output residual integrated weather prediction model structure of the present invention.

[0039] Figure 2 This is a comparison chart showing the correlation coefficients between the present invention and the NVIDIA FourCastNet model over six-hour intervals and eight days for the Z500, T850, U10, and V10 weather variables.

[0040] Figure 3 This is a comparison chart showing the root mean square error results of the present invention and the NVIDIA FourCastNet model over six-hour intervals and eight days for the Z500, T850, U10, and V10 weather variables.

[0041] Figure 4 This is a comparison chart showing the correlation coefficient and root mean square error of the total precipitation forecast anomalies of the present invention and the NVIDIA FourCastNet model over a six-hour interval and five days. Detailed Implementation

[0042] The present invention will be further described below with reference to the accompanying drawings and embodiments. The embodiments of the present invention include, but are not limited to, the following embodiments.

[0043] This invention discloses a multi-output residual integrated weather forecasting method for global applications, comprising the following steps:

[0044] First, such as Figure 1As shown, a multi-output residual integrated weather prediction model is established. The model includes a naive method module, a visual converter, a linear decoder, m residual modules, an integration module, and an adder.

[0045] The weather image input to the model will be used as the baseline value for the weather forecast output using the naive method mechanism. The naive method mechanism can save the original weather image input, representing the weather conditions at the initial moment, and add it to the residual representing the weather changes to obtain the predicted weather conditions at the next moment.

[0046] Next, the weather image input to the model will also serve as input to the Vision Transformer (ViT) and m residual modules. The input weather image is defined as a high-latitude tensor, i.e. Where C represents the quantity of atmospheric variables, and H and W represent height and width, respectively. ViT segments the input image into 2D patch blocks. The sequence, each patch block is [size missing]. Total number of segments Then, linear projection is used to flatten and compress the patch into D-dimensional space, the expression of which is:

[0047] ,

[0048] Represents a linear projection vector; It is position embedding. This is the input to the linear encoder. The input image, after position embedding, is output to the linear encoder. The linear encoder consists of a multi-head self-attention mechanism (MSA) and a multilayer perceptron (MLP), and its expression is:

[0049] ,

[0050] ;

[0051] In the formula, This is the output of MSA. This is the output of the MLP; LN stands for normalization symbol. Each MLP contains two fully connected layers, each using ReLU as the activation function, and layer normalization (LN) is performed before each block. The entire process loops through L layers, finally passing through a linear decoder to obtain the output.

[0052] In this embodiment, each residual module consists of three fully connected layers. ReLU is used as the activation function in the first two layers, while the last fully connected layer does not use any activation function. Here, the definition is... If the input is the residual module, then:

[0053] ,

[0054] In the formula, Represents the output of each residual module. This represents the residual module; therefore, the process of outputting the residual block is as follows:

[0055] ,

[0056] ,

[0057] ,

[0058] ;

[0059] In the formula, The matrix representing the linear transformation. The variable represents bias, FC represents a fully connected layer, ReLU represents the activation function, and L represents a linear layer. The output of the previous fully connected layer serves as the input to the next fully connected layer. The input weather image passes through m residual modules, and each residual module outputs a residual block, ultimately resulting in m residual blocks.

[0060] Then, the vision converter uses the main residual block output from the linear decoder and the m residual blocks output from the m residual modules as input to the integration module. The integration module integrates and constrains these residual blocks, outputting the integrated residual. The integration operation in the integration module can be represented by the following equation:

[0061]

[0062] in This represents the residual of the ViT output. Indicates the first Each residual module, Indicates the first The input of each residual module, Indicates the first The output of each residual module This represents the output after integration. After integrating all the residuals, ViT is used to train a tensor of the same size as the input weather image to constrain the integrated residuals.

[0063] Finally, the constrained residuals output by the integration module are added to the baseline values ​​saved by the naive method mechanism to obtain the final weather forecast result from the model. The weather forecast result can be expressed by the following equation:

[0064]

[0065] in This represents the baseline value preserved by the naive method mechanism. The constraint process representing the residual. This represents the final prediction result.

[0066] In weather forecasting, the weather conditions at the next moment largely depend on the weather conditions at the previous moment. Residuals, by simulating the change in weather conditions between two moments, enable the weather forecasting model in this embodiment to capture the inherent relationships between weather data, thus allowing for better model training using weather data. The multi-output residual ensemble model W-MRI for global weather forecasting provided by this invention improves upon current mainstream data-driven weather forecasting models, enhancing forecast accuracy. This embodiment also uses the Anomaly Correlation Coefficient (ACC) and Root Mean Squared Error (RMSE) to evaluate the model, comparing it with NVIDIA's FourCastNet weather forecasting model when trained using the same dataset. Experimental results are as follows... Figure 2 As shown in Figures 3 and 4. In the figures, Z500, T850, U10, V10, and TP represent challenging weather forecasting variables.

[0067] Numerous experimental results demonstrate that the W-MRI model exhibits strong stability and superior performance in short- to medium-term weather forecasting.

[0068] The evaluation results show that this invention is more suited to the characteristics of weather data and can better handle heterogeneous integrated data, thereby improving the accuracy of short- and medium-term weather forecasts. Furthermore, once trained, the model provided by this invention can quickly output weather forecast results in a short time, without requiring huge computational costs and avoiding the difficulties of solving nonlinear physical processes, thus better handling complex and ever-changing weather and climate conditions.

[0069] The above embodiments are merely one of the preferred embodiments of the present invention and should not be used to limit the scope of protection of the present invention. Any modifications or refinements made to the main design concept and spirit of the present invention that are not of substantial significance, but solve the same technical problem as the present invention, should be included within the scope of protection of the present invention.

Claims

1. A multi-output residual integrated weather forecasting method for global application, characterized in that, Includes the following steps: S1. Establish a multi-output residual ensemble weather prediction model. The model includes a naive method module, a visual converter, a linear decoder, m residual modules, an ensemble module, and an adder. The linear decoder consists of a multi-head self-attention mechanism and a multilayer perceptron. Each multilayer perceptron contains two fully connected layers, which use ReLU as the activation function and perform layer normalization before each block. The composition expression of the linear decoder is as follows: ， ； In the formula, This is the output of MSA. This is the output of the MLP; LN is the normalization symbol. The residual module consists of three fully connected layers. ReLU is used as the activation function in the first two layers, while the last fully connected layer does not use any activation function. Wherein, the definition is... If the input is the residual module, then: ， In the formula, Represents the output of each residual module. This represents the residual module; the process of outputting the residual block is as follows: ， ， ， ； In the formula, The matrix representing the linear transformation. The input weather image is passed through m residual modules. Each residual module outputs a residual block, resulting in m residual blocks. S2 inputs the weather image into the Naive Method module, the Visual Converter, and the Residual Module; where: The weather image, processed by the naive method module, is used as the base output for weather forecasting; that is, the naive method module divides the input image into 2D patch blocks. The sequence, each patch block is [size missing]. Total number of segments Linear projection is used to flatten and compress the patch into a D-dimensional space; the expression for flattening and compressing the patch into a D-dimensional space is: Here, the input weather image is defined as a high-latitude tensor, i.e. C represents the quantity of atmospheric variables, and H and W represent height and width, respectively; , representing a linear projection vector; It is position embedding. It is the input of the linear encoder; After being trained by a visual converter, the weather image is output as a master residual block by a linear decoder. The weather image is processed through m residual modules to output m residual blocks; S3 takes m residual blocks as input to the integration module, integrates and constrains the residual blocks in the integration module, and outputs the integrated residual; wherein, the expression for integrating and constraining the residual blocks in the integration module is: ， In the formula, This represents the residual output of the vision converter. Indicates the first Each residual module, Indicates the first The input of each residual module, Indicates the first The output of each residual module The output after integration is represented by a visual transformer. After integrating all the residuals, a trainable tensor of the same size as the input weather image is trained to constrain the integrated residuals. S4, the residual output from the integration module and the basic output saved by the naive method mechanism of the naive method module are added together by an adder to obtain the final weather forecast result.

2. The multi-output residual integrated weather forecasting method for global coverage according to claim 1, characterized in that, In step S4, the expression for the final weather forecast result is output as follows: ， In the formula, This represents the baseline value preserved by the naive method mechanism. The constraint process representing the residual. This represents the final prediction result.

Images