A variational bayesian based ensemble weather forecast method

Abstract

The application discloses a set weather forecast method based on variational Bayes, comprising the following steps: 1, preprocessing global climate fifth generation atmospheric reanalysis dataset ERA5 and dividing windows; 2, establishing a three-dimensional neural network based on Swin-Transformer; 3, quantifying the probability distribution of model parameters by using a variational Bayes method; 4, model sampling is performed on the probability distribution, set forecast reasoning is realized, and the uncertainty of set forecast results is given. The application can improve the prediction skill of the current numerical method-based weather forecast model, and significantly improve the reasoning speed and efficiency, and reduce energy consumption.

Description

Technical Field

[0001] This invention relates to the fields of weather forecasting and artificial intelligence technology, specifically to an ensemble weather forecasting method based on variational Bayes. Background Technology

[0002] Weather forecasting is a crucial foundational technology for disaster prevention and mitigation, transportation, energy dispatch, agricultural production, and marine activities. Traditional weather forecasting primarily relies on numerical weather prediction methods, which predict future atmospheric conditions by numerically solving atmospheric motion equations and physical processes. To describe the inherent uncertainties in the evolution of weather systems, existing numerical weather prediction systems typically employ ensemble forecasting methods. This involves perturbing the initial field, boundary conditions, or model physical processes to generate multiple ensemble members, which are then used to assess the stability and uncertainty of the prediction results. However, traditional numerical ensemble forecasting usually requires costly numerical integration of multiple ensemble members separately, resulting in high computational resource consumption, low inference efficiency, and difficulty in balancing high timeliness with the demands of large-scale operational applications. Furthermore, as the number of ensemble members increases, computational and storage costs rise further, limiting its application in high-frequency, rapidly updating scenarios.

[0003] In recent years, with the development of deep learning technology, neural network-based weather forecasting methods have gradually become a research hotspot. These methods train deep neural network models using historical reanalysis or observational data to achieve rapid predictions of future weather conditions. Compared to traditional numerical weather prediction methods, deep learning methods have significant advantages in inference speed and computational efficiency, significantly shortening prediction time and reducing deployment costs. However, most existing deep learning-based weather forecasting methods primarily output single deterministic prediction results, making it difficult to directly provide uncertainty information like ensemble forecasting methods. For weather forecasting tasks, the atmospheric system exhibits significant nonlinearity, chaos, and multi-scale coupling characteristics; providing only a single prediction result often fails to meet the needs of risk warning and decision analysis for confidence range, dispersion, and extreme situation identification.

[0004] There is still a lack of a technical solution that balances prediction accuracy, inference efficiency, and probabilistic interpretability, as well as the ability to efficiently model spatiotemporal features using deep neural networks and the uncertainty quantification required for ensemble forecasting. Summary of the Invention

[0005] The present invention addresses the shortcomings of the existing technology by proposing a variational Bayes-based ensemble weather forecasting method. This method aims to quantify the uncertainty of ensemble forecasting inference and prediction results, thereby improving the forecasting skills of weather forecasting models, significantly increasing inference speed and efficiency, and reducing energy consumption.

[0006] To achieve the above-mentioned objectives, the present invention adopts the following technical solution: The present invention provides a variational Bayes-based ensemble weather forecasting method, characterized by the following steps: Step 1: Obtain the global climate dataset and segment it to obtain... N indivual T The global climate time series of frames is normalized to obtain the normalized global climate time series set, denoted as . ,in, Indicates the first n A normalized global gas time series, and ; express The Middle t Frame atmospheric state; Will The first to t The atmospheric state of a frame is denoted as and as the first n Atmospheric state sequence under a historical window; Will The first in t+1 arrive T The atmospheric state of a frame is denoted as and as the first n A sequence of actual atmospheric conditions under a prediction window; Step 2: Establish a 3D neural network model based on Swin-Transformer, including: a dimensionality reduction embedding module, a downsampling module, a FuserLayer module, an upsampling module, and a dimensionality enhancement and restoration module, and then... Processing is performed to obtain the first... n Atmospheric forecast state sequence under one forecast window; Step 3: Quantize the parameters of the three-dimensional neural network model using the variational Bayesian method. Based on the probability distribution and combined with the atmospheric prediction state sequence, an evidence lower bound loss function is designed, thereby utilizing the probability backpropagation method to... The variational posterior distribution is used for training to obtain the posterior distribution of the trained model parameters; where, This represents the total number of parameters in the 3D neural network model. Step 4: Sample the posterior distribution of the trained model parameters multiple times to obtain three-dimensional neural network models corresponding to multiple sets of model parameters, which are used to realize ensemble forecast inference and calculate the uncertainty of the ensemble forecast results.

[0007] The ensemble weather forecasting method based on variational Bayes as described in this invention is characterized in that step 2 includes the following steps: Step 2.1, the dimensionality reduction embedding module... After performing block partitioning, rearrangement, and feature mapping, we obtain the first... n Initial embedded feature sequence under a historical window ; Step 2.2, the downsampling module... Perform hierarchical encoding and spatial downsampling to obtain the first... n A historical window Coding features at each scale ,in, Indicates the first n The first historical window Encoding features at each scale, Indicates the number of levels in the coding layer; Step 2.3, the FuserLayer module is composed of The layers are composed of Swin-Transformer basic blocks based on the window attention mechanism, and... Normalization, window partitioning, and local window self-attention processing are performed to obtain the first... n A historical window Fusion features at various scales ,in, Indicates the first n The first historical window Fusion features at various scales; Step 2.4, the upsampling module... Perform layer-by-layer upsampling and compare with the encoded features at the corresponding scale. By merging, we obtain the first... n Feature representation after resolution restoration under each historical window ,in, Indicates the first Feature representation after layer resolution restoration; Step 2.5, the dimensionality restoration module... Perform rearrangement and mapping, and output the first... n Atmospheric forecast state sequence under one forecast window ,in, express The Middle t+ 1 frame of atmospheric prediction status.

[0008] Furthermore, step 2.3 includes the following steps: Step 2.3.1, the Layer Swin-Transformer base block pair After rearranging, we get the first... n The first historical window Three-dimensional feature representation ,in, , , They represent The vertical dimension, the latitudinal dimension, and the longitudinal dimension. express Channel dimension; Step 2.3.2, according to window size Will Along the vertical dimension, the latitudinal dimension and the longitudinal dimension are divided into The first local window, and the second The features of each local window are denoted as Thus, by using equation (1), the first... n The first historical window The first scale Query matrix , No. Key matrix and the Value matrix ,in, This represents the total number of local windows, and : (1) In equation (1), , , They represent the first The query mapping matrix, key-value mapping matrix, and value mapping matrix of the base block of the Swin-Transformer layer; Step 2.3.3: Use equation (2) to obtain the first... n The first historical window The first scale Window attention features : (2) In equation (2), Indicates the first The feature dimensions of each attention head in the layer. Indicates the first The first scale Relative position information within a local window; T indicates transpose; Indicates the activation function; Step 2.3.4, the first n The first historical window The attention features of each local window at each scale are restored and recombined according to their original spatial positions to obtain the first... n The first historical window Intermediate features after interaction of scale features Thus, by using equation (3), the first... Layer fusion features : (3) In equation (3), Indicates residual connection, This indicates a feedforward mapping.

[0009] Furthermore, step 2.4 includes the following steps: Step 2.4.1 Initialization , order the Feature representation after layer resolution restoration ; Step 2.4.2, the first Feature representation after layer resolution restoration Upsampling to The corresponding resolution yields the first Layer sampling features ; Step 2.4.3, will and Concatenate along the feature dimension, then map the concatenated features to obtain the first feature. Layer recovery Layer-coded feature representation ; Step 2.4.4, will Assign to Then, return to step 2.4.2 until... until.

[0010] Furthermore, step 3 includes the following steps: Step 3.1: Set the j-th parameter of the 3D neural network model. The prior distribution is The variational posterior distribution is Therefore, the lower bound loss function of evidence is calculated according to equation (1). ;in, It follows a normal distribution; Represents the j-th parameter The prior mean, Represents the j-th parameter The prior variance, Represents the j-th parameter The posterior mean, Represents the j-th parameter The posterior variance, : (4) In equation (1), For hyperparameters; Step 3.2: According to the backpropagation rule, calculate the loss function using equation (5). right and the j-th parameter Standard deviation gradient: (5) In equation (5), Indicates the losses from reconstruction, and ; Step 3.2: Update using stochastic gradient descent. and and in Once the model stabilizes, the variational posterior distribution is trained, resulting in the posterior distribution of the trained model parameters.

[0011] Furthermore, step 4 includes the following steps: Step 4.1: Perform analysis on the posterior distribution of the trained model parameters. Second sampling, obtained Group model parameters , Indicates the first Group model parameters, Indicates the number of groups; Step 4.2, Enter them separately Inference is performed in the 3D neural network model corresponding to the group model parameters to obtain the corresponding results. ensemble forecast results; Step 4.3: Use equation (6) to perform statistical analysis on the ensemble forecast results to obtain the ensemble average forecast result. and forecast dispersion To indicate uncertainty; (6) In equation (6), Indicates the first The ensemble forecast results, and ; Indicates the first The three-dimensional neural network model pairs corresponding to the group model parameters The output of the first t+ 1 frame of atmospheric prediction status, .

[0012] The present invention provides an electronic device, including a memory and a processor, characterized in that the memory is used to store a program supporting the processor in performing the method described therein, and the processor is configured to execute the program stored in the memory.

[0013] The present invention discloses a computer-readable storage medium storing a computer program, characterized in that the computer program is executed by a processor to perform the steps of the method described thereon.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention uses a three-dimensional neural network model based on Swing-Transformer to model the atmospheric state, effectively extracting multi-level spatiotemporal features from meteorological data, and improving the ability to represent and predict complex atmospheric evolution processes.

[0015] 2. This invention uses variational Bayesian probabilistic modeling on the parameters of a three-dimensional neural network model, enabling the model to output not only the prediction results but also the uncertainty information corresponding to the prediction results, thereby enhancing the application value of meteorological forecast results for risk warning, extreme weather identification, and decision support.

[0016] 3. This invention learns the probability distribution of model parameters within the same model framework and generates ensemble members by sampling the model parameters multiple times during the inference phase. Unlike traditional ensemble forecasting or multi-model ensemble methods, it does not require training and deploying multiple independent models separately, thereby reducing the complexity of model training and deployment, and reducing storage overhead and inference resource consumption. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0018] In this embodiment, an ensemble weather forecasting method that combines three-dimensional spatiotemporal feature modeling and probabilistic modeling capabilities enables the model to maintain high inference efficiency while generating multiple ensemble members through learning and sampling the probability distribution of model parameters. This gives the model parametric probabilistic modeling capabilities, thereby quantifying the uncertainty of ensemble forecast inference and prediction results. Specifically, as... Figure 1 As shown, the method includes the following steps: Step 1: Obtain the global climate dataset and segment it to obtain... N indivual T The global climate time series of frames is normalized to obtain the normalized global climate time series set, denoted as . ,in, Indicates the first n A normalized global gas time series, and ; express The Middle t In this embodiment, the interval between two adjacent atmospheric states is 6 hours.

[0019] Will The first to t The atmospheric state of a frame is denoted as and as the first n In this embodiment, the atmospheric state sequence under a historical window is described. t =2; Will The first in t+ 1 to T The atmospheric state of a frame is denoted as and as the first n In this embodiment, the atmospheric real-state sequence under a prediction window is shown. T =3.

[0020] Step 2: Establish a 3D neural network model based on Swin-Transformer, including: a dimensionality reduction embedding module, a downsampling module, a FuserLayer module, an upsampling module, and a dimensionality enhancement and restoration module, and then... Processing is performed to obtain the first... n Atmospheric forecast state sequence under one forecast window; Step 2.1, Dimensionality Reduction Embedding Module After performing block partitioning, rearrangement, and feature mapping, we obtain the first... n Initial embedded feature sequence under a historical window ; Step 2.2, the downsampling module... Perform hierarchical encoding and spatial downsampling to obtain the first... n A historical window Coding features at each scale ,in, Indicates the first n The first historical window Encoding features at each scale, In this embodiment, the level of the coding layer is indicated. =2.

[0021] Step 2.3, the FuserLayer module is... The layers are composed of Swin-Transformer basic blocks based on the window attention mechanism, and... Normalization, window partitioning, and local window self-attention processing are performed to obtain the first... n A historical window Fusion features at various scales ,in, Indicates the first n The first historical window Fusion features at various scales.

[0022] Step 2.3.1, the Layer Swin-Transformer base block pair After rearranging, we get the first... n The first historical window Three-dimensional feature representation ,in, , , They represent The vertical dimension, the latitudinal dimension, and the longitudinal dimension. express The channel dimension.

[0023] Step 2.3.2, according to window size Will Divided along the vertical dimension, the latitudinal dimension, and the longitudinal dimension into A partial window, in which and the The features of each local window are denoted as Thus, by using equation (1), the first... n The first historical window The first scale Query matrix , No. Key matrix and the Value matrix : (1) In equation (1), , , They represent the first The query mapping matrix, key-value mapping matrix, and value mapping matrix of the Swin-Transformer base block; in this embodiment, the window size .

[0024] Step 2.3.2: Use equation (2) to obtain the first... n The first historical window The first scale Window attention features : (2) In equation (2), Indicates the first The feature dimensions of each attention head in the layer. Indicates the first The first scale Relative position information within a local window; T indicates transpose; This represents the activation function.

[0025] Step 2.3.3, the first n The first historical window The attention features of each local window at each scale are restored and recombined according to their original spatial positions to obtain the first... n The first historical window Intermediate features after interaction of scale features Thus, by using equation (3), the first... Layer fusion features : (3) In equation (3), Indicates residual connection, This indicates a feedforward mapping.

[0026] Step 2.4, Upsampling module for Perform layer-by-layer upsampling and compare with the encoded features at the corresponding scale. By merging, we obtain the first... n Feature representation after resolution restoration under each historical window ,in, Indicates the first Feature representation after layer resolution restoration.

[0027] Step 2.4.1 Initialization , order the Feature representation after layer resolution restoration ; Step 2.4.2, the first Feature representation after layer resolution restoration Upsampling to The corresponding resolution yields the first Layer sampling features .

[0028] Step 2.4.3, will and Concatenate along the feature dimension, then map the concatenated features to obtain the first feature. Layer recovery Layer-coded feature representation ; Step 2.4.4, will Assign to Then, return to step 2.4.2 until... until.

[0029] Step 2.5, Dimensional Upgrade Recovery Module Perform rearrangement and mapping, and output the first...n Atmospheric forecast state sequence under one forecast window ,in, express The Middle t+ 1 frame of atmospheric prediction status; Step 3: Quantize the parameters of the 3D neural network model using the variational Bayesian method. Based on the probability distribution and combined with the atmospheric prediction state sequence, an evidence lower bound loss function is designed, thereby utilizing the probability backpropagation method to... The variational posterior distribution is used for training to obtain the posterior distribution of the trained model parameters; where, This represents the total number of parameters in the 3D neural network model.

[0030] Step 3.1: Set the j-th parameter of the 3D neural network model. The prior distribution is The variational posterior distribution is Therefore, the lower bound loss function of evidence is calculated according to equation (1). ;in, It follows a normal distribution; Represents the j-th parameter The prior mean, Represents the j-th parameter The prior variance, in this embodiment, ; Represents the j-th parameter The posterior mean, Represents the j-th parameter The posterior variance, : (4) In equation (1), As a hyperparameter, in this embodiment, =1e-4.

[0031] Step 3.2: According to the backpropagation rule, calculate the loss function using equation (5). right and the j-th parameter Standard deviation gradient: (5) In equation (5), Indicates the losses from reconstruction, and .

[0032] Step 3.2: Update using stochastic gradient descent. and and in Once the model stabilizes, the variational posterior distribution is trained, resulting in the posterior distribution of the trained model parameters.

[0033] Step 4: Sample the posterior distribution of the trained model parameters multiple times to obtain three-dimensional neural network models corresponding to multiple sets of model parameters, which are used to realize ensemble forecast inference and calculate the uncertainty of the ensemble forecast results; Step 4.1: Perform analysis on the posterior distribution of the trained model parameters. Second sampling, obtained Group model parameters , Indicates the first Group model parameters, Indicates the number of groups, in this example =48, which is sufficient to capture the uncertainty of atmospheric conditions.

[0034] Step 4.2, Enter them separately Inference is performed in the 3D neural network model corresponding to the group model parameters to obtain the corresponding results. The ensemble forecast results.

[0035] Step 4.3: Use equation (6) to perform statistical analysis on the ensemble forecast results to obtain the ensemble average forecast result. and forecast dispersion To indicate uncertainty; (6) In equation (6), Indicates the first The ensemble forecast results, and ; Indicates the first The three-dimensional neural network model pairs corresponding to the group model parameters The output of the first t+ 1 frame of atmospheric prediction status, .

[0036] In this embodiment, an electronic device includes a memory and a processor. The memory stores a program that supports the processor in executing the above-described method, and the processor is configured to execute the program stored in the memory.

[0038] In this embodiment, a computer-readable storage medium stores a computer program, which is executed by a processor to perform the steps of the above method.

Claims

1. A variational Bayes-based ensemble weather forecasting method, characterized in that, The steps include the following: Step 1: Obtain the global climate dataset and segment it to obtain... N indivual T The global climate time series of frames is normalized to obtain the normalized global climate time series set, denoted as . ,in, Indicates the first n A normalized global gas time series, and ; express The Middle t Frame atmospheric state; Will The first to t The atmospheric state of a frame is denoted as and as the first n Atmospheric state sequence under a historical window; Will The first in t+1 arrive T The atmospheric state of a frame is denoted as and as the first n A sequence of actual atmospheric conditions under a prediction window; Step 2: Establish a 3D neural network model based on Swin-Transformer, including: a dimensionality reduction embedding module, a downsampling module, a FuserLayer module, an upsampling module, and a dimensionality upsampling recovery module, and then... Processing is performed to obtain the first... n Atmospheric state sequence under one forecast window; Step 3: Quantize the parameters of the three-dimensional neural network model using the variational Bayesian method. Based on the probability distribution and combined with the atmospheric prediction state sequence, an evidence lower bound loss function is designed, thereby utilizing the probability backpropagation method to... The variational posterior distribution is used for training to obtain the posterior distribution of the trained model parameters; where, This represents the total number of parameters in the 3D neural network model. Step 4: Sample the posterior distribution of the trained model parameters multiple times to obtain three-dimensional neural network models corresponding to multiple sets of model parameters, which are used to realize ensemble forecast inference and calculate the uncertainty of the ensemble forecast results.

2. The ensemble weather forecasting method based on variational Bayes as described in claim 1, characterized in that, Step 2 includes the following steps: Step 2.1, the dimensionality reduction embedding module... After performing block partitioning, rearrangement, and feature mapping, we obtain the first... n Initial embedded feature sequence under a historical window ; Step 2.2, the downsampling module... Perform hierarchical encoding and spatial downsampling to obtain the first... n A historical window Coding features at each scale ,in, Indicates the first n The first historical window Encoding features at each scale, Indicates the number of levels in the coding layer; Step 2.3, the FuserLayer module is composed of The layers are composed of Swin-Transformer basic blocks based on the window attention mechanism, and... Normalization, window partitioning, and local window self-attention processing are performed to obtain the first... n A historical window Fusion features at various scales ,in, Indicates the first n The first historical window Fusion features at various scales; Step 2.4, the upsampling module... Perform layer-by-layer upsampling and compare with the encoded features at the corresponding scale. By merging, we obtain the first... n Feature representation after resolution restoration under each historical window ,in, Indicates the first Feature representation after layer resolution restoration; Step 2.5, the dimensionality restoration module... Perform rearrangement and mapping, and output the first... n Atmospheric forecast state sequence under one forecast window ,in, express The Middle t+ 1 frame of atmospheric prediction status.

3. The ensemble weather forecasting method based on variational Bayes as described in claim 2, characterized in that, Step 2.3 includes the following steps: Step 2.3.1, the Layer Swin-Transformer base block pair After rearranging, we get the first... n The first historical window Three-dimensional feature representation ,in, , , They represent The vertical dimension, the latitudinal dimension, and the longitudinal dimension. express Channel dimension; Step 2.3.2, by window size Will Along the vertical dimension, the latitudinal dimension and the longitudinal dimension are divided into The first local window, and the second The features of each local window are denoted as Thus, by using equation (1), the first... n The first historical window The first scale Query matrix , No. Key matrix and the Value matrix ,in, This represents the total number of local windows, and : (1) In equation (1), , , They represent the first The query mapping matrix, key-value mapping matrix, and value mapping matrix of the Swin-Transformer base block; Step 2.3.3: Using equation (2), obtain the first... n The first historical window The first scale Window attention features : (2) In equation (2), Indicates the first The feature dimensions of each attention head in the layer. Indicates the first The first scale Relative position information within a local window; T indicates transpose; Indicates the activation function; Step 2.3.4, the first n The first historical window The attention features of each local window at each scale are restored and recombined according to their original spatial positions to obtain the first... n The first historical window Intermediate features after interaction of scale features Thus, by using equation (3), the first... Layer fusion features : (3) In equation (3), Indicates residual connection, This indicates a feedforward mapping.

4. The ensemble weather forecasting method based on variational Bayes as described in claim 3, characterized in that, Step 2.4 includes the following steps: Step 2.4.1 Initialization , order the Feature representation after layer resolution restoration ; Step 2.4.2, the first Feature representation after layer resolution restoration Upsampling to The corresponding resolution yields the first Layer sampling features ; Step 2.4.3, will and Concatenate along the feature dimension, then map the concatenated features to obtain the first feature. Layer recovery Layer-coded feature representation ; Step 2.4.4, will Assign to Then, return to step 2.4.2 until... until.

5. The ensemble weather forecasting method based on variational Bayes as described in claim 4, characterized in that, Step 3 includes the following steps: Step 3.1: Set the j-th parameter of the 3D neural network model. The prior distribution is The variational posterior distribution is Therefore, the lower bound loss function of evidence is calculated according to equation (1). ;in, It follows a normal distribution; Represents the j-th parameter The prior mean, Represents the j-th parameter The prior variance, Represents the j-th parameter The posterior mean, Represents the j-th parameter The posterior variance, : (4) In equation (1), For hyperparameters; Step 3.2: According to the backpropagation rule, calculate the loss function using equation (5). right and the j-th parameter Standard deviation gradient: (5) In equation (5), Indicates the losses from reconstruction, and ; Step 3.2: Update using stochastic gradient descent. and and in Once the model stabilizes, the variational posterior distribution is trained, resulting in the posterior distribution of the trained model parameters.

6. The ensemble weather forecasting method based on variational Bayes as described in claim 5, characterized in that, Step 4 includes the following steps: Step 4.1: Perform analysis on the posterior distribution of the trained model parameters. The second sampling yielded Group model parameters , Indicates the first Group model parameters, Indicates the number of groups; Step 4.2, Enter them separately Inference is performed in the 3D neural network model corresponding to the group model parameters to obtain the corresponding results. ensemble forecast results; Step 4.3: Use equation (6) to perform statistical analysis on the ensemble forecast results to obtain the ensemble average forecast result. and forecast dispersion To indicate uncertainty; (6) In equation (6), Indicates the first The ensemble forecast results, and ; Indicates the first The three-dimensional neural network model pairs corresponding to the group model parameters The output of the first t+ 1 frame of atmospheric prediction status, .

7. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store a program that supports a processor in executing the method of any one of claims 1-6, the processor being configured to execute the program stored in the memory.

8. A computer-readable storage medium storing a computer program thereon, characterized in that, The computer program is executed by a processor to perform the steps of the method according to any one of claims 1-6.

Images