Regional weather forecasting method based on terrain field and meteorological field for symmetric enhancement

Abstract

This invention provides a regional weather forecasting method based on symmetric enhancement of topographic and meteorological fields, belonging to the field of data processing technology. The method includes: inputting the current multivariate meteorological field into an autoencoder to obtain historical latent variables; inputting a static topographic field into a topographic encoder to obtain topographic features; encoding the historical latent variable sequence into a context feature sequence, and injecting the topographic features into the context feature sequence through a gating method to obtain topographically enhanced context features; inputting the topographically enhanced context features and noise scale into a conditional diffusion denoising network to obtain future prediction latent variables, decoding to obtain the predicted meteorological field; during network training, applying the same symmetric enhancement group effect to the historical latent variable sequence, noisy target latent variables, and topographic features; and continuously predicting to obtain the predicted meteorological field for multiple consecutive time periods. This method enables high-precision regional weather forecasts with long time horizons and for multiple consecutive time periods.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and more specifically, to a regional weather forecasting method based on symmetrical enhancement of topographic and meteorological fields. Background Technology

[0002] Current regional weather forecasting technology mainly relies on numerical weather prediction models to build its technical system. Among them, the WRF mesoscale weather model is the core representative of regional numerical models. By solving the primitive atmospheric equations, parameterizing turbulent and radiation physical processes, and combining regional boundary conditions and topographic forcing, it can achieve numerical integral extrapolation of the weather evolution process in a specific region. This type of traditional numerical forecasting method has a clear physical mechanism and standardized process characterization, and is the mainstream basic technology for regional weather forecasting. However, it has significant technical shortcomings, including high computational resource consumption, cumbersome model parameter debugging, high sensitivity to boundary condition adaptation, and high cost of ensemble forecast implementation, resulting in significant limitations in its application.

[0003] High-precision, long-term forecasting under complex local meteorological conditions is a core challenge in regional meteorological forecasting and a key bottleneck restricting accurate early warning and long-term meteorological prediction of severe weather. Regional meteorological evolution is influenced by multiple coupling factors, including complex topography, local ground surface characteristics, land-sea boundaries, and strong nonlinear convection processes. This places stringent demands on the spatial resolution and the ability to finely represent local meteorological structures in forecast models. The shortcomings of traditional numerical models in terms of real-time performance and scalability are further highlighted. Although the short-term average forecast error is relatively controllable, their ability to capture the structural characteristics of key severe weather events such as severe convection is insufficient, leading to significant forecast biases. Furthermore, they cannot balance computational efficiency and forecast accuracy, making it difficult to meet the high-precision regional meteorological forecasting needs of complex terrain areas and long-term autoregressive scenarios. Summary of the Invention

[0004] The problem this invention aims to solve is that existing methods are insufficient to meet the demand for high-precision regional weather forecasts in complex terrain areas and long-term autoregressive scenarios.

[0005] To address the aforementioned problems, in a first aspect, the present invention provides a regional weather forecasting method based on symmetrical enhancement of topographic and meteorological fields, comprising: The current multivariate weather field is input into a pre-trained autoencoder to obtain historical latent variables; The static terrain field is input into the terrain encoder to obtain the terrain features; The historical latent variable sequence is encoded into a context feature sequence, and the terrain features are injected into the context feature sequence through a gating method to obtain the terrain-enhanced context features. The enhanced contextual features and noise scale are input into a conditional diffusion denoising network pre-trained with symmetric enhancement to obtain future prediction latent variables. The future prediction latent variables are then input into a decoder to obtain the predicted weather field. During the training phase of the conditional diffusion denoising network, the same symmetric enhancement group action is applied to the historical latent variable sequence, the noisy target latent variable, and the terrain features. The corresponding inverse group action is applied to the network output before the denoising loss is calculated. By injecting future latent variables into the historical latent variable sequence for the next moment, and continuously predicting future latent variables, the predicted meteorological field for multiple consecutive moments is obtained.

[0006] Optionally, the terrain encoder includes a linear lifting layer, a physical resolution spectral convolutional block, a spectral domain downsampling layer, a latent spatial resolution spectral convolutional block, and a projection head; in, This represents the initial hidden features after linear enhancement. Indicates a linear lifting layer. Represents a static terrain field. Indicates the first The output of a physical resolution spectral convolution block Indicates the first -1 physical resolution spectral convolution block output Indicates the first A physical resolution spectral convolution block This indicates the number of spectral convolutional blocks stacked at the physical resolution stage. This represents the hidden features that fall within the latent spatial resolution after spectral domain downsampling. Indicates the first -1 physical resolution spectral convolution block output This indicates a downsampling layer in the spectral domain. This represents the output of the r-th latent spatial resolution spectral convolution block. This represents the output of the (r-1)th latent spatial resolution spectral convolution block. This represents the input to the first latent spatial resolution spectral convolutional block. , This represents the r-th latent spatial resolution spectral convolutional block. This represents the number of spectral convolutional blocks stacked in the latent spatial resolution stage. This represents the terrain features output by the projector. Indicates the first -1 latent spatial resolution spectral convolution block output This refers to the projection head.

[0007] Optionally, the output of the spectral convolution block is in, This represents the output of the spectral convolution block. This represents the input to the spectral convolution block. ( ) represents the discrete spherical Fourier transform. ( ) represents the discrete spherical inverse Fourier transform. This represents a learnable frequency domain filter. This represents a channel-by-channel, mode-by-mode product. This represents a multilayer perceptron.

[0008] Optionally, the enhanced terrain context features are: in, This represents the contextual features after terrain enhancement. Represents the context feature sequence, This represents the sigmoid activation function. Indicates a learnable scalar. Represents the lexical embedding map, Indicates terrain features.

[0009] Optionally, the conditional diffusion denoising network can be pre-trained, and pre-training methods include: Noise is added to the target latent variable to obtain the noisy target latent variable; During training, a translation is randomly generated according to a uniform distribution, and the symmetric augmented random cluster element corresponding to the translation is selected from a finite cyclic group. Symmetric augmented random cluster elements are applied to the noisy target latent variables and the terrain-enhanced context features, and then input into the denoising network layer. A translation inverse group operation is applied to the output of the denoising network layer to obtain the future prediction latent variables. The loss of the conditional diffusion denoising network is obtained based on the future prediction latent variables and the target latent variables. The parameters of the conditional diffusion denoising network are adjusted based on the loss until the loss converges, resulting in the trained conditional diffusion denoising network.

[0010] Optionally, during the inference phase, the latent variables for future prediction are: During the training phase, the latent variables for future prediction are: in, Indicates latent variables for future prediction. This indicates the translation inverse group operation. ( ) represents a denoising network layer. Represents symmetric enhanced random cluster elements. Represents the latent variables of a noisy target. This represents the maximum value of the noise scale. Indicates the noise level. This represents the contextual features after terrain enhancement.

[0011] Secondly, the present invention also provides a regional weather forecasting system based on topographic and meteorological fields with symmetrical enhancement, comprising: The meteorological field encoding module is used to input the current multivariate meteorological field into a pre-trained autoencoder to obtain historical latent variables; The terrain field encoding module is used to input the static terrain field into the terrain encoder to obtain terrain features; The gated fusion module is used to encode the historical latent variable sequence into a context feature sequence, and inject the terrain features into the context feature sequence through a gated method to obtain the terrain-enhanced context features. The prediction module is used to input the enhanced terrain context features and noise scale into the conditional diffusion denoising network pre-trained by symmetric enhancement to obtain future prediction latent variables. In the training phase of the conditional diffusion denoising network, the same symmetric enhancement group action is applied to the historical latent variable sequence, the noisy target latent variable and the terrain features, and the corresponding inverse group action is applied to the network output before calculating the denoising loss. The meteorological decoding module is used to input the future prediction latent variables into the decoder to obtain the predicted meteorological field; The loop module is used to inject future latent variables into the historical latent variable sequence for the next time moment, continuously predicting future latent variables to obtain the predicted meteorological field for multiple consecutive time moments.

[0012] Thirdly, the present invention provides an electronic device, including a memory and a processor; The memory is used to store computer programs; The processor is configured to, when executing the computer program, implement the regional weather forecasting method based on topographic and meteorological fields with symmetric enhancement as described in the first aspect.

[0013] Fourthly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the regional weather forecasting method based on topographic and meteorological fields for symmetric enhancement as described in the first aspect.

[0014] This invention provides a regional weather forecasting method based on symmetrical enhancement of topographic and meteorological fields. Compared with existing technologies, it has the following advantages: By introducing a pre-trained autoencoder and a conditional diffusion denoising network, high-dimensional meteorological data is compressed into low-dimensional latent variables for processing, significantly reducing computational complexity and improving forecast efficiency. Static terrain features are extracted using a terrain encoder and injected into the context feature sequence using a gating method to generate terrain-enhanced context features. This mechanism can capture the modulating effect of terrain on local weather more precisely and dynamically, thereby improving forecast accuracy in complex terrain areas. The conditional diffusion denoising network, combined with symmetric enhancement and inverse group operations, enhances the model's robustness to spatial variations in meteorological models, improving prediction accuracy and stability. By continuously forecasting future latent variables by injecting them into the historical latent variable sequence for the next time step, high-precision regional weather forecasts over long timeframes and multiple consecutive time steps can be achieved, effectively solving the problem of traditional models struggling to guarantee accuracy in long-term autoregressive scenarios. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart illustrating a regional weather forecasting method based on topographic and meteorological fields for symmetrical enhancement, provided in an embodiment of the present invention. Figure 2 A data processing flowchart of a regional weather forecasting method based on topographic and meteorological fields for symmetrical enhancement, provided in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the symmetric enhancement of the discrete cyclic group effect along the longitude direction, provided in an embodiment of the present invention. Figure 4 This is a schematic diagram of the structure of a regional weather forecasting system based on topographic and meteorological fields for symmetrical enhancement, provided as an embodiment of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions in the embodiments of this application are described clearly and completely. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific implementation methods.

[0019] like Figure 1 As shown in the embodiment of this application, a regional weather forecasting method based on topographic and meteorological fields for symmetrical enhancement is provided, comprising: S1: Input the current multivariate weather field into the pre-trained autoencoder to obtain historical latent variables.

[0020] Specifically, a multivariate meteorological field typically refers to a dataset consisting of multiple meteorological elements (such as temperature, humidity, wind speed, air pressure, precipitation, etc.) at a specific region and time point. This data can be observed values, analyzed values, or model output values, used to describe comprehensive information about atmospheric conditions. An autoencoder is a neural network model that learns effective encoding (i.e., latent variables) and decoding of input data. By compressing and reconstructing the input data, an autoencoder can extract core features from the data, achieving a dimensionality-reduced representation. A pre-trained autoencoder can be an encoder and decoder composed of multiple fully connected layers. The encoder compresses high-dimensional meteorological field data into a low-dimensional latent variable representation, while the decoder reconstructs the latent variables back into the meteorological field. Acquiring historical latent variables aims to transform complex meteorological information into a more compact and easier-to-process representation.

[0021] S2: Input the static terrain field into the terrain encoder to obtain terrain features.

[0022] Specifically, static topographic fields refer to relatively fixed and unchanging topographic information in geographic space, such as elevation, slope, aspect, and land cover type. This information has a long-term and stable impact on regional meteorological processes. A topographic encoder is a neural network module specifically designed to process static topographic field data. It extracts physically meaningful and predictive topographic features from raw topographic data, such as spatial features at different scales, mountain range orientation, and valley distribution, to help the model understand the impact of topography on meteorology. A topographic condition encoder based on the GSNO operator can be used for topographic feature extraction. The GSNO operator can also be replaced with other multi-scale spectral methods (such as the standard Fourier neural operator FNO, the adaptive Fourier neural operator AFNO, and the spherical Fourier neural operator SFNO). S3: Encode the historical latent variable sequence into a context feature sequence, and inject the terrain features into the context feature sequence through a gating method to obtain the terrain-enhanced context features.

[0023] Specifically, the gating method can be a simple multiplicative gating unit, in which terrain features are multiplied element-wise with a sequence of contextual features (e.g., an encoding or embedding vector representing the current time step) to modulate the influence of terrain features on the overall context. The sequence of contextual features can contain time information (such as month, date, hour), seasonal information, or historical latent variables, providing dynamic background information for terrain features so that the influence of terrain can be adjusted according to the specific context.

[0024] S4: Input the enhanced terrain context features and noise scale into the conditional diffusion denoising network pre-trained with symmetric enhancement to obtain future prediction latent variables, and input the future prediction latent variables into the decoder to obtain the predicted meteorological field; wherein, during the training phase of the conditional diffusion denoising network, the same symmetric enhancement group action is applied to the historical latent variable sequence, the noisy target latent variable and the terrain features, and the corresponding inverse group action is applied to the network output before calculating the denoising loss.

[0025] Specifically, the conditional diffusion denoising network is a generative neural network based on the principle of diffusion models. It can employ a DiT architecture with two streams plus a single stream Transformer block, a pure Transformer structure, a Diffusion U-Net structure, or a sequence modeling backbone based on a state-space model (such as Mamba). The key requirement is that it can accept the contextual features after terrain enhancement, noisy target latent variables, and noise scale as input. Furthermore, the diffusion process can use an EDM-style Karras lognormal sigma distribution, or DDPM linear scheduling, cosine scheduling, or v-prediction parameterization. The sampler can use DDIM, DPM-Solver, or EDM multi-step solvers, all of which can achieve the invention's objectives. For example, the historical latent variable sequence can consist of meteorological latent variables from the past 24 hours or longer. The conditional diffusion denoising network can be a deep neural network based on a U-Net architecture, which learns how to recover the original data from noisy data during training, guided by conditional information in the process. Symmetric enhancement operations can include random rotation, flipping, or translation of the input data to improve the model’s robustness to spatial variations in weather patterns.

[0026] The decoder can have the same structure as the decoder in the autoencoder, and is responsible for reconstructing low-dimensional future prediction latent variables into high-dimensional, interpretable predictive weather field data.

[0027] S5: Inject future latent variables into the historical latent variable sequence for the next time moment to continuously predict future latent variables, obtaining the predicted meteorological field for multiple consecutive time moments. Specifically, the continuous prediction process can adopt an autoregressive approach, where the prediction result at each time moment is used as the input condition for the next time moment, thereby achieving long-term regional weather forecasting.

[0028] In this optional embodiment, by introducing a pre-trained autoencoder and a conditional diffusion denoising network, high-dimensional meteorological data is compressed into low-dimensional latent variables for processing, significantly reducing computational complexity and improving forecast efficiency. For example, when processing meteorological data for location A, the use of latent variables enables the model to learn and infer meteorological evolution patterns more efficiently.

[0029] Static terrain features are extracted by a terrain encoder and then gated into a context feature sequence to generate terrain-enhanced context features. This mechanism can capture the modulating effect of terrain on local weather more precisely and dynamically, thereby improving forecast accuracy in complex terrain areas. For example, when predicting localized heavy precipitation at location A, the terrain-enhanced context features can help the model better identify the impact of terrain on water vapor transport and convection triggering, compensating for the shortcomings of traditional models in terms of fine-grained representation.

[0030] The conditional diffusion denoising network, combined with symmetric augmentation and inverse group operation, enhances the model's robustness to spatial variations in meteorological models, improving prediction accuracy and stability. For example, when facing a variable meteorological system at location A, symmetric augmentation enables the model to learn meteorological models from different perspectives, thus exhibiting stronger generalization ability in actual predictions.

[0031] By using future predictable latent variables as inputs for continuous forecasting, this method enables high-precision regional weather forecasts over long periods and multiple consecutive timeframes, effectively addressing the issue of traditional models struggling to guarantee accuracy in long-term autoregressive scenarios. For example, when making 24-hour continuous forecasts for location A, the iterative forecasting mechanism of this method maintains high forecast consistency and accuracy, providing more reliable technical support for severe weather warnings.

[0032] The following is a detailed description of each step, such as Figure 2 As shown.

[0033] S1: Input the current multivariate weather field into the pre-trained autoencoder to obtain historical latent variables.

[0034] Specifically, the latent space coding part uses a pre-trained autoencoder to compress the high-dimensional multivariate meteorological field. Let the multivariate meteorological field of the study area be... It is composed of commonly used regional forecasting variables such as geopotential height, specific humidity, temperature, wind field components, near-surface temperature, sea-level pressure, and total precipitation. Encoder With decoder The mapping between the physical field and the latent space is completed respectively.

[0035] In a typical implementation, the dimension of the latent variable is . .

[0036] S2: Input the static terrain field into the terrain encoder to obtain terrain features.

[0037] Specifically, the static topographic field is This includes land and sea masks and multi-scale terrain height. A GSNO-based terrain encoder. Using a static terrain field as input, the long-range correlation structure is directly manipulated through frequency domain methods. Frequency domain constraints are applied at both the physical resolution and latent space resolution scales to achieve scale matching from physical space to latent space. The terrain encoder consists of five stages: a linear lifting layer, a physical resolution spectral convolutional block, a spectral downsampling layer, a latent space resolution spectral convolutional block, and a projection head. Formalized: (2) (3) (4) (5) (6) in, This represents the initial hidden features after linear enhancement. Indicates a linear lifting layer. Represents a static terrain field. Indicates the first The output of a physical resolution spectral convolution block Indicates the first -1 physical resolution spectral convolution block output Indicates the first A physical resolution spectral convolution block This indicates the number of spectral convolutional blocks stacked at the physical resolution stage. This represents the hidden features that fall within the latent spatial resolution after spectral domain downsampling. Indicates the first -1 physical resolution spectral convolution block output This indicates a downsampling layer in the spectral domain. This represents the output of the r-th latent spatial resolution spectral convolution block. This represents the output of the (r-1)th latent spatial resolution spectral convolution block. This represents the input to the first latent spatial resolution spectral convolutional block. , This represents the r-th latent spatial resolution spectral convolutional block. This represents the number of spectral convolutional blocks stacked in the latent spatial resolution stage. This represents the terrain features output by the projector. Indicates the first -1 latent spatial resolution spectral convolution block output This refers to the projection head.

[0038] Furthermore, a single spectral convolutional block consists of three branches, and the output of the spectral convolutional block is... (7) in, This represents the output of the spectral convolution block. This represents the input to the spectral convolution block. ( ) represents the discrete spherical Fourier transform. ( ) represents the discrete spherical inverse Fourier transform. This represents a learnable frequency domain filter. This represents a channel-by-channel, mode-by-mode product. This represents a multilayer perceptron. The residual pass-through preserves the identity path of the input signal, while the spectral convolution branch passes through... In the frequency domain, MLP branches linearly control different modal responses and provide a single-layer global receptive field, while in the spatial domain, they perform nonlinear feature recombination position by position.

[0039] The spectral convolutional block adopts a three-branch structure that includes residual pass-through, frequency domain learnable filtering, and pointwise nonlinear hybrid MLP. It achieves decoupling between linear spectral filtering (cross-space multi-scale selection) and nonlinear feature recombination (cross-channel hybridization), adapting to the characteristics of static terrain fields with strong multi-scale structures.

[0040] Spectral domain downsampling first performs low-pass truncation in the frequency domain and then resamples to a coarser grid. The spectral coefficients are then processed according to the truncation order. and Perform low-pass truncation.

[0041] (8) in, These are the spectral coefficients after low-pass truncation. The input features are in ( Spectral coefficients (spherical harmonic or two-dimensional Fourier coefficients) on the m) mode. Total wavenumber / latitudinal cutoff order. m is the wavenumber in the longitude direction. L' and M' are the cutoff thresholds (maximum retained wavenumbers), determined according to the latent space dimensions Hz and Wz, with L'=Hz / 2 and M'=Wz / 2.

[0042] The difference between the spectral domain downsampling operator in this application and the general spectral neural operator is that it serves static condition coding in regional multivariate latent space diffusion forecasting and needs to complete scale matching between physical resolution and latent space resolution. Spectral domain downsampling is introduced as a key operator connecting physical resolution and latent space resolution.

[0043] The terrain condition encoder based on the Generalized Spherical Neural Operator (GSNO) transforms the static terrain field into terrain features consistent with the scale of the latent space backbone through five stages: linear lifting layer, physical resolution spectral convolution, spectral domain low-pass downsampling, latent space spectral convolution thinning, and projection head. These features are then injected into the conditional backbone using a learnable gating method. Spectral domain downsampling, as a key operator bridging the physical resolution and latent space resolution, avoids aliasing of high-frequency components of the static terrain during scale transformation in the Nyquist sense.

[0044] S3: Encode the historical latent variable sequence into a context feature sequence, and inject the terrain features into the context feature sequence through a gating method to obtain the terrain-enhanced context features.

[0045] Historical latent variable series Context feature sequences are generated through lexical embedding. Topographic features are obtained through word embedding. Then, through learnable scalar gating parameters With context feature sequence After fusion, the contextual features after terrain enhancement are: (9) in, This represents the contextual features after terrain enhancement. This represents the context feature sequence, the current layer's context feature sequence, which is composed of historical latent variable temporal embeddings; This represents the sigmoid activation function. Indicates a learnable scalar. This represents a lexical embedding mapping, where terrain features are processed using PatchEmbed+LayerNorm. This design represents terrain features. The terrain injection intensity can be automatically learned during the training process, adaptively adjusting the influence of the terrain branch for different variables, regions, and forecast lead times.

[0046] S4: Input the enhanced terrain context features and noise scale into the conditional diffusion denoising network pre-trained with symmetric enhancement to obtain future prediction latent variables, and input the future prediction latent variables into the decoder to obtain the predicted meteorological field; wherein, during the training phase of the conditional diffusion denoising network, the same symmetric enhancement group action is applied to the historical latent variable sequence, the noisy target latent variable and the terrain features, and the corresponding inverse group action is applied to the network output before calculating the denoising loss.

[0047] In the inference phase, the latent variables for future prediction are: in, Indicates latent variables for future prediction. ( ) represents a denoising network layer. This represents the maximum value of the noise scale. This represents the contextual features after terrain enhancement.

[0048] Specifically, the conditional diffusion denoising network needs to be pre-trained, and pre-training methods include: Noise is added to the target latent variable to obtain the noisy target latent variable.

[0049] By introducing random perturbations to simulate the uncertainty or degradation of real-world data, the network's ability to recover original information from damaged inputs can be enhanced. For example, Gaussian noise, with a mean of zero and a variance that gradually decreases with the number of training steps or time steps, can be added to historical latent variables; or, Poisson noise can be applied to simulate random fluctuations in count data.

[0050] During training, a translation is randomly generated according to a uniform distribution, and the symmetric augmented random cluster element corresponding to the translation is selected from a finite cyclic group.

[0051] It should be noted that by randomly generating translation amounts and selecting corresponding cluster elements, various possible spatial displacements of data can be simulated during training. For example, the translation amount can be a randomly selected pixel offset (Δx, Δy) on the image plane, where Δx and Δy are uniformly distributed within a preset range (e.g., -5 to 5 pixels). The corresponding symmetric augmentation random cluster element is a function or matrix that performs this (Δx, Δy) translation operation.

[0052] Symmetrically enhanced random cluster elements are applied to the noisy target latent variables and the terrain-enhanced contextual features, and then input into the denoising network layer. A translation inverse group operation is applied to the output of the denoising network layer to obtain the future prediction latent variables.

[0053] It should be noted that the symmetry enhancement mechanism can be based on discrete cyclic groups in the longitude direction, or it can use finite symmetry groups in the latitude direction, or use finite groups in multiple directions as joint symmetry enhancement; in addition, the random sampling distribution of the group elements can also be replaced by a weighted distribution instead of a uniform distribution.

[0054] The loss of the conditional diffusion denoising network is obtained based on the future prediction latent variables and the target latent variables.

[0055] The step of obtaining the loss of the conditional diffusion denoising network based on the future predicted latent variables and the target latent variables is used to quantify the difference between the model's predictions and the true target, providing direction for the network's parameter updates. The design of the loss function directly affects the model's learning performance and convergence. For example, mean squared error (MSE) can be used as the loss function to calculate the squared difference between the future predicted latent variables and the corresponding true (unnoised) latent variables; or, mean absolute error (MAE) can be used to measure the prediction error.

[0056] The parameters of the conditional diffusion denoising network are adjusted based on the loss until the loss converges, resulting in the trained conditional diffusion denoising network.

[0057] For example, the training process stops when the loss value reaches a stable state or a preset condition is met. The Adam optimizer can be used, with an initial learning rate of 0.0001, and a learning rate scheduling strategy (such as cosine annealing) can be used to dynamically adjust the learning rate; alternatively, a stochastic gradient descent (SGD) optimizer can be used, with a momentum term set to accelerate convergence.

[0058] In the pre-training of the conditional diffusion denoising network, symmetric enhancement and inverse group operations were introduced, enabling the network to learn predictive capabilities with equivariance to common spatial translation transformations in meteorological fields. This training method significantly enhances the robustness and generalization ability of the conditional diffusion denoising network, allowing it to maintain high-precision predictive performance even when faced with small spatial displacements in meteorological data caused by observation errors, model discretization, or actual physical processes. This pre-training method effectively avoids the network's over-reliance on features specific to a particular spatial location, instead learning more fundamental and physically meaningful patterns, thereby improving the accuracy and stability of regional weather forecasts, especially when dealing with complex and variable meteorological phenomena, providing more reliable prediction results.

[0059] In a specific embodiment, the symmetric enhancement mechanism of the discrete cyclic group effect in the longitude direction, such as... Figure 3 As shown. Let the latent variable be... Define group for Finite cyclic group of order, group element ( The corresponding group action is a cyclic translation along the longitude direction (the last dimension). Grid points: During training, each batch (block) is sampled with a random shift amount in a uniform distribution. (K is the half-width of the symmetric range, and the complete finite cyclic group is) Whole-group sampling (As an alternative implementation), it can simultaneously apply to noisy target latent variables, historical latent variable sequences, and terrain-enhanced contextual features. During enhancement, terrain features are cyclically translated along the longitude direction with the same k; after forward propagation, an inverse group effect is applied to the model output. Then calculate the loss.

[0060] Future predictive latent variables are (11) in, Indicates latent variables for future prediction. This indicates the translation inverse group operation. ( ) represents a denoising network layer. Represents symmetric enhanced random cluster elements. This represents the sequence of historical latent variables at time t-1. Represents the latent variables of a noisy target. Indicates the noise level. The context features represent the terrain enhancement, and the network can learn parameters are θ.

[0061] The loss of the conditional diffusion denoising network is (12) in, Representing historical latent variables, Let w(σ) represent the latent variable for future prediction. We perform weighted L2 regression on the prediction error to balance the variance of the training signal across different noise scales.

[0062] This mechanism leverages the approximate statistical stationarity of regional meteorological data along the longitude direction, forcing the network to learn robust representations to longitude-direction positional perturbations through data augmentation on a finite swarm. Its advantages include: (a) lightweight implementation, introducing no additional learnable parameters, only adding lightweight tensor transformation overhead such as cyclic translations; (b) compact swarm structure, with an actual sampling space of 2K+1 translations (or expandable to a complete swarm). (c) It does not rely on the isovariance assumption of the encoder, but only utilizes the statistical uniformity of the longitude dimension of the latent space.

[0063] Karras lognormal sigma sampling and preconditioning sample regression in the EDM style: the forward process adds noise to the historical latent variables to obtain the noisy target latent variable. The inverse denoising network predicts the denoised historical latent variables based on historical latent variable sequences, terrain features, noisy target latent variables, and noise scale. If parameterization such as DDPM / DDIM is used, this objective can also be equivalently written in the standard noise regression form. After introducing discrete symmetric enhancement along the longitude direction, the unified training objective is as follows: (13) Among them, the expectation is for Joint sampling, σ follows a Karras lognormal noise scale distribution. Standard Gaussian noise, The discrete translation amount used in training (which can also be extended to full group sampling) The target is structurally fully compatible with standard conditional diffusion, and existing diffusion training techniques, time-step sampling strategies, and samplers can be directly inherited.

[0064] A longitude-direction discrete cyclic translation symmetric enhancement mechanism based on finite cyclic groups applies stochastic discrete group effects simultaneously to conditional latent variables, noisy historical latent variables, and terrain features during the training phase. After forward propagation, an inverse group effect is applied to the model output, and EDM preconditioning denoising loss or equivalent noise regression loss is calculated. This mechanism does not introduce any additional learnable parameters and does not rely on the encoder's equivariance assumption.

[0065] S5: Inject future latent variables into the historical latent variable sequence for the next moment, continuously predict future latent variables, and obtain the predicted meteorological field for multiple consecutive moments.

[0066] In the inference phase, a conditional autoregressive generative paradigm is employed to achieve rolling forecasts. For each future time step, the model samples from the denoising network based on historical latent variable sequences, terrain features, and other conditional information. The generated samples are then concatenated back into the input sequence to recursively generate longer-term forecasts. Multiple independent samplings can generate different forecast trajectories, forming an ensemble forecast. Terrain features continuously participate in each generation step, rather than just playing a role in the first step, thus ensuring the effectiveness of complex terrain constraints in long-term forecasts.

[0067] The combination of conditional diffusion and conditional autoregressive generation paradigms enables terrain conditions to continuously participate in each step of the rolling forecast generation, thereby maintaining the effectiveness of complex terrain constraints in long-term autoregressive inference and naturally supporting ensemble forecasts through multiple independent sampling.

[0068] like Figure 4As shown in the embodiment of this application, a regional weather forecasting system based on topographic and meteorological fields with symmetrical enhancement is provided, comprising: The meteorological field encoding module is used to input the current multivariate meteorological field into a pre-trained autoencoder to obtain historical latent variables.

[0069] The terrain field encoding module is used to input the static terrain field into the terrain encoder to obtain terrain features.

[0070] The gated fusion module is used to encode the historical latent variable sequence into a context feature sequence, and inject the terrain features into the context feature sequence through a gated method to obtain the terrain-enhanced context features.

[0071] The prediction module is used to input the enhanced terrain context features and noise scale into the conditional diffusion denoising network pre-trained with symmetric enhancement to obtain future prediction latent variables. During the training phase of the conditional diffusion denoising network, the same symmetric enhancement group action is applied to the historical latent variable sequence, the noisy target latent variable, and the terrain features, and the corresponding inverse group action is applied to the network output before calculating the denoising loss.

[0072] The meteorological decoding module is used to input the future prediction latent variables into the decoder to obtain the predicted meteorological field.

[0073] The loop module is used to inject future latent variables into the historical latent variable sequence for the next time moment, continuously predicting future latent variables to obtain the predicted meteorological field for multiple consecutive time moments.

[0074] In this optional embodiment, by combining topographic features and meteorological fields in a gated fusion manner, and introducing a conditional diffusion denoising network to apply symmetric enhancement and inverse group operations, the modulating effect of complex terrain on local weather is effectively captured, and the model robustness is improved. This achieves the effects of reducing computational resource consumption and improving forecast accuracy and timeliness. The gated fusion mechanism allows topographic features to dynamically adjust their weights according to the context. For example, it can enhance the influence of terrain on water vapor transport during periods of strong convection and weaken its effect under stable weather conditions, enabling the model to simulate local meteorological processes more precisely. The conditional diffusion denoising network increases the diversity of training data through symmetric enhancement operations (such as random translation and rotation), improving the model's generalization ability, and ensures the spatial consistency of prediction results through inverse group operations, avoiding forecast bias caused by sensitivity to boundary conditions in traditional numerical models.

[0075] Through the aforementioned technical solutions, the system significantly reduces computational resource requirements, avoids the cumbersome parameter tuning process in traditional numerical models, and enhances its ability to capture structural features of severe weather events such as strong convection. For example, in mountainous heavy precipitation forecasting scenarios, the system can accurately identify the guiding role of terrain on water vapor transport, generating high-resolution precipitation distribution predictions and providing reliable support for disaster early warning. Furthermore, the autoregressive prediction mechanism of the cyclic module ensures the continuity and stability of long-term forecasts, effectively solving the problem of difficulty in guaranteeing accuracy in long-term autoregressive scenarios using traditional methods.

[0076] Experimental results on the ERA5 reanalysis test set show that in the long-term series (240-hour duration) test from July to August: adding the topographic spectrum prior reduces the overall RMSE by 6.7%, CRPS by 11.2%, and the spread / skill ratio of the ensemble probability calibration improves from 0.835 to 1.000; the benefits are most significant in complex terrain areas (mountains and coastal zones). After superimposing the discrete cyclic swarm interaction symmetric enhancement along the longitude direction, the RMSE is further reduced by 0.80% in mountainous areas and 0.42% in coastal zones, showing a consistent directional gain at both short and long durations; the overall average index is relatively similar to that of the pure topographic model. The discrete cyclic swarm interaction scheme systematically outperforms the Fourier continuous phase rotation scheme (RMSE is 1.3% lower in July-August, 1.8% lower in marine areas, and 1.9% lower in winter).

[0077] In summary, compared with existing technologies, it has the following beneficial effects: 1. Compared with traditional regional numerical weather prediction models such as WRF and deterministic deep neural network prediction schemes, this invention does not require iteratively solving the original equations during the inference stage. It has fast inference speed and low computational overhead. Furthermore, it naturally supports ensemble prediction through conditional diffusion modeling, and can directly output multiple possible future trajectories, thereby improving the real-time performance, scalability, and uncertainty expression capabilities of operational deployment.

[0078] 2. Compared with spatial domain fusion schemes based on micro-topography, short-term precipitation ensemble schemes, Diffusion-ViT downscaling schemes, and spatiotemporal graph neural network downscaling schemes, this invention extracts multi-scale topographic frequency domain priors through spectral convolution and spectral domain low-pass downsampling, and injects them into the diffusion backbone in a learnable gating manner after matching the latent spatial resolution; it performs latent spatial rolling ensemble forecasts for regional multivariate meteorological fields, achieving an overall RMSE reduction of 6.7% and a CRPS reduction of 11.2% in long-term series tests from July to August, and improving the ensemble calibration Spread / Skill Ratio from 0.835 to 1.000.

[0079] 3. Compared with single-variable and local correction schemes such as wind speed correction for complex terrain and existing common spatial domain condition fusion methods, this invention proposes discrete symmetric enhancement on a finite cyclic group in the longitude direction. It does not introduce additional learnable parameters, does not rely on the strict isovariability assumption of the autoencoder, and has a compact and stable training sampling space. Experimentally, it is more robust than the continuous Fourier phase rotation scheme in global, marine and winter tests.

[0080] An electronic device provided in this application includes a memory and a processor; the memory is used to store a computer program; the processor is used to implement the regional weather forecasting method based on topographic field and meteorological field symmetry enhancement as described above when the computer program is executed.

[0081] This application provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the regional weather forecasting method based on topographic and meteorological fields with symmetric enhancement as described above.

[0082] In this embodiment, the beneficial effects of the electronic device and the computer-readable storage medium are similar to those of the regional weather forecasting method based on symmetrical enhancement of topographic and meteorological fields described above, and will not be repeated here.

[0083] The present invention describes electronic devices that can serve as servers or clients of this application, which are examples of hardware devices that can be applied to various aspects of this application. Electronic devices are intended to represent various forms of digital electronic computer devices, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. Electronic devices can also represent various forms of mobile devices, such as personal digital assistant devices, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the application described and / or claimed herein.

[0084] Electronic devices include a computing unit that can perform various appropriate actions and processes based on a computer program stored in read-only memory (ROM) or loaded from a storage unit into random access memory (RAM). The RAM can also store various programs and data required for device operation. The computing unit, ROM, and RAM are interconnected via a bus. Input / output (I / O) interfaces are also connected to the bus.

[0085] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc. In this application, the separately described modules may or may not be physically separate. Some or all of the modules can be selected to achieve the purpose of the embodiments of this application according to actual needs. Furthermore, the functional modules in the various embodiments of this application can be integrated into one processing unit, or each module can exist physically separately, or two or more modules can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0086] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0087] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A regional weather forecasting method based on symmetrical enhancement of topographic and meteorological fields, characterized in that, include: The current multivariate weather field is input into a pre-trained autoencoder to obtain historical latent variables; The static terrain field is input into the terrain encoder to obtain the terrain features; The historical latent variable sequence is encoded into a context feature sequence, and the terrain features are injected into the context feature sequence through a gating method to obtain the terrain-enhanced context features. The enhanced contextual features and noise scale are input into a conditional diffusion denoising network pre-trained with symmetric enhancement to obtain future prediction latent variables. The future prediction latent variables are then input into a decoder to obtain the predicted weather field. During the training phase of the conditional diffusion denoising network, the same symmetric enhancement group action is applied to the historical latent variable sequence, the noisy target latent variable, and the terrain features. The corresponding inverse group action is applied to the network output before the denoising loss is calculated. By injecting future latent variables into the historical latent variable sequence for the next moment, and continuously predicting future latent variables, the predicted meteorological field for multiple consecutive moments is obtained.

2. The regional weather forecasting method based on topographic and meteorological fields with symmetrical enhancement as described in claim 1, characterized in that, The terrain encoder includes a linear lifting layer, a physical resolution spectral convolutional block, a spectral domain downsampling layer, a latent spatial resolution spectral convolutional block, and a projection head; in, This represents the initial hidden features after linear enhancement. Indicates a linear lifting layer. Represents a static terrain field. Indicates the first The output of a physical resolution spectral convolution block Indicates the first -1 physical resolution spectral convolution block output Indicates the first A physical resolution spectral convolution block This indicates the number of spectral convolutional blocks stacked at the physical resolution stage. This represents the hidden features that fall within the latent spatial resolution after spectral domain downsampling. Indicates the first -1 physical resolution spectral convolution block output Indicates a spectral domain downsampling layer. This represents the output of the r-th latent spatial resolution spectral convolution block. This represents the output of the (r-1)th latent spatial resolution spectral convolution block. This represents the input to the first latent spatial resolution spectral convolutional block. , This represents the r-th latent spatial resolution spectral convolutional block. This represents the number of spectral convolutional blocks stacked in the latent spatial resolution stage. This represents the terrain features output by the projector. Indicates the first -1 latent spatial resolution spectral convolution block output This refers to the projection head.

3. The regional weather forecasting method based on topographic and meteorological fields for symmetrical enhancement as described in claim 2, characterized in that, The output of the spectral convolution block is in, This represents the output of the spectral convolution block. This represents the input to the spectral convolution block. ( ) represents the discrete spherical Fourier transform. ( ) represents the discrete spherical inverse Fourier transform. This indicates a learnable frequency domain filter. This represents a channel-by-channel, mode-by-mode product. This represents a multilayer perceptron.

4. The regional weather forecasting method based on topographic and meteorological fields for symmetrical enhancement as described in claim 1, characterized in that, The enhanced terrain context features are in, This represents the contextual features after terrain enhancement. Represents the context feature sequence, This represents the sigmoid activation function. Indicates a learnable scalar. Represents the lexical embedding mapping, Indicates terrain features.

5. The regional weather forecasting method based on topographic and meteorological fields with symmetrical enhancement as described in claim 1, characterized in that, Pre-training of the conditional diffusion denoising network is performed using methods including: Noise is added to the target latent variable to obtain the noisy target latent variable; During training, a translation is randomly generated according to a uniform distribution, and the symmetric augmented random cluster element corresponding to the translation is selected from a finite cyclic group. Symmetric augmented random cluster elements are applied to the noisy target latent variables and the terrain-enhanced context features, and then input into the denoising network layer. A translation inverse group operation is applied to the output of the denoising network layer to obtain the future prediction latent variables. The loss of the conditional diffusion denoising network is obtained based on the future prediction latent variables and the target latent variables. The parameters of the conditional diffusion denoising network are adjusted based on the loss until the loss converges, resulting in the trained conditional diffusion denoising network.

6. The regional weather forecasting method based on topographic and meteorological fields for symmetrical enhancement as described in claim 5, characterized in that, In the inference phase, the latent variables for future prediction are: During the training phase, the latent variables for future prediction are: in, Indicates latent variables for future prediction. This indicates the translation inverse group operation. ( ) represents a denoising network layer. Represents symmetric enhanced random cluster elements. Represents the latent variables of a noisy target. This represents the maximum value of the noise scale. Indicates the noise level. This represents the contextual features after terrain enhancement.

7. A regional weather forecasting system based on symmetrical enhancement of topographic and meteorological fields, characterized in that, include: The meteorological field encoding module is used to input the current multivariate meteorological field into a pre-trained autoencoder to obtain historical latent variables; The terrain field encoding module is used to input the static terrain field into the terrain encoder to obtain terrain features; The gated fusion module is used to encode the historical latent variable sequence into a context feature sequence, and inject the terrain features into the context feature sequence through a gated method to obtain the terrain-enhanced context features. The prediction module is used to input the enhanced terrain context features and noise scale into the conditional diffusion denoising network pre-trained by symmetric enhancement to obtain future prediction latent variables. In the training phase of the conditional diffusion denoising network, the same symmetric enhancement group action is applied to the historical latent variable sequence, the noisy target latent variable and the terrain features, and the corresponding inverse group action is applied to the network output before calculating the denoising loss. The meteorological decoding module is used to input the future prediction latent variables into the decoder to obtain the predicted meteorological field; The loop module is used to inject future latent variables into the historical latent variable sequence for the next time moment, continuously predicting future latent variables to obtain the predicted meteorological field for multiple consecutive time moments.

8. An electronic device, characterized in that, Including memory and processor; The memory is used to store computer programs; The processor is configured to, when executing the computer program, implement the regional weather forecasting method based on topographic and meteorological fields with symmetric enhancement as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the regional weather forecasting method based on topographic and meteorological fields for symmetrical enhancement as described in any one of claims 1 to 6.

Images