Meteorological forecasting method and medium based on feature fusion and region revision
The meteorological forecasting method based on feature fusion and regional correction utilizes a transformer network with 3D convolution and window attention mechanisms to generate high-resolution forecast results, solving the problems of insufficient regional accuracy and insufficient information fusion in existing technologies, and achieving seamless transformation between high-precision meteorological forecasts and operational decisions.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SINE SPACE (ANHUI) TECHNOLOGY CO LTD
- Filing Date
- 2026-01-15
- Publication Date
- 2026-07-21
Smart Images
Figure CN121524958B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of artificial intelligence and weather forecasting technology, and in particular to a weather forecasting method and medium based on feature fusion and regional correction. Background Technology
[0002] Numerical weather prediction is the core of meteorological operations. Traditional methods are based on atmospheric physical equations and are solved discretely on supercomputers. Although they can predict large-scale circulation, the computational cost is extremely high, and they are affected by errors in parameterization schemes, limiting their ability to predict small- and medium-scale weather processes and detailed elements of specific regions.
[0003] In recent years, end-to-end weather forecasting models based on deep learning have made groundbreaking progress. These models learn complex spatiotemporal evolution patterns directly from historical meteorological reanalysis data, enabling the generation of global or regional forecasts with extremely low inference costs. Existing technical solutions mainly fall into two categories: The first category consists of global forecasting models, such as GraphCast and the Pangu meteorological model. These models typically employ architectures based on graph neural networks or Vision Transformers, demonstrating capabilities comparable to or even surpassing traditional numerical models in global medium-range forecasts several days in advance. However, the output resolution of these general-purpose large models is usually relatively coarse (e.g., 0.25° × 0.25°). Their design goal is to capture global atmospheric circulation, making it difficult to directly provide kilometer-level high-precision forecasts of specific elements (such as wind speed, precipitation, and temperature) that meet the operational needs of specific regions such as watersheds, wind farms, and urban clusters, resulting in a "not clearly visible" problem. The second category comprises deep learning downscaling models targeting specific regions or elements. These models use the output of the first type of large models or low-resolution analysis fields as input, and then refine the spatial representation through convolutional neural networks. However, these methods are often used as independent post-processing modules, which are separate from the pre-prediction models. They fail to make full use of the intermediate features and multi-scale information flow within the prediction models, and usually lack effective fusion modeling of key environmental auxiliary information such as multi-source real-time observations of the target area and complex terrain. This results in limited correction capabilities, inaccurate predictions, and weak integration with local physical mechanisms.
[0004] In summary, the existing technologies have the following obvious shortcomings: (1) The general meteorological large model is not capable of depicting regional details, and the forecast accuracy and practicality are poor when applied directly; (2) The independent downscaling post-processing model fails to form an organic synergy with the basic forecast model, and the feature utilization is insufficient; (3) It fails to systematically and dynamically integrate the multi-source environmental information of the target area into the correction process, which limits the pertinence and accuracy of the forecast; (4) The forecast results are disconnected from the business decision indicators of vertical industries such as agriculture and energy, requiring users to perform secondary conversion, and the application threshold is high.
[0005] Therefore, there is an urgent need for an end-to-end intelligent meteorological forecasting scheme that can couple high-precision basic forecasts with dynamic regional corrections and directly output operationally usable results. Summary of the Invention
[0006] The purpose of this invention is to overcome the shortcomings of the existing technology. To achieve the above objectives, a weather forecasting method and medium based on feature fusion and regional correction are adopted to solve the problems mentioned in the background technology.
[0007] A weather forecasting method based on feature fusion and regional correction includes the following steps: Obtain the initial weather forecast result generated by the basic weather forecast model, which is configured to generate the initial weather forecast result by fusing spatiotemporal features and modeling long-range dependencies; The initial weather forecast results and multi-source environmental auxiliary data of the target area are input into the regional fine correction module; The regional refinement correction module performs deviation correction and spatial downscaling on the initial weather forecast results based on the multi-source environmental auxiliary data to generate high-resolution meteorological element forecast results for the target area.
[0008] As a further aspect of the present invention: the basic large-scale weather forecasting model includes: A 3D convolutional encoder-decoder architecture is used to extract and fuse multi-scale spatiotemporal meteorological features; and The core modeling layer, coupled with the encoder-decoder architecture, employs a transformer network based on a window attention mechanism to enhance the modeling of remote spatial dependencies in meteorological data.
[0009] As a further aspect of the present invention: the 3D convolution-based encoder-decoder architecture includes: The encoder employs a multi-level downsampling structure, with each level containing a 3D convolutional layer. Its convolutional kernels use a stride greater than 1 in the spatial dimension and maintain a stride of 1 in the temporal dimension. The decoder employs a multi-level upsampling structure symmetrical to the encoder, with each level containing a 3D transposed convolutional layer, and fuses features of the corresponding level of the encoder through skip connections.
[0010] As a further aspect of the present invention: each level output of the encoder is connected to a multi-scale feature fusion module, the multi-scale feature fusion module being configured as follows: The application uses feature extraction paths with different receptive fields in parallel and employs an attention weighting mechanism to fuse the features extracted by each path.
[0011] As a further aspect of the present invention: the transformer network based on the window attention mechanism is SwinTransformer or a variant thereof, which introduces an adaptive relative position bias that incorporates meteorological prior knowledge into its self-attention calculation.
[0012] As a further aspect of the present invention: the region refinement correction module adopts an encoder-decoder framework, which is configured as follows: The input initial weather forecast result sequence and the multi-source environmental auxiliary data are encoded and feature-represented. The encoded features are modeled temporally using an improved transformer module, and the cross-attention mechanism is used in the transformer module to achieve interactive fusion of weather forecast features and environmental auxiliary features. Based on the fused feature prediction correction, and by combining the feature prediction correction with the initial weather forecast result, the corrected forecast result is obtained.
[0013] As a further aspect of the present invention, the improved converter module introduces a position bias based on meteorological prior knowledge in the attention calculation.
[0014] As a further aspect of the present invention: the feature prediction correction is implemented through a fully connected output layer, and the final correction result is obtained by adding the initial weather forecast result to the correction amount scaled by learnable parameters.
[0015] As a further aspect of the present invention, it also includes: The high-resolution meteorological element forecast results are converted into decision indicators corresponding to specific business scenarios through predefined industry intelligent mapping rules.
[0016] The second aspect of the technical solution: a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the method as described in any of the above.
[0017] Compared with the prior art, the present invention has the following technical advantages: The aforementioned technical solution generates initial forecast results using a large-scale meteorological forecasting model capable of fusing spatiotemporal features and modeling remote dependencies. Subsequently, the initial forecast results, along with multi-source environmental auxiliary data for the target area, are input into a regional refinement correction module. This module, through deep learning algorithms, deeply fuses environmental data to perform bias correction and spatial downscaling on the initial forecast, ultimately outputting high-resolution meteorological element forecast results for the target area. On one hand, it overcomes the shortcomings of general-purpose large-scale models in specific areas, which are often "unclear and inaccurate," generating high-precision, highly targeted forecast results at the kilometer level through dynamic regional correction. On the other hand, this method can provide downstream operations with directly applicable high-resolution element products. By integrating industry mapping interfaces, it can seamlessly transform meteorological forecasts into operational decision indicators, achieving end-to-end connectivity from raw data to operational value, significantly improving its practical efficiency and deployment value in vertical industries. Attached Figure Description
[0018] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings: Figure 1 This is a schematic diagram illustrating the steps of the weather forecasting method according to an embodiment of this application; Figure 2 This is a schematic diagram of a large-scale meteorological forecasting model based on an embodiment of this application. Figure 3 This is a schematic diagram of a region refinement correction module according to an embodiment of this application. Detailed Implementation
[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0020] Please refer to Figure 1 In this embodiment of the invention, a weather forecasting method based on feature fusion and regional correction includes the following steps: Step S1: Obtain the initial weather forecast result generated by the basic weather forecast model, wherein the basic weather forecast model is configured to generate the initial weather forecast result by fusing spatiotemporal features and modeling long-range dependencies; like Figure 2 As shown, the diagram is a schematic of the forecast model; In this embodiment, the basic weather forecasting model includes: A 3D convolutional encoder-decoder architecture is used to extract and fuse multi-scale spatiotemporal meteorological features; and The core modeling layer, coupled with the encoder-decoder architecture, employs a transformer network based on a window attention mechanism to enhance the modeling of remote spatial dependencies in meteorological data.
[0021] In this embodiment, the encoder-decoder architecture based on 3D convolution includes: The encoder employs a multi-level downsampling structure, with each level containing a 3D convolutional layer. Its convolutional kernels use a stride greater than 1 in the spatial dimension and maintain a stride of 1 in the temporal dimension. The decoder employs a multi-level upsampling structure symmetrical to the encoder, with each level containing a 3D transposed convolutional layer, and fuses features of the corresponding level of the encoder through skip connections.
[0022] In a specific implementation, a 3D convolutional encoder-decoder architecture design is shown. The encoder employs a four-level downsampling structure, with each level containing two 3D convolutional layers. The convolutional kernel size is 3×3×3, and the stride is 1×2×2. The temporal dimension maintains a stride of 1 to preserve temporal information, while the spatial dimension is downsampled with a stride of 2. The number of encoder channels is [32, 64, 128, 256], and the feature map spatial size is halved after each downsampling level.
[0023] The decoder employs a four-level upsampling structure, symmetrical to the encoder. Each level contains two 3D transposed convolutional layers with a kernel size of 3×3×3, a stride of 1×2×2, and channel numbers of [256, 128, 64, 32]. The decoder fuses with the corresponding level features of the encoder through skip connections to ensure effective recovery of detailed information.
[0024] In this embodiment, each level output of the encoder is connected to a multi-scale feature fusion module, which is configured as follows: The application uses feature extraction paths with different receptive fields in parallel and employs an attention weighting mechanism to fuse the features extracted by each path.
[0025] Specifically, the multi-scale feature fusion mechanism is as follows: At each stage output of the encoder, multiple dilated convolutional modules with different dilation rates (1, 2, 4, and 8) are applied in parallel to capture meteorological features from different receptive fields. After adjusting the number of channels using 1×1×1 convolutions, feature maps at each scale are fused using a channel attention-weighted feature fusion method.
[0026] in, The channel attention weights are obtained through learning. This represents the feature map at the i-th scale. The multi-scale feature fusion module outputs a feature map with the same size as the corresponding layer of the encoder, ensuring compatibility with skip connections.
[0027] In this embodiment, the transformer network based on the window attention mechanism is the Swing Transformer or a variant thereof, which introduces an adaptive relative position bias that incorporates meteorological prior knowledge into its self-attention calculation.
[0028] In a specific implementation, the core modeling layer based on the Swing Transformer variant is: Perform a layered Swing Transformer block design; The core modeling layer contains L Swing Transformer blocks, each consisting of alternating window multi-head self-attention (W-MSA) and shifted window multi-head self-attention (SW-MSA).
[0029] Based on meteorological adaptive relative position offset; To address the spatial continuity of meteorological data, an adaptive relative position offset for meteorological data is designed:
[0030] in, This is a learnable relative position bias matrix. To determine the location bias based on prior meteorological knowledge, factors such as geographical elevation and climate zoning are considered.
[0031] Specifically, a self-developed large-scale weather forecasting model was constructed as the basic framework. Key architectural settings were implemented for this model: (1) An encoder-decoder architecture based on 3D convolution is adopted to effectively capture the spatiotemporal evolution characteristics of atmospheric variables through three-dimensional convolution kernels; a multi-scale feature fusion mechanism is introduced to extract meteorological features of different scales in parallel during the encoding process, which can simultaneously model the interaction between large-scale circulation systems and small- and medium-scale weather processes. (2) A variant architecture based on Swing Transformer is adopted in the core modeling layer. Its hierarchical attention mechanism and shift window design are used to enhance the model’s ability to model remote spatial dependencies and boundary conditions of complex weather systems. (3) Based on this self-developed large model, a high-precision regional correction module is further integrated, namely the regional fine correction module. This module takes the output of the previous large model as input, deeply integrates the real-time observation, topography and other multi-source data of the target area, and performs deviation correction and spatial downscaling through dynamic adaptive algorithm, and finally generates regional fine meteorological element forecasts at the kilometer level or even higher resolution.
[0032] Step S2: Input the initial weather forecast results and multi-source environmental auxiliary data of the target area into the regional fine correction module; Step S3: Using the regional refinement correction module, the initial weather forecast results are corrected for deviations and spatially downscaled based on the multi-source environmental auxiliary data to generate high-resolution meteorological element forecast results for the target area.
[0033] like Figure 3 As shown in the figure, this is a schematic diagram of the regional fine-grained correction module; In this embodiment, the region refinement correction module adopts an encoder-decoder framework, which is configured as follows: The input initial weather forecast result sequence and the multi-source environmental auxiliary data are encoded and feature-represented. The encoded features are modeled temporally using an improved transformer module, and the cross-attention mechanism is used in the transformer module to achieve interactive fusion of weather forecast features and environmental auxiliary features. Based on the fused feature prediction correction, and by combining the feature prediction correction with the initial weather forecast result, the corrected forecast result is obtained.
[0034] In this embodiment, the improved converter module introduces a position bias based on meteorological prior knowledge in the attention calculation.
[0035] In this embodiment, the feature prediction correction is implemented through a fully connected output layer, and the final correction result is obtained by adding the initial weather forecast result to the correction amount scaled by learnable parameters.
[0036] This embodiment also includes: The high-resolution meteorological element forecast results are converted into decision indicators corresponding to specific business scenarios through predefined industry intelligent mapping rules.
[0037] In a specific implementation, the regional fine-grained correction module adopts a hierarchical hybrid architecture and uses deep learning methods to fine-grained correct the initial large model prediction results.
[0038] The correction module employs an encoder-decoder framework, using the prediction sequence from the preceding large model as the core input, combined with environmental auxiliary features, and predicts the correction amount through a deep learning model. The overall processing flow can be represented as follows:
[0039] in, This is the initial large model forecast sequence. As an environmental auxiliary feature, The correction function obtained through learning, θFor the learnable parameters of the correction function, This is the final correction result.
[0040] The module receives two main input sources: the forecast time series data of the preceding large model and environmental auxiliary features. The forecast sequence is first converted into a high-dimensional representation through a fully connected embedding layer, while the environmental features are encoded through a multilayer perceptron to form a unified feature space.
[0041] The encoded features are input into an improved Transformer module for deep temporal modeling. This module contains multi-head self-attention layers and feedforward neural network layers, and training stability is ensured through layer normalization and residual connections. In particular, a relative position bias matrix of meteorological priors is introduced into the attention calculation, enabling the model to better distinguish the meteorological significance at different time steps.
[0042] The features after temporal modeling are used to achieve deep interaction between forecast features and environmental features through a cross-attention mechanism, forming a fused feature representation. This fused feature is then used to predict corrections through a fully connected output layer.
[0043]
[0044] in, and This is the weight matrix of the fully connected layer. and This is the bias vector of the fully connected layer. The weighted fusion features are the input. For activation function, This is the correction amount. The final result is:
[0045] in, For learnable scaling parameters, This is the initial large model forecast sequence. This is the final correction result.
[0046] The second aspect of the technical solution: a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the method as described in any of the above.
[0047] This invention fundamentally improves the ability to characterize key weather processes through a self-developed multi-scale basic model architecture; combined with subsequent regional correction modules for dynamic downscaling and bias correction, it achieves accurate forecasts at the kilometer level or even higher resolution in target areas (such as wind farms and agricultural production areas), effectively overcoming the core defects of existing models that are "unclear and inaccurate". The solution offers significant end-to-end business value and strong practicality. It innovatively incorporates an industry-specific intelligent mapping interface, seamlessly transforming high-precision weather forecasts into direct business decision-making indicators (such as power generation and irrigation recommendations), achieving end-to-end value transfer from "meteorological data" to "production capacity insights." This not only eliminates the complex process of secondary modeling for users but also enables forecast results to directly support operations and decision-making, greatly enhancing its practical value and application efficiency in vertical industries.
[0048] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention. The scope of the invention is defined by the appended claims and their equivalents, all of which should be included within the scope of protection of the invention.
Claims
1. A weather forecasting method based on feature fusion and regional correction, characterized in that, Includes the following steps: Obtain the initial weather forecast result generated by the basic weather forecast model, which is configured to generate the initial weather forecast result by fusing spatiotemporal features and modeling long-range dependencies; The initial weather forecast results and multi-source environmental auxiliary data of the target area are input into the regional fine correction module; The region refinement correction module adopts an encoder-decoder framework, which is configured as follows: The input initial weather forecast result sequence and the multi-source environmental auxiliary data are encoded and feature-represented. The encoded features are modeled temporally using an improved transformer module, and the cross-attention mechanism is used in the transformer module to realize the interactive fusion of meteorological forecast features and environmental auxiliary features. The improved transformer module introduces a position bias based on meteorological prior knowledge in the attention calculation. Based on the fused feature prediction correction, and by combining the feature prediction correction with the initial weather forecast result, the corrected forecast result is obtained; The regional refinement correction module performs deviation correction and spatial downscaling on the initial weather forecast results based on the multi-source environmental auxiliary data to generate high-resolution meteorological element forecast results for the target area.
2. The weather forecasting method based on feature fusion and regional correction according to claim 1, characterized in that, The aforementioned basic large-scale weather forecasting model includes: A 3D convolutional encoder-decoder architecture is used to extract and fuse multi-scale spatiotemporal meteorological features; and The core modeling layer, coupled with the encoder-decoder architecture, employs a transformer network based on a window attention mechanism to enhance the modeling of remote spatial dependencies in meteorological data.
3. The weather forecasting method based on feature fusion and regional correction according to claim 2, characterized in that, The 3D convolution-based encoder-decoder architecture includes: The encoder employs a multi-level downsampling structure, with each level containing a 3D convolutional layer. Its convolutional kernels use a stride greater than 1 in the spatial dimension and maintain a stride of 1 in the temporal dimension. The decoder employs a multi-level upsampling structure symmetrical to the encoder, with each level containing a 3D transposed convolutional layer, and fuses features of the corresponding level of the encoder through skip connections.
4. The weather forecasting method based on feature fusion and regional correction according to claim 3, characterized in that, Each stage output of the encoder is connected to a multi-scale feature fusion module, which is configured as follows: The application uses feature extraction paths with different receptive fields in parallel and employs an attention weighting mechanism to fuse the features extracted by each path.
5. The weather forecasting method based on feature fusion and regional correction according to claim 2, characterized in that, The transformer network based on the window attention mechanism is a variant of the Swing Transformer, which introduces an adaptive relative position bias that incorporates meteorological prior knowledge into its self-attention computation.
6. The weather forecasting method based on feature fusion and regional correction according to claim 1, characterized in that, The feature prediction correction is implemented through a fully connected output layer, and the final correction result is obtained by adding the initial weather forecast result to the correction amount scaled by learnable parameters.
7. The weather forecasting method based on feature fusion and regional correction according to claim 1, characterized in that, Also includes: The high-resolution meteorological element forecast results are converted into decision indicators corresponding to specific business scenarios through predefined industry intelligent mapping rules.
8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 7.