A physical perception fusion bridging model, a weather prediction method and related devices

Abstract

This application belongs to the field of artificial intelligence and numerical weather prediction fusion technology. Addressing the technical problems in existing technologies where there is insufficient physical consistency, poor multivariate synergy, inadequate observation fusion, and difficulty in effectively suppressing systematic biases between the output of large-scale AI weather prediction models and the input of regional numerical models, this application proposes a physical perception fusion bridging model, a weather prediction method, and related devices. A multi-scale encoder extracts multi-scale meteorological field features from the output of the large-scale AI weather prediction model; a physical constraint decoder decodes the meteorological field features to generate a high-resolution meteorological field that conforms to physical conservation laws, with a physical conservation penalty term introduced during training; and an observation fusion layer fuses multi-source observation data and the high-resolution meteorological field to obtain a high-resolution initial field. This application achieves seamless integration between the large-scale AI weather prediction model and the regional numerical model, ensuring physical consistency and fusing multi-source observation data.

Description

Technical Field

[0001] This application belongs to the field of artificial intelligence and numerical weather prediction fusion technology, specifically involving a physical perception fusion bridging model, a meteorological prediction method and related devices. Background Technology

[0002] With the development of artificial intelligence (AI) technology, large-scale AI weather forecasting models have made significant progress in global-scale weather prediction. For example, models such as Pangu, Fuxi, and GraphCast can achieve highly efficient global weather field predictions. However, in practical operational applications, refined regional weather forecasts still mainly rely on regional numerical weather prediction models, such as WRF (Weather Research and Forecasting Model) and MPAS (Model for Prediction Across Scales). This is because regional numerical models still play a crucial role in local high-resolution forecasting, adaptation to complex terrain, and operationally refined forecasting.

[0003] However, significant differences remain between the output of large-scale AI weather forecasting models and the input requirements of regional numerical models, making direct and effective integration difficult. Specifically: First, large-scale AI weather forecasting models typically output only a limited number of meteorological variables, while regional numerical models usually require a complete three-dimensional initial field, resulting in variable mismatch. Second, the output of large-scale AI weather forecasting models may exhibit physical inconsistencies, such as failure to meet constraints like mass and energy conservation. Directly using these as the initial field for regional numerical models can easily lead to instability in the regional numerical model's integrals. Third, the output of large-scale AI weather forecasting models often does not fully integrate real-time observation data, resulting in deviations between the generated initial field and the current actual atmospheric conditions, affecting the timeliness and accuracy of forecasts. Fourth, the output resolution of large-scale AI weather forecasting models is typically low, while regional numerical models require a higher-resolution input field, thus also presenting a resolution mismatch problem.

[0004] In existing technologies, methods such as horizontal interpolation, vertical interpolation, variable completion, data assimilation, and bias correction are commonly used to transform the output of large AI weather forecasting models to address the aforementioned integration issues. However, most of these methods only address single aspects of resolution matching, variable completion, or local error correction, lacking a holistic bridging mechanism that meets the input requirements of regional numerical models. Therefore, problems remain, including insufficient physical consistency, poor multivariate synergy, inadequate observational fusion, and difficulty in effectively suppressing systematic biases, making it difficult to provide reliable high-resolution initial fields and boundary conditions for the stability of regional numerical models.

[0005] Therefore, there is an urgent need for a technical solution that can effectively bridge the gap between large AI weather forecasting models and regional numerical models to ensure physical consistency, integrate multi-source observation data, and have dynamic correction capabilities, thereby providing more reliable high-resolution initial fields and boundary conditions for regional numerical models. Summary of the Invention

[0006] This application addresses the technical problems in existing technologies, such as insufficient physical consistency, poor multivariate synergy, inadequate observation fusion, and difficulty in effectively suppressing systematic biases, by providing a physical perception fusion bridging model, a weather forecasting method, and related devices.

[0007] To achieve the above objectives, this application adopts the following technical solution: Firstly, this application proposes a physical perception fusion bridging model, which is set between an AI weather forecasting large-scale model and a regional numerical model, including: A multi-scale encoder is used to extract multi-scale meteorological field features from the output of a large AI weather forecasting model. A physical constraint decoder is used to decode the meteorological field features and generate a high-resolution meteorological field that conforms to the physical conservation law. The observation fusion layer is used to fuse multi-source observation data and high-resolution meteorological fields to obtain a high-resolution initial field as input to the regional numerical model. The physical constraint decoder introduces a physical conservation penalty term during training; the physical conservation penalty term includes at least one of energy conservation, mass conservation, and momentum conservation.

[0008] Secondly, this application proposes a weather forecasting method, including: Meteorological field data is obtained through a large AI-powered weather forecasting model. The meteorological field data is processed using the aforementioned physical perception fusion bridging model to obtain a high-resolution initial field as input for the regional numerical model. High-resolution initial fields are input into the regional numerical model to obtain regional weather forecast results.

[0009] Thirdly, this application proposes a weather forecasting system, comprising: The data acquisition module is used to acquire meteorological field data through the AI ​​meteorological forecasting model; The bridging processing module is used to process the meteorological field data using the aforementioned physical perception fusion bridging model to obtain a high-resolution initial field as input to the regional numerical model. The forecast output module is used to input high-resolution initial fields into the regional numerical model to obtain regional weather forecast results.

[0010] Fourthly, this application proposes a computer program product, which includes a computer program that, when executed by a processor, implements the above-mentioned weather forecasting method.

[0011] Compared with the prior art, this application has the following beneficial effects: This application proposes a physical perception fusion bridging model, including a multi-scale encoder, a physical constraint decoder, and an observation fusion layer. The multi-scale encoder extracts multi-scale meteorological field features from the output of a large AI meteorological prediction model, and the physical constraint decoder decodes the meteorological field features to generate a high-resolution meteorological field that conforms to the physical conservation law. The observation fusion layer then fuses multi-source observation data with the high-resolution meteorological field to obtain a high-resolution initial field that can be directly used by regional numerical models. On the one hand, this application introduces at least one physical conservation penalty term from energy conservation, mass conservation, and momentum conservation terms during the training process of the physical constraint decoder. This ensures that the generated results balance data expressiveness and physical consistency during high-resolution reconstruction, effectively alleviating the model integral instability and spin-up oscillation problems caused by physical inconsistencies when the output of a large AI meteorological forecast model is directly integrated into a regional numerical model. On the other hand, extracting multi-scale meteorological field features through a multi-scale encoder helps retain large-scale background information and local details. Combining this with the physical constraint decoder to generate a high-resolution meteorological field improves the adaptability of the bridging results to the needs of refined regional forecasts. Furthermore, by fusing observational data from radar, satellites, ground stations, and radiosondes through an observational fusion layer, the obtained high-resolution initial field is closer to the current real atmospheric state, thereby improving the timeliness and reliability of the initial field and reducing the deviation between the initial field and the actual situation. According to actual experimental results, in a plain area of ​​a certain country, the average temperature error predicted by the stations decreased from 1.61 to 0.96, a reduction of approximately 40%, and the average wind speed error decreased from 0.72 to 0.41, a reduction of approximately 43%. In the central mountainous region of another area, the time-averaged temperature error decreased from 3.13 to 1.80, a reduction of approximately 42%, and the station-averaged wind speed error decreased from 0.87 to 0.32, a reduction of approximately 63%. Simultaneously, compared to methods that do not fuse observational data, the deviation between the initial field and the actual situation decreased by approximately 35%. This indicates that this application not only improves the forecast accuracy of key meteorological elements such as temperature and wind speed, but also has good applicability under different terrain conditions, including plains and complex mountainous areas, and its effect on constructing regional numerical model input fields is particularly prominent in complex regions.

[0012] Furthermore, this application utilizes a dynamic residual correction module, particularly based on the forecast error of the first 6 hours, to perform dynamic error correction and eliminate the systematic bias of the physical perception fusion bridging model.

[0013] Furthermore, based on the aforementioned physical perception fusion bridging model, this application proposes a meteorological forecasting method that possesses all the advantages of the aforementioned physical perception fusion bridging model. In particular, the generated high-resolution initial field can achieve a spatial resolution of 0.01°×0.01°, and its vertical structure includes 37 pressure layers, which can directly drive regional numerical models for refined forecasting, meeting operational needs such as wind farm power forecasting and local severe convection early warning.

[0014] Furthermore, this application also proposes a weather forecasting system and a computer program product, both of which possess all the advantages of the aforementioned physical perception fusion bridging model. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of the physical perception fusion bridging model in Embodiment 1 of this application; Figure 2 This is a schematic diagram illustrating how the physical conservation constraint mechanism optimizes model parameters through backpropagation during training in Embodiment 1 of this application. Figure 3 This is a schematic diagram of the physical perception fusion bridging model in Embodiment 2 of this application; Figure 4 This is a schematic diagram of the meteorological forecasting method described in this application; Figure 5 This is a comparison chart of the meteorological forecasting method in this application and EC HRES (European Centre for Medium-Range Weather Forecasts) in terms of regional temperature forecast errors; Figure 6 This is a schematic diagram of the meteorological forecasting system of this application. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of 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] In practical operations, large-scale forecast results typically require intermediate processing before being used to initiate regional numerical models, resulting in initial fields and boundary conditions suitable for the model. This intermediate processing involves not only spatial grid conversion but also the completion of multiple meteorological elements, vertical layer data organization, consistency coordination between different variables, and adjustments to match the current real-time conditions. Especially in complex weather processes, localized severe convective weather, areas significantly affected by mountainous terrain, and short-term nowcasting scenarios, regional numerical models are more sensitive to the quality of the input field. If the input field cannot simultaneously reflect the large-scale weather background, local fine structure, and real-time atmospheric conditions, it often affects the accuracy and stability of subsequent forecast results.

[0019] In existing processing methods, several relatively independent techniques are typically employed to convert front-end prediction results into the input field of regional numerical models. For example, spatial interpolation can be performed on low-resolution prediction results to obtain a finer-grid data format; then, based on empirical relationships or preset rules, elements required by the regional numerical model but not directly given in the front-end results are estimated; subsequently, corrections, adjustments, or local modifications are used to make the processed results as close as possible to the usage requirements of the regional numerical model. In some business scenarios, real-time observation data is also introduced to correct local areas, thereby improving the consistency between the input field and the current reality. While these approaches can achieve a certain degree of data format transition, they are mostly step-by-step and modular processing methods with limited coupling between different stages. They often focus more on the correction of a single objective and struggle to address multiple needs such as resolution improvement, variable reconstruction, physical relationship constraints, and reality fusion within a unified framework.

[0020] Regional models require input fields that are not limited to numerical values, but also consider the interrelationships between multiple variables and the continuity of spatial structure. For example, wind, pressure, temperature, and humidity fields are often intrinsically linked, and their distribution characteristics should be relatively consistent across different altitudes. If only a single variable is interpolated or corrected without simultaneously considering the relationships between other related variables, the processed result may improve a certain index, but the overall field structure may not be suitable for use in regional numerical models. Especially under high-resolution model operation, regional numerical models respond more significantly to local gradients, boundary layer structure, topographic effects, and small- and medium-scale weather systems. Therefore, if the input field lacks sufficient structural consistency, it can easily reduce the regional numerical model's ability to represent key weather processes. In addition, from an operational timeliness perspective, the input field not only needs to reflect forecast information but also needs to be as close as possible to the actual atmospheric conditions at the time of regional numerical model startup. Currently, there are abundant sources of actual observations, including satellite remote sensing, weather radar, automatic weather stations, radiosonde systems, and other types of observation data. Different observation methods have their own characteristics in terms of spatial coverage, temporal update frequency, and types of meteorological elements. How to retain the overall trend information of the front-end forecast results while fully absorbing the real-time state characteristics represented by multi-source observations, so that the resulting input field has both a large-scale forecast background and strong real-world fit, is one of the key issues that needs to be considered in the existing forecasting chain. Without an effective fusion mechanism, the forecast background field and local observation information can easily become disconnected, thus affecting the ability of regional numerical models to depict the current weather situation. Furthermore, in regional refined forecasting scenarios, there are significant differences in topographic conditions, underlying surface features, and weather system types between different regions, which places higher demands on the intermediate conversion stage. For example, there are significant differences in the spatial distribution characteristics and evolution patterns of meteorological elements between plains and mountains, coastal and inland areas, and arid and humid regions. If the intermediate processing methods lack the ability to adapt to regional characteristics, it may be difficult to fully reflect local structural information when introducing the front-end forecast results into the regional numerical model, affecting the final refined forecasting effect. Therefore, in the process of converting large-scale prediction results into regional numerical model input fields, a technical solution is needed that can take into account multi-scale information representation, variable organization and coordination, real-world information absorption, and regional adaptability, so as to improve the availability and reliability of regional numerical model input fields.

[0021] Therefore, based on the aforementioned business needs and current technological status, it is necessary to construct a bridging mechanism for the conversion of forecast results into regional numerical model input fields. This mechanism would enable the model to complete meteorological field feature extraction, high-resolution reconstruction, physical relation constraints, and multi-source observation fusion processing within a unified framework, thereby providing regional numerical models with input data that is more suitable for operational requirements.

[0022] Example 1 like Figure 1 The diagram shown is a schematic of the physical perception fusion bridging model of Embodiment 1, which may include: (1) Multiscale encoder.

[0023] The multi-scale encoder receives meteorological field data output from the AI ​​weather prediction model and extracts multi-scale features. It converts the raw meteorological information output by the AI ​​weather prediction model into encoded results that can characterize meteorological features at different spatial scales, so as to facilitate subsequent decoding, reconstruction, and observation fusion processing.

[0024] In practical applications, the meteorological field data output by large AI weather forecasting models can include variables such as geopotential height, temperature, specific humidity, and wind speed. In some embodiments of this application, the multi-scale encoder can adopt a layered architecture, including a bottom-level feature extraction layer, a middle-level feature fusion layer, and a high-level semantic encoding layer. This layered architecture can model the input meteorological field from three levels: local details, cross-scale correlations, and global semantics, thereby improving the ability to express complex meteorological field structures. Specifically, the bottom-level feature extraction layer uses convolutional kernels to extract local features; the middle-level feature fusion layer uses SwinTransformer blocks for multi-scale feature fusion; and the high-level semantic encoding layer uses a self-attention mechanism to extract global semantic features. The bottom-level feature extraction layer can perceive the local spatial distribution patterns in the input meteorological field, extracting texture features and gradient change features of variables such as temperature, wind speed, and humidity within local regions; the middle-level feature fusion layer can use SwinTransformer blocks to interactively fuse features at different scales, thereby enhancing the correlation between large-scale weather background and local disturbance information; and the high-level semantic encoding layer can further model long-distance spatial dependencies through a self-attention mechanism to extract global semantic features representing the evolution of the overall meteorological situation.

[0025] The output of the multi-scale encoder is a multi-scale feature pyramid. In practical applications, the large-scale AI weather prediction model can be the Pangu weather model, or other large-scale AI weather prediction models such as Fuxi and GraphCast. The multi-scale feature pyramid can be understood as the feature representation retained at different resolution levels. It contains both coarse-scale circulation background information and fine-scale local variation information, thus providing sufficient feature support for the subsequent physical constraint decoder to generate a high-resolution weather field.

[0026] (2) Physical constraint decoder.

[0027] The physical constraint decoder is used to decode the extracted features and generate a high-resolution meteorological field that conforms to the laws of physical conservation. Its function is to map the feature pyramid output by the multi-scale encoder into a high-resolution meteorological field that meets the input requirements of regional numerical models, while introducing physical constraints during training to make the output more consistent with the basic laws of atmospheric motion.

[0028] In some embodiments of this application, the physical constraint decoder may adopt a U-Net architecture, including an upsampling decoding layer, a skip connection layer, and a physical constraint layer. Specifically, the upsampling decoding layer is used to progressively restore spatial resolution, reconstructing the abstract features formed in the encoding stage into a high-resolution meteorological field; the skip connection layer is used to fuse the shallow detail features retained in the encoding stage with the high-level semantic features in the decoding stage to improve the ability of the reconstruction result to express local structure and boundary details; and the physical constraint layer is used to introduce physical conservation constraints during model training to make the generated result more numerically reasonable.

[0029] As an example, the formula for calculating the physical conservation penalty term can be:

[0030] in, This is a physical conservation penalty term. For energy conservation terms, For the mass conservation term, For momentum conservation, for The weighting coefficients, for The weighting coefficients, for The weighting coefficients.

[0031] The aforementioned physical conservation penalty term constrains whether the generated meteorological field features satisfy energy conservation, mass conservation, and momentum conservation, enabling the physical perception fusion bridging model to consider both reconstruction accuracy and physical consistency of the output results. The weight coefficients can be set according to training requirements to balance the influence of different physical constraints on the total loss.

[0032] Total loss function during training of the physical constraint decoder for:

[0033] in, This is the L2 reconstruction loss.

[0034] The above total loss function shows that, during the training of the physical perception fusion bridging model, on the one hand, the numerical difference between the generated result and the target high-resolution meteorological field is constrained by the L2 reconstruction loss, and on the other hand, the generated result is constrained by the physical conservation penalty term to meet the preset physical laws, so that the output of the physical constraint decoder has both reconstruction accuracy and physical rationality.

[0035] like Figure 2 The diagram illustrates how the physical conservation constraint mechanism optimizes model parameters through backpropagation during training, enabling the physical constraint decoder to learn and generate meteorological fields that conform to physical conservation laws. In other words, in each training iteration, the physical constraint decoder calculates the error based on the total loss function and continuously adjusts its parameters through backpropagation to gradually improve the reconstruction accuracy and physical consistency of the output high-resolution meteorological field, ultimately obtaining high-resolution meteorological field results suitable for regional numerical model inputs.

[0036] (3) Observation fusion layer.

[0037] The observation fusion layer is used to fuse multi-source observation data with the decoded high-resolution meteorological field. Its role is to further incorporate real-time observation information into the high-resolution meteorological field output by the physically constrained decoder, thereby improving the output result's ability to represent the current true atmospheric state.

[0038] In some embodiments of this application, the observation fusion layer can employ variational assimilation methods or attention mechanisms to fuse multi-source observation data. Specifically, when using a variational assimilation method, the decoded high-resolution meteorological field can be biased based on multi-source observation data by constructing an optimized relationship between the background field and the observation field; when using an attention mechanism, the importance of observation data from different sources can be adaptively weighted, thereby enhancing the contribution of key observation information to the fusion result. Through these methods, the observation fusion layer can achieve effective coupling between multi-source observation data and the high-resolution meteorological field.

[0039] As an example, multi-source observation data can include radar observations, satellite observations, ground station observations, and radiosonde observations. Different types of observation data can reflect atmospheric state information at different spatial scales, altitudes, and temporal resolutions. For instance, radar observations are useful for reflecting local precipitation systems and echo structures, satellite observations are useful for reflecting large-scale cloud systems and radiation characteristics, ground station observations can reflect near-surface temperature, wind speed, and air pressure, and radiosonde observations can provide vertical atmospheric profile information. Therefore, by fusing these multi-source observation data, the completeness and accuracy of the output meteorological field can be further improved.

[0040] The output of the observation fusion layer is a high-resolution initial field that fuses multi-source observation data. The fused high-resolution initial field can be used as input for regional numerical models, enabling them to inherit large-scale forecast information provided by AI meteorological prediction models at startup, while also fully incorporating the current atmospheric state reflected in real-time observations, thereby improving the accuracy and stability of subsequent refined regional forecasts.

[0041] Example 2 like Figure 3 The diagram shown is a schematic of the physical perception fusion bridging model in Embodiment 2. The difference from Embodiment 1 is that it also includes a dynamic residual correction module.

[0042] The dynamic residual correction module is used to perform dynamic residual correction on the output of the physical perception fusion bridging model based on short-term forecast errors, so as to further improve the consistency between the generated high-resolution initial field and the actual atmospheric state, and reduce the systematic bias in subsequent forecasts of regional numerical models.

[0043] The method for obtaining short-term forecast errors includes: using the high-resolution initial field obtained through the physical sensing fusion bridging model as input to obtain the short-term weather forecast results generated by the regional numerical model; comparing the generated short-term weather forecast results with multi-source observation data for the corresponding time period to obtain the short-term forecast error of the regional numerical model. Therefore, instead of statically correcting based solely on the output of the physical sensing fusion bridging model itself, a feedback-based dynamic residual correction mechanism is further constructed by utilizing the differences between the short-term forecast results generated after the actual operation of the regional numerical model and the actual observations. This allows the output of the physical sensing fusion bridging model to better adapt to the actual operating characteristics of the regional numerical model.

[0044] As an example, the short-term forecast error of a regional numerical model is the forecast error over the first 6 hours. The forecast error over the first 6 hours can well reflect the recent integral performance of the regional numerical model under the initial field driving force, and the error within this time range is usually strongly correlated with the quality of the initial field. Therefore, selecting the forecast error over the first 6 hours as the basis for dynamic residual correction is beneficial for more targeted identification of the bias components still existing in the output of the physical perception fusion bridging model. The short-term forecast error over the first 6 hours can include one or more of the following: temperature error, wind speed error, humidity error, and air pressure error; it can also be a comprehensive error index constructed based on multiple meteorological elements.

[0045] As one embodiment, during dynamic residual correction, the output of the physical perception fusion bridging model is dynamically corrected using learnable affine transformation parameters. Specifically, the learnable affine transformation parameters may include scaling and translation parameters, which are used to adaptively adjust the high-resolution initial field output by the physical perception fusion bridging model. This allows the dynamically corrected result to maintain the original spatial structure characteristics while further reducing the deviation associated with short-term forecast errors in regional numerical models. By introducing learnable affine transformation parameters, the dynamic residual correction module can adaptively correct the output of the physical perception fusion bridging model based on error feedback from different regions, different weather processes, and different forecast times, without relying on manual adjustments using fixed empirical coefficients.

[0046] In this embodiment, the dynamic residual correction module can be set after the observation fusion layer. It is used to receive the high-resolution meteorological field after fusing multi-source observation data and perform further dynamic residual correction on the high-resolution meteorological field in combination with the acquired regional numerical model short-term forecast error. The result processed by the dynamic residual correction module can be used as the final output high-resolution initial field input to the regional numerical model, or used to update the output of the physical sensing fusion bridging model in the next time interval online. Thus, this embodiment further introduces an error correction mechanism based on the regional numerical model operation feedback on the aforementioned physical sensing fusion bridging model. This enables the physical sensing fusion bridging model to not only realize the conversion from AI meteorological forecast results to regional model input fields, but also to dynamically optimize according to the actual short-term forecast performance of the regional numerical model, thereby improving the adaptability of the bridging results and the accuracy of subsequent regional refined forecasts.

[0047] Example 3 like Figure 4 The diagram shown is a schematic representation of the meteorological forecasting method proposed in this application. It may include: S101 obtains weather forecast field data through an AI weather prediction big data model; S102, the meteorological field data is processed using the above-mentioned physical perception fusion bridging model to obtain a high-resolution initial field as input to the regional numerical model.

[0048] S103 inputs the high-resolution initial field into the regional numerical model to obtain regional weather forecast results.

[0049] The following illustrates a specific application flow of the physical perception fusion bridging model of this application in a meteorological forecasting system. The meteorological field data output from the large AI meteorological forecasting model is processed by the physical perception fusion bridging model and converted into a high-resolution initial field suitable for regional numerical model input, further yielding regional meteorological forecast results. Specifically: (1) Obtain global-scale meteorological field data from AI meteorological prediction big data model.

[0050] First, the AI-powered weather forecasting model outputs global-scale or large-area-scale meteorological field data, which serves as the input data source for the physical perception fusion bridging model. Global-scale meteorological field data can include one or more meteorological elements such as geopotential height, temperature, specific humidity, wind speed, and air pressure, used to characterize the large-scale weather background and basic meteorological conditions at future moments.

[0051] (2) Extract multi-scale features through a multi-scale encoder to generate a feature pyramid.

[0052] The acquired global-scale meteorological field data is input into a multi-scale encoder, which performs hierarchical feature extraction on the input data. By encoding meteorological field features at different spatial scales, a feature pyramid is generated, which includes local detail features, mesoscale structural features, and large-scale background features. The feature pyramid can provide multi-level feature support for subsequent decoding and reconstruction of high-resolution meteorological fields.

[0053] (3) A high-resolution meteorological field with a resolution of 0.01°×0.01° and a vertical structure of 37 layers is generated by a physical constraint decoder.

[0054] The feature pyramid is input into the physical constraint decoder, which performs step-by-step decoding and spatial resolution recovery of the meteorological field features, generating a high-resolution meteorological field with a resolution of 0.01°×0.01° and a 37-layer vertical structure. It should be noted that this high-resolution meteorological field not only meets the requirements for refined regional forecasting in terms of horizontal resolution, but also adapts to the input requirements of regional numerical models in terms of vertical hierarchical structure. Furthermore, during training, the physical constraint decoder can constrain the output results through a physical conservation penalty term, making the generated high-resolution meteorological field more consistent with physical conservation laws.

[0055] (4) The observation data from multiple sources, such as radar, satellite, and ground station, are fused through the observation fusion layer.

[0056] The high-resolution meteorological field output from the physical constraint decoder is input into the observation fusion layer and fused with multi-source observation data from radar, satellites, and ground stations. By introducing real-time multi-source observation data, the high-resolution meteorological field can further approximate the current real atmospheric state while retaining the large-scale prediction trends provided by the AI ​​meteorological prediction model, thereby improving the timeliness and reliability of the initial field. In practical applications, in addition to multi-source observation data from radar, satellites, and ground stations, other types of observation data, such as radiosonde observations, can also be included.

[0057] (5) Dynamic residual correction based on the forecast error of the first 6 hours.

[0058] Dynamic residual correction is further applied to the high-resolution meteorological field after observation fusion. Dynamic residual correction can be based on the forecast error of the regional numerical model 6 hours in advance. That is, by using the difference between the short-term forecast results of the regional numerical model and the multi-source observation data of the corresponding period, possible deviations in the current high-resolution initial field can be identified, and the output of the physical sensing fusion bridging model can be adjusted accordingly. This can further reduce systematic errors and improve the adaptability of the high-resolution initial field to the regional numerical model.

[0059] (6) Input the high-resolution initial field into the regional numerical model and perform numerical integration to obtain the regional weather forecast results.

[0060] The high-resolution initial field obtained after the aforementioned steps is input into the regional numerical model. The regional numerical model then performs numerical integration calculations based on its dynamic framework and physical process parameterization scheme to finally obtain the regional meteorological forecast results for the target area. Specifically, the regional meteorological forecast results may include one or more forecast elements such as temperature field, wind field, precipitation field, humidity field, and pressure field, which can be used for subsequent operational forecasts or decision support.

[0061] It should be noted that the AI ​​weather forecasting large-scale model in this embodiment can be the same model as the AI ​​weather forecasting base model in the published invention patent. This AI weather forecasting base model can be the Pangu weather forecasting large-scale model or other AI weather forecasting-related base models. The regional numerical model can be mainstream numerical models such as WRF, MPAS, and GRAPES (Global / Regional Assimilation and Prediction System). The physical perception fusion bridging model in this embodiment has good model compatibility, adapting to different types of AI weather forecasting large-scale models and different types of regional numerical models, thus facilitating deployment and application in different weather forecasting operational systems.

[0062] Example 4 Example 4 illustrates a running example of the physical sensing fusion bridging model of this application. This example uses the Pangu-Weather model as the AI ​​weather forecasting model to obtain global-scale weather field data. WRF v4.5 (Weather Research and Forecasting Model version 4.5) is used as the regional numerical model in this example.

[0063] (1) Model building.

[0064] Multi-scale encoder: The Swin Transformer–U-Net structure is used to downsample the 1°×1° global scale meteorological field output by the Pangu-Weather model and extract multi-scale meteorological field features.

[0065] Physically Constrained Decoder: Upsampling generates a 37-layer initial field of 0.01° × 0.01°. The loss function includes L2 reconstruction loss and a physical conservation penalty term. The physical conservation penalty term includes energy conservation, mass conservation, and momentum conservation terms.

[0066] Observation fusion layer: During the fusion process, radar reflectivity, satellite brightness temperature, and ground station temperature, pressure, humidity and wind data are introduced as multi-source observation data and fused using the 3D-Var variational assimilation method.

[0067] Dynamic residual correction module: Runs every 6 hours to calculate the RMSE (Root Mean Square Error) of the forecast and observation of WRF v4.5 in the previous 6 hours. It generates dynamic residual correction factors through a lightweight MLP (Multi-Layer Perceptron) network and feeds them back to the physical perception fusion bridging model.

[0068] (2) Model training.

[0069] Using ECMWF (European Centre for Medium-Range Weather Forecasts) reanalysis data from 1980–2020 as ground truth, and the output of the Pangu-Weather model as the input to the physical perception fusion bridging model, the physical perception fusion bridging model was trained.

[0070] (3) Run.

[0071] Running every 6 hours, the Pangu-Weather model outputs a global-scale meteorological field. The physical sensing fusion bridging model generates an initial field of 0.01°×0.01° for a certain region of a certain country. Then, it drives WRF v4.5 to run and outputs a refined forecast of 0-72 hours. Finally, the dynamic residual correction module updates the parameters.

[0072] Compared to the Pangu-Weather model, this application demonstrates significant advantages in key forecast indicators, as verified. Taking a certain region as an example, this application reduces the station-average temperature error from 1.61 to 0.96, a reduction of approximately 40%, and the station-average wind speed error from 0.72 to 0.41, a reduction of approximately 43%. In another region, the time-average temperature error decreased from 3.13 to 1.80, a reduction of approximately 42%, and the station-average wind speed error decreased from 0.87 to 0.32, a reduction of approximately 63%. The results indicate that this application significantly improves the accuracy of temperature and wind speed forecasts using the Pangu-Weather model, with the most significant improvement in wind speed forecasting.

[0073] like Figure 5 The figure shown is a comparison of the regional temperature forecast error between the meteorological forecasting method of this application and EC HRES (European Centre for Medium-Range Weather Forecasts, High Resolution Forecasting). Specifically, Figure 5 Figures 'a' and 'c' show the comparison curves of the site-averaged 2m temperature (T2M (K)) over time. The horizontal axis "Time (h)" represents the forecast time in hours, and the vertical axis "T2M (K)" represents the 2m temperature in Kelvin. The black dots "Stations" in the figure represent the measured temperature data of the stations, the curve "PXW" represents the prediction result of the meteorological forecasting method in this application, and the curve "EC HRES" represents the prediction result of EC HRES. From the curve fit, it can be seen that the PXW curve corresponding to the method in this application is generally closer to the black station observation points, especially in the process of temperature peaks, troughs, and periodic changes, and the tracking of the measured temperature evolution trend is more accurate. In contrast, the EC HRES curve deviates significantly in some periods, indicating that its regional temperature forecast error is relatively larger. Figure 5 Figures 'b' and 'd' in the graph show a scatter plot comparing model predictions and observed values, using the regional average 2m temperature as the evaluation object. The horizontal axis, "Model data," represents model prediction data, and the vertical axis, "Observation data," represents observed data. The black dashed line in the graph represents the ideal consistency reference line, i.e., the distribution position corresponding to when the model predictions and observed values ​​are completely equal. The light green scatter points represent the results of the PXW method in this application, and the blue-purple scatter points represent the results of EC HRES. The closer the scatter points are to this dashed line, the more consistent the model predictions and observed values ​​are, and the smaller the forecast error. Furthermore, Figure 5Figures b and d also show the root mean square error (RMSE). Specifically, in a plain region of a certain country, the RMSE for PXW is 0.96, while the RMSE for EC HRES is 2.28; in a mountainous region in the central part of the country, the RMSE for PXW is 0.69, while the RMSE for EC HRES is 1.28. It should also be noted that... Figure 5 In the diagram, 'a' and 'b' correspond to a plain region in a certain country, while 'c' and 'd' correspond to a mountainous region in the central part of a certain area. The above results show that, regardless of whether it is a plain or mountainous region, the temperature prediction error of the meteorological forecasting method in this application is significantly lower than that of EC HRES, indicating that the method in this application can more accurately characterize regional temperature change features, and has higher regional temperature forecast accuracy and better environmental adaptability.

[0074] In summary, this application extracts low-resolution meteorological field features from the output of the AI ​​meteorological forecasting model through a multi-scale encoder. A physically constrained decoder introduces soft constraints such as energy conservation and mass continuity during the generation of a high-resolution complete initial field. A fusion layer integrates multi-source observation data to improve local accuracy. A dynamic residual correction module dynamically updates model parameters based on short-term forecast errors from regional numerical models. This application solves the variable mismatch and physical inconsistency problems between the output of the AI ​​meteorological forecasting model and the input of the regional numerical model, achieving seamless integration of the AI ​​meteorological forecasting model and the regional numerical model. Experiments show that after driving the WRF model, surface air temperature and wind speed are improved by 40% compared to the AI ​​meteorological forecasting model, significantly improving the accuracy and stability of refined meteorological forecasts.

[0075] Example 5 like Figure 6 The diagram shown is a schematic representation of the meteorological forecasting system of this application, which may include: The data acquisition module is used to acquire meteorological field data through the AI ​​meteorological forecasting model; The bridging processing module is used to process the meteorological field data using the aforementioned physical perception fusion bridging model to obtain a high-resolution initial field as input to the regional numerical model. The forecast output module is used to input high-resolution initial fields into the regional numerical model to obtain regional weather forecast results.

[0076] The meteorological forecasting system of this application can be implemented using software, hardware, or a combination of both. Specifically, in some embodiments, the data acquisition module, bridging processing module, and forecast output module can be implemented by a processor calling program instructions stored in memory, i.e., deployed as software programs on servers, workstations, cloud platforms, or other computing devices. In other embodiments, each module can also be implemented using dedicated hardware circuits, programmable logic devices, or dedicated computing units to improve data processing efficiency and system stability. In still other embodiments, a software and hardware collaborative approach can be used, for example, a general-purpose computing platform performs data scheduling and flow control, while GPUs, AI accelerators, or other dedicated hardware complete model inference, feature processing, and numerical calculation tasks. The system can be configured according to different business scenarios, computing power conditions, and deployment requirements.

[0077] Example 6 Embodiments of this application also provide a computer program product, including a computer program that, when executed by a processor, implements the steps of the methods described in various embodiments of this application; the computer program product further includes a non-volatile computer-readable storage medium storing the computer program, which, when executed by a processor, implements the steps of the weather forecasting methods described in various embodiments of this application.

[0078] The embodiments described above are merely preferred embodiments of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various improvements and substitutions without departing from the technical principles of this application, and these improvements and substitutions should also be considered within the scope of protection of this application. Therefore, the scope of protection of this patent application should be determined by the scope of the claims.

Claims

1. A physical perception fusion bridging model, characterized in that, The physical perception fusion bridging model is set between the AI ​​weather forecasting large model and the regional numerical model, including: A multi-scale encoder is used to extract multi-scale meteorological field features from the output of a large AI weather forecasting model. A physical constraint decoder is used to decode the meteorological field features and generate a high-resolution meteorological field that conforms to the physical conservation law. The observation fusion layer is used to fuse multi-source observation data and high-resolution meteorological fields to obtain a high-resolution initial field as input to the regional numerical model. The physical constraint decoder introduces a physical conservation penalty term during training; the physical conservation penalty term includes at least one of energy conservation, mass conservation, and momentum conservation.

2. The physical perception fusion bridging model according to claim 1, characterized in that, It also includes a dynamic residual correction module; The dynamic residual correction module is used to perform dynamic residual correction on the output of the physical perception fusion bridging model based on short-term forecast errors; the method for obtaining the short-term forecast errors includes: Using the high-resolution initial field obtained through the physical perception fusion bridging model as input, the short-term weather forecast results generated by the regional numerical model are obtained; The generated short-term weather forecast results are compared with multi-source observation data for the corresponding time period to obtain the short-term forecast error of the regional numerical model.

3. The physical perception fusion bridging model according to claim 2, characterized in that, In the dynamic residual correction module, the short-term forecast error of the regional numerical model is the forecast error of the previous 6 hours. When performing dynamic residual correction, the output of the physical perception fusion bridging model is dynamically corrected using learnable affine transformation parameters.

4. The physical perception fusion bridging model according to claim 1, characterized in that, The multi-scale encoder employs a hierarchical Transformer architecture or a convolutional neural network architecture.

5. The physical perception fusion bridging model according to claim 1, characterized in that, The physical conservation penalty terms include energy conservation terms, mass conservation terms, and momentum conservation terms; The formula for calculating the physical conservation penalty term includes: in, This is a physical conservation penalty term. For energy conservation terms, For the mass conservation term, For momentum conservation, for The weighting coefficients, for The weighting coefficients, for The weighting coefficients.

6. The physical perception fusion bridging model according to claim 1, characterized in that, The observation fusion layer employs variational assimilation or attention mechanisms for fusion.

7. A weather forecasting method, characterized in that, include: Meteorological field data is obtained through a large AI-powered weather forecasting model. The meteorological field data is processed using the physical perception fusion bridging model described in any one of claims 1 to 6 to obtain a high-resolution initial field as input to the regional numerical model. High-resolution initial fields are input into the regional numerical model to obtain regional weather forecast results.

8. The weather forecasting method according to claim 7, characterized in that, The high-resolution initial field has a spatial resolution of 0.01°×0.01°, and its vertical structure includes 37 pressure layers.

9. A weather forecasting system, characterized in that, include: The data acquisition module is used to acquire meteorological field data through the AI ​​meteorological forecasting model; The bridging processing module is used to process the meteorological field data using the physical perception fusion bridging model as described in any one of claims 1 to 6, to obtain a high-resolution initial field as input to the regional numerical model. The forecast output module is used to input high-resolution initial fields into the regional numerical model to obtain regional weather forecast results.

10. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the weather forecasting method as described in claim 7 or 8.

Images