A meteorological prediction model training and application method, system, terminal device and medium based on multi-scale feature fusion

Abstract

This invention discloses a method, system, terminal device, and medium for training and applying a meteorological forecasting model based on multi-scale feature fusion, relating to the field of meteorological forecasting technology. The method includes: acquiring raw meteorological data; preprocessing the raw meteorological data to obtain a global meteorological field; slicing the global meteorological field using at least two different scale slicing rules to obtain several slice sets; extracting features from each slice set to obtain meteorological features; fusing and decoding the meteorological features through a multi-scale fusion structure to obtain a multi-scale predicted meteorological field; calculating the loss based on the global meteorological field and the multi-scale predicted meteorological field and updating the parameters of the meteorological forecasting model. This invention, under limited training sample conditions, processes the meteorological field through multi-scale feature fusion, while simultaneously considering local detail representation and large-scale background modeling, effectively improving the model's convergence speed, prediction accuracy, and generalization ability.

Description

Technical Field

[0001] This invention relates to the field of meteorological forecasting technology, and in particular to a method, system, terminal equipment, and medium for training and applying meteorological forecasting models based on multi-scale feature fusion. Background Technology

[0002] Traditional numerical weather prediction supports meteorological disaster early warning and social governance. In recent years, data-driven weather prediction has become an important research direction in the field due to its efficient reasoning speed, and can achieve rapid prediction by learning the statistical laws of meteorological data.

[0003] However, existing data-driven weather forecasting models face several technical challenges when predicting high-resolution meteorological fields. First, the increased spatial resolution leads to a surge in the number of meteorological grid points, significantly increasing data storage and training computation costs, making the models reliant on massive samples and long training periods. Furthermore, the models are highly dependent on the size of the training sample set; with limited samples, they cannot simultaneously capture both local fine-structure information and large-scale background circulation features, easily resulting in insufficient feature extraction, decreased prediction accuracy, and weak generalization ability.

[0004] Therefore, there is an urgent need for a method that can perform joint modeling of multi-scale meteorological features with limited training samples to fill the gaps in existing technologies. Summary of the Invention

[0005] The technical problem this invention aims to solve is that, in the field of data-driven weather forecasting, high-resolution weather field prediction suffers from high model training costs, high dependence on massive samples, and an inability to simultaneously represent local details and model large-scale background circulation under limited sample conditions, resulting in insufficient model convergence efficiency, prediction accuracy, and generalization ability. Therefore, an effective solution is urgently needed to address these technical problems.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: In a first aspect, the present invention provides a method for training a meteorological forecasting model based on multi-scale feature fusion, the method comprising: Obtain raw meteorological data, preprocess the raw meteorological data, and obtain the global meteorological field; The global meteorological field is sliced ​​using at least two different scale slicing rules to obtain several slice block sets. Each slice block set corresponds to a slicing rule of one scale. All slice blocks in each slice block set have the same scale, and any slice block set completely constitutes the global meteorological field. Feature extraction is performed on each slice set to obtain the meteorological features corresponding to each slice set. The meteorological features are then fused and decoded through a multi-scale fusion structure to obtain a multi-scale predicted meteorological field. The loss is calculated and the parameters of the weather forecast model are updated based on the global meteorological field and the multi-scale predicted meteorological field.

[0007] In one implementation, the step of acquiring raw meteorological data and preprocessing the raw meteorological data to obtain a global meteorological field includes: Obtain raw meteorological data and convert the raw meteorological data into a format that is uniform; By interpolating to complete the data, missing values ​​in the uniformly formatted meteorological data are filled in to obtain complete meteorological data. The complete meteorological data is normalized, time-aligned, and gridded to obtain a standard global meteorological field.

[0008] In one implementation, the step of extracting features from each set of slice blocks to obtain meteorological features corresponding to each set of slice blocks, and then fusing and decoding the meteorological features through a multi-scale fusion structure to obtain a multi-scale predicted meteorological field, includes: Meteorological feature extraction is performed independently on each slice set to obtain the meteorological features corresponding to the slice set; Using the same shared core feature operator, global feature modeling and parallel computation are performed on the meteorological features corresponding to each slice set to obtain the global meteorological features corresponding to each slice set. Decode the global meteorological features corresponding to each slice set independently to obtain the single-scale predicted meteorological field corresponding to each slice set at the target time. The single-scale predicted meteorological fields corresponding to all slice blocks are weighted and fused to obtain the multi-scale predicted meteorological field at the target time.

[0009] In one implementation, the step of calculating the loss and updating the parameters of the weather forecast model based on the global meteorological field and the multi-scale predicted meteorological field includes: Obtain the true value label of the global meteorological field at the target time; Calculate the loss between the true value label and the multi-scale predicted meteorological field to obtain the loss value; Based on the loss value, update the independent meteorological feature extraction branch parameters at each scale, the independent decoding branch parameters at each scale, the weighted fusion learnable weight parameters, and the shared core feature operator parameters.

[0010] In one implementation, the step of extracting features from each set of slice blocks to obtain meteorological features corresponding to each set of slice blocks, and then fusing and decoding the meteorological features through a multi-scale fusion structure to obtain a multi-scale predicted meteorological field, includes: Meteorological feature extraction is performed independently on each slice set to obtain the meteorological features corresponding to each slice set; By using the expert feature branches corresponding to each slice set, global feature modeling and parallel computation are performed on the meteorological features corresponding to each slice set to obtain the global meteorological features corresponding to each slice set. Among them, each slice set at each scale corresponds to a unique expert feature branch. The global meteorological features corresponding to each slice block set are weighted and fused to obtain the fused global meteorological features; A unified decoding process is performed on the fused global meteorological features to obtain a multi-scale predicted meteorological field.

[0011] In one implementation, the step of calculating the loss and updating the parameters of the weather forecast model based on the global meteorological field and the multi-scale predicted meteorological field includes: Obtain the real meteorological labels for the target time corresponding to the global meteorological field; Calculate the loss between the real weather label and the multi-scale predicted weather field to obtain the loss value; Based on the loss value, update the independent expert feature branch parameters for each scale, as well as the learnable weight parameters for weighted fusion.

[0012] Secondly, embodiments of the present invention also provide a method for applying a meteorological forecasting model based on multi-scale feature fusion, applied to a meteorological forecasting model obtained by the training method for a meteorological forecasting model based on multi-scale feature fusion as described in any of the above schemes, the application method comprising: Raw meteorological data is acquired and preprocessed to obtain the global meteorological field; The global meteorological field is input into the trained meteorological forecasting model to obtain the multi-scale predicted meteorological field at the target time.

[0013] Thirdly, embodiments of the present invention also provide a meteorological forecasting model system based on multi-scale feature fusion, the system comprising: The preprocessing module is used to acquire raw meteorological data, preprocess the raw meteorological data, and obtain the global meteorological field. The slicing module is used to slice the global meteorological field using at least two different scale slicing rules to obtain several slice block sets. Each slice block set corresponds to a slicing rule of one scale. All slice blocks in each slice block set have the same scale, and any slice block set completely constitutes the global meteorological field. The prediction module is used to extract features from each set of slice blocks to obtain the meteorological features corresponding to each set of slice blocks, and to perform fusion and decoding of the meteorological features through a multi-scale fusion structure to obtain a multi-scale predicted meteorological field. The training module is used to calculate the loss and update the parameters of the weather forecast model based on the global meteorological field and the multi-scale predicted meteorological field.

[0014] Fourthly, embodiments of the present invention also provide a terminal device, the terminal device including a memory, a processor, and a meteorological forecast model training program based on multi-scale feature fusion stored in the memory and executable on the processor. When the processor executes the meteorological forecast model training program based on multi-scale feature fusion, it implements the steps of the meteorological forecast model training method based on multi-scale feature fusion as described in any of the above schemes.

[0015] Fifthly, embodiments of the present invention also provide a computer-readable storage medium storing a meteorological forecasting model training program based on multi-scale feature fusion. When the meteorological forecasting model training program based on multi-scale feature fusion is executed by a processor, it implements the steps of the meteorological forecasting model training method based on multi-scale feature fusion as described in any of the above schemes.

[0016] Beneficial Effects: This invention discloses a method, system, terminal device, and medium for training and applying a meteorological forecasting model based on multi-scale feature fusion, relating to the field of meteorological forecasting technology. The method first acquires raw meteorological data and preprocesses it to obtain a global meteorological field. Then, using at least two different scale slicing rules, the global meteorological field is sliced ​​to obtain several slice set sets. Each slice set corresponds to a slicing rule of one scale, all slices in each slice set have the same scale, and any slice set completely constitutes the global meteorological field. Subsequently, features are extracted from each slice set to obtain the meteorological features corresponding to each slice set. These features are then fused and decoded using a multi-scale fusion structure to obtain a multi-scale predicted meteorological field. Finally, the loss is calculated based on the global meteorological field and the multi-scale predicted meteorological field, and the parameters of the meteorological forecasting model are updated.

[0017] This invention achieves multi-scale division of the global meteorological field through multi-scale slicing rules, which can fully extract meteorological feature information at different scales. With the help of multi-scale fusion structure, feature fusion and decoding are completed. Under limited training sample conditions, the model's ability to learn and represent multi-dimensional features of the meteorological field can be enhanced, the training process of the meteorological forecast model can be optimized, the output accuracy of multi-scale meteorological field prediction can be improved, the model prediction results can be more consistent with the real meteorological field, and the accuracy and stability of meteorological forecasts can be effectively improved. Attached Figure Description

[0018] Figure 1 The flowchart illustrates a specific implementation method for training a meteorological forecast model based on multi-scale feature fusion, as provided in this embodiment of the invention.

[0019] Figure 2This is a schematic diagram of the basic model encoding and decoding process for the meteorological forecast model training method based on multi-scale feature fusion provided in this embodiment of the invention.

[0020] Figure 3 This is a schematic diagram of the spatial multi-scale scheme fusion process of the meteorological forecast model training method based on multi-scale feature fusion provided in the embodiments of the present invention.

[0021] Figure 4 This is a schematic diagram of the expert multi-scale scheme fusion process for the meteorological forecast model training method based on multi-scale feature fusion provided in the embodiments of the present invention.

[0022] Figure 5 A comparison chart of the loss in training iterations of the meteorological forecast model training method based on multi-scale feature fusion provided in the embodiments of the present invention.

[0023] Figure 6 The RMSE variation diagram of key variables in the meteorological forecast model training method based on multi-scale feature fusion provided in the embodiments of the present invention.

[0024] Figure 7 This is a schematic diagram of the principle of the meteorological forecast model training device based on multi-scale feature fusion provided in the embodiments of the present invention.

[0025] Figure 8 This is a block diagram illustrating the internal structure of the terminal device provided in an embodiment of the present invention. Detailed Implementation

[0026] To make the objectives, technical solutions, and effects of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0027] The flowchart shown in the attached diagram is for illustrative purposes only and does not necessarily include all content, operations, or steps, nor does it require execution in the described order. For example, some operations or steps can be broken down, combined, or partially merged, so the actual execution order may change depending on the actual situation.

[0028] It should be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0029] It should be understood that, in order to clearly describe the technical solutions of the embodiments of the present invention, the terms "first" and "second" are used in the embodiments of the present invention to distinguish identical or similar items with essentially the same function and effect. For example, "first control information" and "second control information" are only used to distinguish different control information and do not limit their order.

[0030] Those skilled in the art will understand that the words "first" and "second" do not limit the quantity or the order of execution, and that the words "first" and "second" do not necessarily imply that they are different.

[0031] It should also be understood that the term “and / or” as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0032] Traditional numerical weather prediction technology has evolved from manual calculation to large-scale numerical simulation supported by modern high-performance computing since the 20th century. With the improvement of computing power, the refinement of physical process characterization, and advancements in observation and data assimilation technologies, forecast accuracy has significantly improved, making it an important technical means to support disaster early warning, production scheduling, and social governance. In recent years, deep learning technology has developed rapidly, and data-driven weather forecasting methods have gradually become a core research direction in this field. These methods achieve rapid prediction of atmospheric conditions by learning statistical regularities from historical reanalysis data, observational data, or climate simulation data. They are computationally efficient and cost-effective in the inference stage, and have significant application potential in scenarios such as ensemble forecasting, probabilistic forecasting, sub-seasonal to seasonal scale forecasting, and extreme weather event prediction.

[0033] However, existing data-driven weather forecasting models face numerous technical challenges when predicting high-resolution meteorological fields. First, the increased spatial resolution leads to a significant increase in the number of meteorological grid points. The superposition of multiple meteorological variables, vertical layers, and temporal dimensions results in a dramatic expansion of training data, placing extremely high demands on data reading, storage management, and input / output throughput, severely limiting model training efficiency. Second, high-resolution input leads to a surge in model parameter computation and training costs, requiring massive training samples and long training periods. This results in a strong dependence on high-performance hardware resources, high trial-and-error costs in model development, and low iteration efficiency. Third, the high dependence on large-scale training samples makes it difficult to simultaneously capture local fine-structure information and large-scale background circulation features under conditions of few samples or limited training data. This can easily lead to insufficient feature extraction, decreased generalization ability, and reduced prediction accuracy.

[0034] Currently, existing technologies cannot achieve both local detail representation and large-scale background modeling under limited training sample conditions. They are also difficult to effectively improve model convergence efficiency, prediction accuracy and generalization performance. There is a lack of multi-scale meteorological feature fusion modeling schemes adapted to scenarios with few samples.

[0035] Therefore, the purpose of this invention is to overcome the problems of long training cycles, high computational costs, and high dependence on massive samples in existing data-driven weather forecasting models for high-resolution forecasting tasks, and to provide a multi-scale forecasting algorithm based on fewer samples. This algorithm, through the construction of a multi-scale patch representation mechanism, a cross-scale feature fusion mechanism, and an expert branch collaboration mechanism, enables the model to effectively extract key meteorological features at different spatial scales under limited training sample conditions, thereby improving the model's convergence efficiency, prediction accuracy, and generalization performance. Addressing the issues of high computational cost and long training cycles in high-resolution meteorological field modeling, a multi-scale modeling mechanism improves feature utilization efficiency, thereby reducing the model's dependence on ultra-large-scale data and long training cycles. Furthermore, addressing the problem that models struggle to simultaneously consider local details and large-scale background information under fewer sample conditions, this invention proposes to improve the model's ability to represent complex atmospheric dynamic processes through joint modeling and adaptive fusion of features at different scales, thus maintaining good prediction performance while reducing the amount of training data.

[0036] This embodiment provides a method for training a meteorological forecasting model based on multi-scale feature fusion, such as... Figure 1 As shown, the specific steps include the following: Step S100: Obtain raw meteorological data, preprocess the raw meteorological data to obtain the global meteorological field.

[0037] In this embodiment, the raw meteorological data refers to the initial observational and reanalysis data used for training meteorological forecasting models. It serves as the foundational data for atmospheric physical state information and can specifically include global or regional meteorological station observation data, atmospheric reanalysis gridded data, and meteorological data retrieved from satellite remote sensing. Preprocessing is a standardized processing procedure performed on the raw meteorological data to eliminate differences between multi-source data, correct data defects, and standardize data format. The global meteorological field is a standardized, gridded dataset covering the target spatial region and containing various meteorological physical quantities.

[0038] Raw meteorological data often suffers from problems such as diverse sources, inconsistent formats, and uneven distribution of missing measurement points. Preprocessing can transform the messy initial data into standardized data, avoiding interference from data defects on feature extraction and model training, and improving the stability and reliability of the overall process.

[0039] In one implementation, the steps of acquiring raw meteorological data and preprocessing the raw meteorological data to obtain a global meteorological field specifically include the following: Step S110: Obtain raw meteorological data and convert the raw meteorological data into a format that is uniform; Step S120: By interpolation, missing values ​​in the uniformly formatted meteorological data are filled in to obtain complete meteorological data; Step S130: Normalize, time-align, and grid the complete meteorological data to obtain a standard global meteorological field.

[0040] In this embodiment, format unification is the process of converting raw meteorological data from different sources, with different storage formats, and different spatial coordinate definitions into a unified standard format. This process unifies the data's storage format, variable units, spatial projection methods, and numerical representations, eliminating format differences between multi-source data. Interpolation completion is a repair operation performed on missing measurement points in the format-unified meteorological data. Spatial interpolation algorithms are used to estimate and fill in missing content based on the spatial distribution patterns of surrounding valid data, ensuring the spatial continuity and completeness of the meteorological data. Specific interpolation methods can include Kriging interpolation, inverse distance weighted interpolation, etc. Normalization is the process of mapping the numerical range of complete meteorological data to the model adaptation interval, reducing the interference of differences in the magnitude of different physical quantities on model training, improving the stability of the gradient descent process, and accelerating model convergence. Time alignment is the process of unifying the timestamps of data from different sources to the same time base, ensuring that all data correspond to the same forecast time and time interval, avoiding feature distortion caused by time misalignment. Gridding is the process of mapping the processed meteorological data to a regular spatial grid, forming standardized grid data covering the target area, i.e., the global meteorological field.

[0041] After the above preprocessing steps of format conversion, interpolation completion, normalization, time alignment, and gridding, the original meteorological data is transformed into a standardized, complete, and unified global meteorological field as input for subsequent model training.

[0042] Step S200: Using at least two different scale slicing rules, slice the global meteorological field to obtain several slice block sets. Each slice block set corresponds to a slicing rule of one scale. All slice blocks in each slice block set have the same scale, and any slice block set completely constitutes the global meteorological field.

[0043] In this embodiment, the slicing rules at different scales refer to multiple spatial segmentation standards set for the same global meteorological field. Specifically, the global meteorological field is sliced ​​using different patch sizes to divide the continuous global meteorological field into local data blocks with different spatial coverage. These can be spatial grid division standards such as 8×8, 16×16, and 24×24. Different patch sizes correspond to different spatial perceptions; smaller-scale patches are better at preserving fine local structural information, while larger-scale patches are better at extracting large-scale background circulation and long-range dependency features. A patch set is a collection of all local data blocks obtained by the same slicing rule, with all data blocks within the same set maintaining a consistent spatial size and coverage. Scale is an indicator used to characterize the spatial coverage and feature perception level of a patch. Smaller scales correspond to fine local space and can capture local meteorological disturbance information, while larger scales correspond to broad background space and can extract large-scale circulation features. By employing at least two different scale slicing rules to divide the global meteorological field, the model can simultaneously perceive meteorological information at different spatial levels, solving the problem that a single scale cannot take into account both local details and large-scale background, and improving the model's ability to learn complex atmospheric motion patterns.

[0044] Step S300: Extract features from each slice set to obtain the meteorological features corresponding to each slice set, and perform fusion and decoding on the meteorological features through a multi-scale fusion structure to obtain a multi-scale predicted meteorological field.

[0045] In this embodiment, meteorological features are high-dimensional representations of information obtained from slices through feature extraction. These features are used to condense the distribution, spatial correlation, and variation patterns of meteorological physical quantities within the slices, reflecting the model's learning of atmospheric motion patterns. The multi-scale fusion structure is a pre-defined architecture in this embodiment used to integrate meteorological features at different scales. It enables the correlated modeling and collaborative utilization of multi-scale information, enhancing the model's feature representation capabilities. Fusion is the operation of associating and integrating independent meteorological features at different scales, eliminating information gaps between scales and achieving complementary and synergistic effects of multi-scale information. Decoding is the operation of restoring high-dimensional abstract meteorological features into physically meaningful meteorological field data, outputting prediction results in the same format as the original global meteorological field. The multi-scale predicted meteorological field is the meteorological prediction result for the target time obtained after fusing multi-scale feature information, simultaneously expressing both small-scale details and large-scale background.

[0046] The process of extracting, decoding, and predicting meteorological features forms the basic framework for calculating the meteorological field prediction results in this embodiment. The choice of the basic framework is not limited to a specific network structure. In particular, various data-driven meteorological forecasting models can be selected, such as convolutional neural networks, visual transformers, neural operators, and hybrid encoder architectures, including mainstream meteorological forecasting models such as FourCastNet, Pangu-Weather, and FuXi.

[0047] In this embodiment, FourCastNet is used as the basic model framework, and its basic encoding and decoding processing flow is as follows: Figure 2 As shown. A token is a high-dimensional feature label obtained after feature embedding and dimensionality mapping, serving as the input to the feature modeling module. The Adaptive Fourier Neural Operator (AFNO) is an efficient feature computation operator suitable for global correlation modeling of meteorological fields. It achieves rapid capture of long-range dependencies through Fourier transform, exhibiting strong parallel computing capabilities and good global modeling performance.

[0048] The standard processing flow of the basic model framework first divides the input meteorological field into several local patches of a fixed size. Then, convolutional embedding operations are used to extract features and map dimensions for each patch, projecting the original local meteorological information into a high-dimensional feature space to obtain the corresponding word representation. Next, the word representations are input into an adaptive Fourier neural operator layer for feature extraction. Finally, the extracted features are decoded to reconstruct the target meteorological field.

[0049] Building upon the basic model framework, this embodiment improves upon the aforementioned basic process by generating multiple sets of slice blocks for the same spatiotemporal meteorological field using at least two different scale slicing rules. Subsequently, convolutional embedding operations are used to extract features and map dimensions for each slice block, projecting the original local meteorological information into a high-dimensional feature space to generate multiple sets of lexical representations to extract atmospheric motion features at different scales. Furthermore, a pre-defined or learnable multi-scale fusion structure is used to collaboratively model multi-scale information, thereby improving the model's ability to represent complex weather systems and its small-sample learning capability.

[0050] Specifically, for the input meteorological reanalysis data or observation data, the first step is to extract features and map dimensions of all slices in the slice set at each scale through convolutional embedding operations, generating a series of lexical representations composed of slices at different scales, i.e. meteorological features. Then, multi-scale joint modeling is achieved through a pre-set fusion strategy, taking into account both local details and large-scale background, and improving the completeness of feature expression.

[0051] Fusion strategy is a method and steps for fusing the feature representations of slices at different scales. It can be a spatial multi-scale scheme (MultiP) fusion strategy or an expert multi-scale scheme (MultiE) fusion strategy.

[0052] The spatial multi-scale scheme involves learning features at each scale independently first, and then fusing them at the output. This scheme helps maintain the independent representational ability of each scale branch and reduces information interference that may be caused by premature fusion.

[0053] The expert multi-scale scheme specifically sets up independent coding branches for different scale slices. Each branch acts as an expert network at a different scale, independently learning the meteorological feature representation at the corresponding scale. Each expert branch and the fusion module can gradually learn the contribution of information at different scales to the forecasting task, achieving better cross-scale feature synergy.

[0054] By extracting features and performing multi-scale fusion decoding, the value of meteorological data at different scales is explored, improving the model's accuracy in representing and predicting meteorological fields, and making the prediction results more consistent with the actual atmospheric motion state.

[0055] Step S400: Calculate the loss based on the global meteorological field and the multi-scale predicted meteorological field, and update the parameters of the meteorological forecast model.

[0056] In this embodiment, the loss is a numerical indicator used to quantify the degree of deviation between the multi-scale predicted meteorological field and the original global meteorological field. It is used to supervise the optimization of model parameters and can be calculated based on root mean square error, mean square error, etc. Parameter update is an optimization process that iteratively adjusts the trainable parameters within the meteorological forecast model based on the loss value. Adaptive optimization of parameters can be achieved through the backpropagation algorithm. Model parameters are learnable weights and biases that support meteorological feature extraction, fusion, and decoding operations.

[0057] The model calculates the loss and updates the parameters based on the global meteorological field and the multi-scale predicted meteorological field, continuously reducing the deviation between the prediction results and the actual meteorological field, allowing the model to gradually converge to the optimal state. Under the condition of limited training samples, the model's convergence efficiency and prediction accuracy are improved, and the model's generalization ability in actual forecast scenarios is enhanced.

[0058] In one implementation, the step of extracting features from each set of slice blocks to obtain meteorological features corresponding to each set of slice blocks, and then fusing and decoding the meteorological features through a multi-scale fusion structure to obtain a multi-scale predicted meteorological field, specifically includes the following steps: Step S310a: Perform meteorological feature extraction independently on each slice set to obtain the meteorological features corresponding to the slice set; Step S320a: Using the same shared core feature operator, perform global feature modeling and parallel computation on the meteorological features corresponding to each slice block set to obtain the global meteorological features corresponding to each slice block set; Step S330a: Independently decode the global meteorological features corresponding to each slice set to obtain the single-scale predicted meteorological field corresponding to each slice set at the target time; Step S340a: Weighted fusion of the single-scale predicted meteorological fields corresponding to all slice blocks to obtain the multi-scale predicted meteorological field at the target time.

[0059] In this embodiment, Figure 3 This paper demonstrates a fusion strategy for spatial multi-scale schemes, which retains the independent representation capabilities of each scale branch while sharing the global feature modeling capability, thus achieving a balance between computational efficiency and feature representation capability.

[0060] In the spatial multi-scale scheme, meteorological feature extraction is performed independently for each set of slices. Convolutional embedding and dimensionality mapping, consistent with the base model, are performed separately for each scale slice, transforming the local slices at each scale into corresponding high-dimensional lexical representations, i.e., meteorological features. The feature extraction processes at different scales are independent of each other and do not share any computational parameters, ensuring that each scale can autonomously learn the specific meteorological features corresponding to its spatial level.

[0061] Meteorological features at all scales are input into the same shared core feature operator to perform global feature modeling and parallel computation. This shared core feature operator is implemented using an adaptive Fourier neural operator (AFNO), which completes the modeling of global spatial correlation and channel correlation based on Fourier transform. The mathematical expression is as follows: ,in This represents the Discrete Fourier Transform, used to convert spatial domain features into frequency domain features. This represents the inverse discrete Fourier transform, used to restore frequency domain features to spatial domain features. This represents a learnable frequency-domain convolution kernel used to extract meteorological features of different frequency components. This represents the multi-scale meteorological characteristics of the input.

[0062] Figure 3 During the calculation process, This indicates the current forecast for the previous moment, on a scale of... The set of slice block data, This indicates the current forecast for the previous moment, with a scale of [missing information]. The set of slice block data, The first term obtained after feature embedding is represented as... Line number The column is represented by a lexical representation. Meteorological features at different scales are computed in parallel within the same shared operator. All parameters of the operator are fully shared across scales, capturing cross-scale atmospheric motion correlations through unified global modeling capabilities, while reducing parameter increments and computational costs associated with multi-scale modeling. After processing by the shared operator, global meteorological features corresponding to each slice set are obtained. These features retain the local characteristics of the corresponding scale and incorporate long-range dependencies in the global space.

[0063] Decoding is performed independently for the global meteorological features corresponding to each slice set. Each scale has its own dedicated decoding branch, and the parameters of each decoding branch are independent and not shared with other scales. The decoding branch gradually restores the high-dimensional abstract global meteorological features into physically meaningful grid data, obtaining the single-scale predicted meteorological field corresponding to each slice set at the target time. All single-scale predicted meteorological fields are weighted and fused, and the prediction results of each scale are integrated by pixel-level weighted averaging. The final predicted value of each spatial grid point is obtained by weighted summation of the predicted values ​​of all scales at that grid point, and the weighting coefficients are learnable parameters automatically learned during model training.

[0064] By sharing core feature operators, global modeling capabilities are reused. Meanwhile, independent feature extraction and decoding branches preserve the independent representation capabilities of each scale, avoiding the loss of detailed information and feature interference caused by premature fusion. Under limited training sample conditions, this effectively improves the model's efficiency in utilizing meteorological features at different scales, accelerates model convergence, and simultaneously ensures the prediction accuracy of both local fine structure and large-scale background circulation.

[0065] In one implementation, the step of calculating the loss and updating the parameters of the weather forecast model based on the global meteorological field and the multi-scale predicted meteorological field specifically includes the following steps: Step S410a: Obtain the true value label of the global meteorological field at the target time; Step S420a: Calculate the loss between the true value label and the multi-scale predicted meteorological field to obtain the loss value; Step S430a: Update the independent meteorological feature extraction branch parameters at each scale, the independent decoding branch parameters at each scale, the weighted fusion learnable weight parameters, and the shared core feature operator parameters based on the loss value.

[0066] In this embodiment, the true value label of the global meteorological field at the target time is the actual observed meteorological field data corresponding to the prediction time, which serves as a supervisory signal for model training and parameter optimization. The loss between the true value label and the multi-scale predicted meteorological field is calculated, and the degree of deviation between the prediction result and the actual data is quantified by a preset loss function to obtain the loss value used to guide parameter updates.

[0067] Parameter updates are performed based on the loss value, using the backpropagation algorithm. The algorithm iteratively adjusts all trainable parameters of the model based on the gradient information of the loss value, causing the model to converge in the direction of reducing prediction error. The updated parameter range fully covers all trainable components of the spatial multi-scale scheme, including parameters of the feature extraction branch at each scale, parameters of the decoding branch at each scale, learnable weight parameters of the weighted fusion of single-scale predicted meteorological fields, and all parameters of the unique shared core feature operator.

[0068] The parameters of each component are updated synchronously, enabling the global modeling capability of the shared core feature operator to be synergistically adapted with the feature extraction and decoding capabilities of independent branches at each scale. Simultaneously, the fusion weights adaptively learn the contribution of prediction results at different scales to different meteorological scenarios. This joint update approach accelerates model convergence under limited training sample conditions, improving the overall accuracy of multi-scale meteorological field predictions and the model's ability to represent complex weather systems.

[0069] In one implementation, the step of extracting features from each set of slice blocks to obtain meteorological features corresponding to each set of slice blocks, and then fusing and decoding the meteorological features through a multi-scale fusion structure to obtain a multi-scale predicted meteorological field, specifically includes the following steps: Step S310b: Perform meteorological feature extraction independently on each slice set to obtain the meteorological features corresponding to each slice set; Step S320b: Through the expert feature branches corresponding to each slice set, perform global feature modeling and parallel computation on the meteorological features corresponding to each slice set to obtain the global meteorological features corresponding to each slice set. Among them, each slice set at each scale corresponds to a unique expert feature branch. Step S330b: Perform weighted fusion processing on the global meteorological features corresponding to each slice block set to obtain fused global meteorological features; Step S340b: Perform unified decoding on the fused global meteorological features to obtain a multi-scale predicted meteorological field.

[0070] In this embodiment, Figure 4 This paper demonstrates a fusion strategy for expert multi-scale schemes, which achieves specialized learning of features at each scale through independent expert branches and combines adaptive fusion at the feature level to further improve the model's generalization ability in scenarios with few samples.

[0071] Meteorological feature extraction is performed independently for each slice set, consistent with the spatial multi-scale scheme. Convolutional embedding and dimension mapping operations are performed separately for each scale slice to generate corresponding high-dimensional word representations, i.e. meteorological features. The input data for the two fusion strategies are the same.

[0072] Global feature modeling and parallel computation are performed on the meteorological features corresponding to each set of slices using expert feature branches. Each scale of slice set corresponds to a unique expert feature branch, and all expert branches use adaptive Fourier neural operators as the core computational component. Figure 4 middle This indicates that the number of expert branches is equal to the number of scales. Each expert branch is completely independent, without sharing any parameters, calculation results, or feature information. They focus on learning the specific meteorological features at their respective scales, avoiding mutual interference between features at different scales.

[0073] After processing by various expert branches, the global meteorological features corresponding to each slice set are obtained. These features reflect the atmospheric motion patterns at the corresponding scale and possess scale specificity. The global meteorological features corresponding to each slice set are then weighted and fused. Learnable weights are assigned to the global meteorological features at each scale, and the integration of multi-scale features is achieved through adaptive adjustment of the weights, resulting in fused global meteorological features. The learnable weights can be scalar weights, vector weights, spatially relevant weights, dynamically adjusted gating weights based on samples, or adaptive weights based on attention mechanisms, automatically optimizing the contribution ratio of each scale according to the feature distribution of different meteorological scenarios.

[0074] As an alternative implementation, one can first independently decode the global meteorological features output by each expert branch to obtain the single-scale predicted meteorological field corresponding to each scale, and then perform weighted fusion of all single-scale predicted meteorological fields to obtain the final multi-scale predicted meteorological field. This implementation can also achieve the collaborative utilization of multi-scale information and is suitable for application scenarios that require high independence of prediction results at each scale.

[0075] A unified decoding process is performed on the fused global meteorological features. A single decoding branch is used to restore the integrated high-dimensional features into physically meaningful grid data, resulting in a multi-scale predicted meteorological field. The unified decoding architecture simplifies the computational complexity of the decoding process while ensuring the continuity and consistency of the prediction results in the spatial dimension. Independent expert branches are used to mine features at each scale, and an adaptive fusion mechanism is combined to integrate cross-scale information. This improves the model's feature learning ability and generalization performance under limited training sample conditions, and suppresses the accumulation of errors in medium- and long-term forecasts.

[0076] In one implementation, the step of calculating the loss and updating the parameters of the weather forecast model based on the global meteorological field and the multi-scale predicted meteorological field specifically includes the following steps: Step S410b: Obtain the real meteorological label corresponding to the target time in the global meteorological field; Step S420b: Calculate the loss between the real weather label and the multi-scale predicted weather field to obtain the loss value; Step S430b: Based on the loss value, update the independent expert feature branch parameters for each scale, as well as the learnable weight parameters for weighted fusion.

[0077] In this embodiment, the real meteorological labels corresponding to the target time in the global meteorological field have the same source and format as the real value labels in the spatial multi-scale scheme, providing a unified supervised training signal for the two fusion strategies, and ensuring the comparability of the optimization objectives of different schemes. The loss between the real meteorological labels and the multi-scale predicted meteorological field is calculated using the same loss calculation logic as the spatial multi-scale scheme to obtain a quantified bias loss value, ensuring the consistency and comparability of the loss values.

[0078] Parameter updates are performed based on loss values, adjusting only the trainable parameters corresponding to the expert multi-scale scheme. These include independent expert feature branch parameters for each scale, as well as learnable weight parameters for weighted fusion of global meteorological features. The core feature operators of the expert multi-scale scheme are not shared, and the parameter update process maintains the independence of each expert branch, avoiding mutual interference between the feature learning of different branches.

[0079] The parameters of each expert branch and the fusion weights are updated synchronously, enabling the feature extraction capabilities of each expert branch and the weight allocation capabilities of the fusion module to be optimized in synergy. This allows the fusion weights to gradually learn the importance of information at different scales in various forecasting tasks, ensuring the professional learning effect of each expert branch. At the same time, it enables adaptive integration of cross-scale features, further improving the convergence efficiency and prediction stability of the model in scenarios with few samples.

[0080] This invention provides a method for applying a meteorological forecasting model based on multi-scale feature fusion, which is applied to the meteorological forecasting model trained in the above embodiments, including: Step S510: Obtain raw meteorological data and perform preprocessing to obtain the global meteorological field; Step S520: Input the global meteorological field into the trained meteorological forecast model to obtain the multi-scale predicted meteorological field at the target time.

[0081] In this embodiment, raw meteorological data is acquired and preprocessed using the same processing flow as the model training phase. The real-time observed raw meteorological data is converted into a standard global meteorological field, which is then input into the trained meteorological forecasting model. The model automatically performs multi-scale slicing, feature extraction, multi-scale fusion, and decoding operations to quickly complete the meteorological forecasting process. The model outputs a multi-scale predicted meteorological field at the target time, possessing both local fine structure and large-scale background circulation features, suitable for meteorological forecasting operations.

[0082] By using a well-trained multi-scale fusion model, high-precision prediction results can be quickly output in real-world forecasting scenarios, adapting to the operational needs of high-resolution weather forecasting. At the same time, it maintains stable forecasting performance under limited data conditions, combining practicality and robustness.

[0083] In summary, this embodiment introduces multi-scale slice representation and cross-scale fusion mechanisms, enabling the model to simultaneously extract local fine structure information and large-scale background circulation information under limited training sample conditions, thereby improving feature learning efficiency. Compared to single-scale modeling methods, this invention effectively improves the model's ability to represent complex weather systems, increases training convergence speed, and reduces prediction errors. Compared to existing single-scale slice fast data-driven weather forecasting methods, this embodiment introduces a multi-scale representation mechanism, enabling the model to simultaneously learn local detailed features and large-scale background features. It also designs multiple cross-scale fusion methods to improve the synergistic utilization efficiency of information at different scales. Furthermore, it proposes an expert multi-scale structure, achieving adaptive cross-scale feature integration through a multi-branch expert network and learnable fusion weights. This method is suitable for training scenarios with few samples, still improving the model's convergence speed, prediction accuracy, and generalization ability under limited training sample conditions. It is also compatible with the main framework of existing weather forecasting models, requiring no complete reconstruction of the original model, and has good engineering feasibility and deployment convenience.

[0084] Specifically, FourCastNet was used as the baseline model. The training data consisted of 0.25-degree resolution reanalysis data from 1996 to 1998, and the validation data consisted of data from 1999 at the same resolution. The batch size was set to 64. The comparative experiment setup included two single-scale baseline schemes, namely P8 (8×8) and P24 (24×24), and two multi-scale schemes, namely the spatial multi-scale scheme (8×8 + 24×24) and the expert multi-scale scheme (8×8 + 24×24).

[0085] Figure 5 The experimental results were presented. Figure 5 (a) in the figure shows the comparison of training loss. Figure 5 (b) shows the comparison of validation losses. In the comparison of training and validation losses, P8 outperforms P24 overall in the single-scale scheme, while the spatial multi-scale scheme outperforms P24, indicating that multi-scale modeling has a significant advantage over the larger single-scale patch scheme. Furthermore, the expert multi-scale scheme exhibits faster convergence speed and lower convergence lower bound in both the training and validation phases, and its overall performance is superior to other comparative schemes.

[0086] Furthermore, Figure 6The RMSE (Root Mean Square Error) of key meteorological variables, including U10 (zonal wind speed at 10 meters altitude), V10 (meridian wind speed at 10 meters altitude), T2M (temperature at 2 meters altitude), T850 (temperature at 850 hPa), and Z500 (geopotential height at 500 hPa), as a function of forecast lead time is presented after 150 training steps. Figure 6 (a) shows U10, Figure 6 (b) shows V10, Figure 6 (c) in the middle shows T2M, Figure 6 (d) in the image shows the Z500. Figure 6 (e) shows T850. The results show that the single-scale scheme has a larger overall error, and the error accumulation is more pronounced with increasing forecast time. The multi-scale scheme, on the other hand, can significantly reduce RMSE and effectively slow down the error growth trend. Among them, the expert multi-scale scheme achieves the lowest RMSE across all variables and forecast periods, and has the smallest error growth slope, demonstrating better stability and robustness. The spatial multi-scale scheme is second best. The single-scale 24×24 scheme performs worse than the single-scale 8×8 scheme in mid-to-late-stage forecasts.

[0087] like Figure 7 As shown in the figure, this embodiment of the invention provides a meteorological forecast model training system based on multi-scale feature fusion. The system includes: a preprocessing module 10, a slicing module 20, a prediction module 30, and a training module 40.

[0088] Specifically, the preprocessing module 10 is used to acquire raw meteorological data, preprocess the raw meteorological data to obtain a global meteorological field; the slicing module 20 is used to slice the global meteorological field using at least two different scale slicing rules to obtain several slice block sets, wherein each slice block set corresponds to a slicing rule of one scale, all slice blocks in each slice block set have the same scale, and any slice block set completely constitutes the global meteorological field; the prediction module 30 is used to extract features from each slice block set to obtain meteorological features corresponding to each slice block set, and perform fusion and decoding of the meteorological features through a multi-scale fusion structure to obtain a multi-scale predicted meteorological field; the training module 40 is used to calculate the loss and update the parameters of the meteorological forecasting model based on the global meteorological field and the multi-scale predicted meteorological field.

[0089] Based on the above embodiments, the present invention also provides a terminal device, the principle block diagram of which can be as follows: Figure 8As shown, the terminal device includes a processor, memory, network interface, display screen, and temperature sensor connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements a weather forecasting model training method based on multi-scale feature fusion. The display screen can be an LCD screen or an e-ink screen. The temperature sensor is pre-installed inside the terminal device to detect the operating temperature of the internal components.

[0090] Those skilled in the art will understand that Figure 8 The schematic diagram shown is only a partial structural diagram related to the present invention and does not constitute a limitation on the terminal device to which the present invention is applied. The specific terminal device may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements.

[0091] In one embodiment, a terminal device is provided, including a memory and one or more programs, wherein one or more programs are stored in the memory and configured to be executed by one or more processors, the one or more programs including instructions for performing operations as described in the embodiments of the methods above.

[0092] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0093] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0094] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A weather forecast model training method based on multi-scale feature fusion, characterized in that, The method includes: Obtain raw meteorological data, preprocess the raw meteorological data, and obtain the global meteorological field; The global meteorological field is sliced ​​using at least two different scale slicing rules to obtain several slice block sets. Each slice block set corresponds to a slicing rule of one scale. All slice blocks in each slice block set have the same scale, and any slice block set completely constitutes the global meteorological field. Feature extraction is performed on each slice set to obtain the meteorological features corresponding to each slice set. The meteorological features are then fused and decoded through a multi-scale fusion structure to obtain a multi-scale predicted meteorological field. The loss is calculated and the parameters of the weather forecast model are updated based on the global meteorological field and the multi-scale predicted meteorological field. 2.The weather forecast model training method based on multi-scale feature fusion according to claim 1, characterized in that, The process of acquiring raw meteorological data and preprocessing the raw meteorological data to obtain a global meteorological field includes: Obtain raw meteorological data and convert the raw meteorological data into a format that is uniform; By interpolating to complete the data, missing values ​​in the uniformly formatted meteorological data are filled in to obtain complete meteorological data. The complete meteorological data is normalized, time-aligned, and gridded to obtain a standard global meteorological field. 3.The weather forecast model training method based on multi-scale feature fusion according to claim 1, characterized in that, The process involves extracting features from each set of slice blocks to obtain the meteorological features corresponding to each set of slice blocks, and then fusing and decoding the meteorological features using a multi-scale fusion structure to obtain a multi-scale predicted meteorological field, including: Meteorological feature extraction is performed independently on each slice set to obtain the meteorological features corresponding to the slice set; Using the same shared core feature operator, global feature modeling and parallel computation are performed on the meteorological features corresponding to each slice set to obtain the global meteorological features corresponding to each slice set. Decode the global meteorological features corresponding to each slice set independently to obtain the single-scale predicted meteorological field corresponding to each slice set at the target time. The single-scale predicted meteorological fields corresponding to all slice blocks are weighted and fused to obtain the multi-scale predicted meteorological field at the target time. 4.The weather forecast model training method based on multi-scale feature fusion according to claim 3, characterized in that, The parameters used to calculate losses and update the weather forecast model based on the global meteorological field and the multi-scale predicted meteorological field include: Obtain the true value label of the global meteorological field at the target time; Calculate the loss between the true value label and the multi-scale predicted meteorological field to obtain the loss value; Based on the loss value, update the independent meteorological feature extraction branch parameters at each scale, the independent decoding branch parameters at each scale, the weighted fusion learnable weight parameters, and the shared core feature operator parameters. 5.The weather forecast model training method based on multi-scale feature fusion according to claim 1, characterized in that, The process involves extracting features from each set of slice blocks to obtain the meteorological features corresponding to each set of slice blocks, and then fusing and decoding the meteorological features using a multi-scale fusion structure to obtain a multi-scale predicted meteorological field, including: Meteorological feature extraction is performed independently on each slice set to obtain the meteorological features corresponding to each slice set; By using the expert feature branches corresponding to each slice set, global feature modeling and parallel computation are performed on the meteorological features corresponding to each slice set to obtain the global meteorological features corresponding to each slice set. Among them, each slice set at each scale corresponds to a unique expert feature branch. The global meteorological features corresponding to each slice block set are weighted and fused to obtain the fused global meteorological features; A unified decoding process is performed on the fused global meteorological features to obtain a multi-scale predicted meteorological field.

6. The meteorological forecasting model training method based on multi-scale feature fusion according to claim 5, characterized in that, The parameters used to calculate losses and update the weather forecast model based on the global meteorological field and the multi-scale predicted meteorological field include: Obtain the real meteorological labels for the target time corresponding to the global meteorological field; Calculate the loss between the real weather label and the multi-scale predicted weather field to obtain the loss value; Based on the loss value, update the independent expert feature branch parameters for each scale, as well as the learnable weight parameters for weighted fusion.

7. A weather forecast model application method based on multi-scale feature fusion, characterized in that, The application method for a meteorological forecast model obtained by the meteorological forecast model training method based on multi-scale feature fusion according to any one of claims 1-6 includes: Raw meteorological data is acquired and preprocessed to obtain the global meteorological field; The global meteorological field is input into the trained meteorological forecasting model to obtain the multi-scale predicted meteorological field at the target time.

8. A weather forecasting model system based on multi-scale feature fusion, characterized in that, The system includes: The preprocessing module is used to acquire raw meteorological data, preprocess the raw meteorological data, and obtain the global meteorological field. The slicing module is used to slice the global meteorological field using at least two different scale slicing rules to obtain several slice block sets. Each slice block set corresponds to a slicing rule of one scale. All slice blocks in each slice block set have the same scale, and any slice block set completely constitutes the global meteorological field. The prediction module is used to extract features from each set of slice blocks to obtain the meteorological features corresponding to each set of slice blocks, and to perform fusion and decoding of the meteorological features through a multi-scale fusion structure to obtain a multi-scale predicted meteorological field. The training module is used to calculate the loss and update the parameters of the weather forecast model based on the global meteorological field and the multi-scale predicted meteorological field.

9. A terminal device, comprising: The terminal device includes a memory, a processor, and a meteorological forecasting model training program based on multi-scale feature fusion stored in the memory and executable on the processor. When the processor executes the meteorological forecasting model training program based on multi-scale feature fusion, it implements the steps of the meteorological forecasting model training method based on multi-scale feature fusion as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a meteorological forecasting model training program based on multi-scale feature fusion. When the meteorological forecasting model training program based on multi-scale feature fusion is executed by a processor, it implements the steps of the meteorological forecasting model training method based on multi-scale feature fusion as described in any one of claims 1-7.

Images