A physically consistent intelligent correction method for sub-seasonal model prediction products

Abstract

This invention discloses an intelligent correction method for the physical consistency of sub-seasonal model forecast products, belonging to the interdisciplinary field of artificial intelligence and meteorological forecasting. Multiple meteorological elements and multi-level circulation elements from the model forecast are used as raw data. First, the deviation field between atmospheric reanalysis data and model forecast data is calculated. Then, a multi-scale feature extraction network is used to obtain deviation features at different spatial scales. These multi-scale deviation features are concatenated with the original input features and input into a U-Net. The U-Net encoder and decoder are used for deep feature fusion and mapping. Simultaneously, a composite loss function that integrates numerical accuracy and physical consistency is designed to guide the collaborative training of the multi-scale feature extraction network and the U-Net network, resulting in a trained intelligent correction model. In application, the intelligent correction model outputs a physically consistent deviation correction amount to intelligently correct real-time model forecast data, ultimately obtaining a sub-seasonal forecast field that combines prediction progress and physical consistency.

Description

Technical Field

[0001] This invention relates to the field of meteorological data processing technology, specifically to an intelligent correction method for the physical consistency of sub-seasonal model forecast products. Background Technology

[0002] Subseasonal-seasonal forecasts (S2S) refer to the prediction of weather and climate on a timescale between conventional weather forecasts and seasonal forecasts, typically ranging from two weeks to two months. Subseasonal forecasts provide strong guidance and information for industries closely related to weather and climate conditions, such as power dispatching, fisheries production, and agricultural development. However, the predictability of numerical weather prediction for subseasonal to seasonal transitions is theoretically limited by the uncertainties in the parameters of physical models and the errors in initial values. Current subseasonal precipitation forecasts still have significant biases and are insufficient for practical operational forecasting.

[0003] To address this, traditional statistical methods or deep learning methods are mainly used to correct errors in S2S sub-seasonal forecast data. Deep learning correction methods, centered on CNNs and U-Nets, transform meteorological field data into gridded features and learn model forecast error characteristics through nonlinear mapping, but these methods often focus on correcting single meteorological elements. Traditional statistical correction methods, such as Model Output Statistics (MOS) and Proportional Prediction (PP), remain widely used in operational practice as fundamental tools. They correct basic biases in sub-seasonal forecasts by constructing statistical relationships between historical model outputs and observed values. In addition, some studies have proposed methods for optimizing meteorological forecasts using multi-scale feature fusion, such as the publication CN115049443A, "An Adaptive Regression Meteorological Medium- and Long-Term Optimization Method and System Based on Multi-Scale Features." This method uses ECMWF S2S forecast data as a foundation, combines reanalysis data and station data, extracts multi-scale features at different grid resolutions using an attention model, and fuses large-scale fluctuation signals with an adaptive regression (AR) algorithm to optimize near-surface meteorological element forecasts.

[0004] While existing subseasonal forecast correction technologies each have their advantages, their core shortcomings remain prominent, hindering forecast accuracy, physical consistency, and operational practicality. Regarding the embedding of physical mechanisms, traditional statistical methods, multi-scale feature fusion methods, and deep learning methods all fail to embed core atmospheric physical laws into the core correction process. They lack consideration of physical constraints and the coupling relationships between elements, leading to physical inconsistencies in the corrected forecast field, such as geostrophic imbalance and abrupt changes in vorticity mimicry, which accumulate with increasing forecast lead time. In terms of capturing multi-scale biases, existing technologies are mismatched with the inherent scale characteristics of the atmospheric system. They either perform simple fusion around grid resolution without designing differentiated extraction strategies for planetary, weather, and convective scales, or lack multi-scale design, only able to fit single-scale errors, resulting in insufficient correction accuracy for key systems such as mesoscale cyclones and fronts.

[0005] Furthermore, existing technologies exhibit significant imbalances in model design. For instance, purely data-driven methods are prone to inadequate dynamics, physical constraint improvement methods suffer from cumbersome architectures and difficulty in business implementation, and statistical methods have limited fitting capabilities. An effective balance between physical constraints and lightweight models has not been found. Regarding generalization ability and business adaptability, existing technologies also lack specific physical modeling for systematic biases, robust design, and generalization capabilities, resulting in sharp performance drops under extreme weather conditions or long-term forecasts of 30-60 days. Moreover, they struggle to balance accuracy and efficiency, with slow inference in complex networks and poor performance of statistical methods, failing to meet the dual requirements of high accuracy and real-time performance from business applications. Summary of the Invention

[0006] The present invention aims to at least partially solve one of the technical problems existing in the related art.

[0007] The purpose of this invention is to provide an intelligent correction method for physical consistency of sub-seasonal model forecast products. It constructs a physics-guided deep learning correction model, takes the original forecast field of the numerical model as input, and outputs a physically consistent correction field. Through innovative model architecture, physical constraint loss function and multi-scale feature extraction, it achieves the unity of systematic bias correction, physical consistency guarantee and operational efficiency, and comprehensively improves the accuracy and practicality of sub-seasonal forecast correction.

[0008] To achieve the above objectives, the present invention provides a method for intelligent correction of the physical consistency of sub-seasonal model forecast products, comprising the following steps:

[0009] S1. Obtain historical atmospheric reanalysis data and corresponding model forecast data, extract multiple meteorological elements, perform preprocessing, and calculate the bias field of each meteorological element on the two types of data.

[0010] S2. The deviation field and multiple meteorological elements extracted from the model forecast data are used as the original input features and input into a multi-scale feature extraction network. The multiple parallel paths of the multi-scale feature extraction network are used to capture the systematic deviation features at different spatial scales respectively. The feature maps output by each parallel path are fused through a gated fusion unit to generate multi-scale deviation features.

[0011] S3. The multi-scale deviation features are concatenated with the original input features and then input into the U-Net network. The core encoder and decoder are used to perform deep feature fusion and mapping, and the deviation correction amount of each meteorological element is output.

[0012] S4. Design a composite loss function that integrates numerical accuracy and physical consistency. Use this composite loss function to guide the training of the multi-scale feature extraction network and the U-Net network to obtain the bias correction model after training.

[0013] S5. Process the real-time model forecast data to be corrected using the same preprocessing method as in step S1, and input it into the trained bias correction model. Output a physically consistent bias correction amount, and use this bias correction amount to correct the real-time model forecast data to be corrected.

[0014] A further preferred technical solution of the present invention is as follows: Step S1 involves acquiring historical atmospheric reanalysis data and corresponding period model forecast data, extracting multiple meteorological elements, performing preprocessing, and calculating the bias field of each meteorological element on the two types of data; specifically:

[0015] Acquire historical atmospheric reanalysis data and corresponding period model forecast data, and extract key meteorological elements from the two types of data, including geopotential height, wind field u component, wind field v component, 2m temperature and near-surface specific humidity;

[0016] Atmospheric reanalysis data are resampled using bilinear interpolation to a resolution consistent with the model forecast data, and only data with the same start time as the model forecast data are retained in time to ensure spatiotemporal matching between the two types of data.

[0017] For each grid point and each forecast lead time, calculate the bias field. , is represented as:

[0018] ;

[0019] in, Meteorological elements for atmospheric reanalysis data. For meteorological elements in model forecast data, the Bt units of each deviation field are consistent with the corresponding elements;

[0020] Each deviation field Meteorological elements corresponding to model forecast data The input features are merged using a channel splicing method to form 10 channels, which are represented as follows: , , , , , , , , and ;

[0021] Standardize all input data.

[0022] Preferably, the multi-scale feature extraction network in step S2 includes three parallel pathways: a planetary scale pathway, a weather scale pathway, and a convective scale pathway.

[0023] The planetary-scale pathway includes two 5×5×5 three-dimensional convolutional layers with 32 and 64 kernels respectively, a stride of 1, a SAME padding method, and a ReLU activation function. A spectral attention mechanism is then applied, projecting the input feature map into 64-dimensional Q, K, and V matrices using a 1×1×1 convolution. Two-dimensional Fourier transforms are then performed on the Q, K, and V matrices to calculate the attention weights.

[0024] The features are then mapped back to the spatial domain through inverse Fourier transform, outputting a 64-channel planetary scale deviation feature map, focusing on extracting circulation system deviations with wavelengths greater than 1000km.

[0025] The weather-scale pathway includes two 3×3×3 deformable 3D convolutional layers with 32 and 64 kernels respectively, a stride of 1, and the activation function LeakyReLU with a negative slope of 0.01. It is accompanied by a 1×1×1 auxiliary convolutional layer with 81 output channels, corresponding to offsets in the x, y, and z directions, with 27 sampling points for each direction. The weather-scale pathway outputs a 64-channel weather-scale deviation feature map, enabling the deformable convolution to adaptively adjust the sampling position and accurately capture the weather system deviation morphology at the 100-1000km scale.

[0026] The convective-scale pathway includes two 3×3×3 three-dimensional convolutional layers with 32 and 64 kernels respectively, a stride of 1, a SAME padding method, and a ReLU activation function. It outputs a 64-channel convective-scale deviation feature map to capture convective-scale deviation features less than 100 km.

[0027] Preferably, in step S2, after obtaining the planetary scale deviation feature map, weather scale deviation feature map, and convective scale deviation feature map using the multi-scale feature extraction network, the gated fusion unit calculates adaptive weight coefficients for the 64-channel feature maps output from the three parallel paths using the sigmoid function, as follows:

[0028] ;

[0029] ;

[0030] ;

[0031] in, , , A 1×1×1 convolution is used to unify the feature dimensions; , and These represent planetary-scale bias characteristic maps, synoptic-scale bias characteristic maps, and convective-scale bias characteristic maps, respectively.

[0032] Then, the feature maps output from the three parallel paths are fused and represented as follows:

[0033] ;

[0034] The final output is a 64-channel multi-scale bias feature. .

[0035] Preferably, in step S3, the multi-scale deviation features are concatenated with the original input features, then input into the U-Net network, and the core encoder and decoder are used for deep feature fusion and mapping to output the deviation correction amount of each meteorological element; specifically:

[0036] The multi-scale bias features are concatenated with the original input features and then fed into the U-Net network;

[0037] The U-Net network encoder performs downsampling through four max pooling operations to gradually extract abstract features. Before each downsampling, the feature expression is enhanced by 3×3 convolution and ReLU activation function.

[0038] The U-Net network decoder performs upsampling through four transposed convolutions to restore spatial details; after each upsampling, it makes skip connections with the feature maps of the corresponding layers of the encoder to preserve the low-level spatial information.

[0039] The U-Net network decoder uses 1×1 convolutions to map features into bias correction fields for each meteorological element.

[0040] As a preferred option, the composite loss function designed in step S4 that integrates numerical accuracy and physical consistency includes a numerical accuracy loss function and a physical consistency loss function.

[0041] The numerical accuracy loss function uses mean squared error to measure the amount of bias correction in the model's predictions. Deviation from reality The difference between them is expressed as:

[0042] ;

[0043] The physical consistency loss function measures the degree to which the corrected field violates key physical laws. It includes two core constraint terms: the geostrophic wind constraint loss, which ensures that the corrected wind field and the height field satisfy approximately geostrophic equilibrium. And the vorticity pseudoenergy constraint loss due to unreasonable abrupt changes in atmospheric vortex intensity during the constraint correction process. ;

[0044] Among them, geostrophic wind constraint loss The root mean square error between the actual wind and the geostrophic wind on the 500 hPa isobaric surface of the correction field is calculated and expressed as:

[0045] ;

[0046] In the formula For actual wind Quantity, For actual wind Quantity, For the earth to turn wind Quantity, For the earth to turn wind The formulas for calculating the two components of geostrophic wind are as follows:

[0047] ;

[0048] ;

[0049] In the formula For potential height, For Coriolis parameters;

[0050] vorticity pseudoenergy-constrained loss Calculate the relative vorticity before and after correction The rate of change of pseudo-energy is expressed as:

[0051] ;

[0052] in, For time step, Indicates the average of the regional space. This is the corrected relative vorticity. The relative vorticity before correction; the formula for calculating relative vorticity is:

[0053] ;

[0054] A composite loss function that combines numerical accuracy and physical consistency , is represented as:

[0055] ;

[0056] in, These are the weighting coefficients.

[0057] Preferably, a composite loss function that integrates numerical accuracy and physical consistency is used to guide the training of the multi-scale feature extraction network and the U-Net network. During training, all convolutional layers are initialized using a He normal distribution; the Adam optimizer is used with an initial learning rate of 0.001. Parameters are updated to optimize the target; Dropout layers are enabled during training to prevent overfitting, while Dropout is disabled on the validation and test sets;

[0058] The training process employs an early stopping strategy, which monitors the changes in validation set loss in real time. When the validation set loss stops decreasing for 10 consecutive epochs, training is immediately stopped, and the model parameters with the best performance on the validation set are saved. The maximum number of iterations is set to 200 epochs.

[0059] Preferably, in step S5, the real-time model forecast data to be corrected is... The same preprocessing method as in step S1 is used, and the preprocessed bias correction model is input to output a physically consistent bias correction amount. Final revision of the forecast field The calculation formula is:

[0060] .

[0061] In another aspect, the present invention provides a non-transitory computer-readable storage medium having computer instructions stored thereon, the computer instructions causing a computer to execute the above-described intelligent correction method for physical consistency of sub-seasonal model forecast products.

[0062] In another aspect, the present invention provides an electronic device, comprising: a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus, and the processor calls logical instructions in the memory to execute the above-mentioned intelligent correction method for physical consistency of sub-seasonal model forecast products.

[0063] In another aspect, the present invention provides a computer program product comprising a computer program stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer performs the aforementioned intelligent correction method for physical consistency of the sub-seasonal model forecast product.

[0064] Beneficial effects: The intelligent correction method for physical consistency of sub-seasonal model forecast products in this invention designs parallel pathways at three scales: planetary, weather, and convection. Combined with a gated fusion unit, it adaptively extracts the deviation characteristics of circulation systems at different scales, thus solving the core deficiency of existing technologies where a single scale cannot adapt to multi-scale errors.

[0065] This invention removes physical constraints from the complex internal network layer, designs a composite loss function of geostrophic wind and eddy energy, and guides model training through backpropagation of the loss. While ensuring physical consistency, it greatly simplifies the model architecture and reduces the difficulty of debugging.

[0066] This invention integrates the pattern field and deviation field characteristics of multiple elements through channel splicing, enabling the model to simultaneously learn the deviation correlation laws of multiple elements, preserving the inherent physical correlation between elements, and solving the physical inconsistency problem ensured by traditional independent correction.

[0067] This invention replaces the original three-dimensional convolution with two-dimensional convolution, simplifies the number of network parameters, and combines U-Net skip connections to enhance feature reuse, thereby improving correction accuracy and achieving efficient business-oriented inference. Attached Figure Description

[0068] Figure 1 A flowchart illustrating the intelligent correction method for physical consistency of next-season model forecast products in an embodiment;

[0069] Figure 2 This example shows a comparison of the predicted and actual precipitation over the next 15-20 days. Figure 2 (a) shows the forecast result before the lead time correction. Figure 2 (b) shows the forecast result after the lead time correction. Figure 2 (c) shows the actual observation results;

[0070] Figure 3 This example shows a comparison of the predicted and actual precipitation over the next 21-30 days. Figure 2 (a) shows the forecast result before the lead time correction. Figure 2 (b) shows the forecast result after the lead time correction. Figure 2 (c) shows the actual observation results;

[0071] Figure 4 The above examples show a comparison of the predicted and actual precipitation over the next 31-60 days. Figure 2 (a) shows the forecast result before the lead time correction. Figure 2 (b) shows the forecast result after the lead time correction. Figure 2 (c) shows the actual observation results. Detailed Implementation

[0072] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, embodiments of this invention, and should not be construed as limiting the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention. In the description of this invention, it should be understood that the terminology used is for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0073] Example 1: This example provides a method for intelligent correction of the physical consistency of sub-seasonal model forecast products, such as... Figure 1 As shown, it includes the following steps:

[0074] S1. Input data preparation and deviation field calculation.

[0075] 1. Data Acquisition: In this embodiment, the reanalysis observation data (Ot) uses the fifth-generation global climate and atmosphere reanalysis dataset (ERA5) provided by the European Centre for Medium-Range Weather Forecasts (ECMWF), covering the period from 2003 to 2023, with a spatial resolution of 0.25°×0.25°. It includes geopotential height (6 standard layers: 200hPa, 300hPa, 500hPa, 700hPa, 850hPa, and 925hPa), wind field (u and v components), 2m temperature, near-surface specific humidity, and other key meteorological elements. The time frequency is 4 times per day (00Z, 06Z, 12Z, and 18Z), and the data format is NetCDF.

[0076] The model forecast data (Ft) were obtained from the S2S forecast system (BCC-S2S v2.0) of the National Climate Center's integrated sub-seasonal-seasonal-interannual scale climate model forecasting operational system. This model began quasi-operational operation in 2018, with historical backcalculation data ranging from 2003 to 2022, a spatial resolution of 0.5°×0.5°, a forecast lead time of 1-60 days, and a daily 00Z start time. The output elements are consistent with the ERA5 reanalysis data, and the data format is GRIB2.

[0077] 2. Data preprocessing: ERA5 data is resampled to a 0.5°×0.5° resolution in BCC-S2S v2.0 mode using bilinear interpolation. Only 00Z data consistent with the mode's reporting time is retained to ensure spatiotemporal matching between the two.

[0078] 3. Deviation field calculation: Calculate the deviation field for each grid point and each forecast lead time (ld = 1-60 days). , is represented as:

[0079] ;

[0080] in, Meteorological elements for atmospheric reanalysis data. The meteorological elements in the model forecast data are defined by the Bt unit for each deviation field, which is consistent with the corresponding element (e.g., height field unit is gpm, wind field unit is m / s, and temperature unit is K).

[0081] S2, Multi-scale bias feature extraction.

[0082] 1. Original input feature acquisition: First, extract each bias field Meteorological elements corresponding to model forecast data The input features are merged using a channel splicing method to form 10 channels, which are represented as follows: , , , , , , , , and ;

[0083] The input data is standardized and represented as follows:

[0084] ;

[0085] in, For a certain meteorological element, This represents the average value of the corresponding meteorological element. This represents the standard deviation of the corresponding meteorological element.

[0086] 2. Multi-scale bias feature extraction: The processed 10-channel raw input features are input into a multi-scale feature extraction network. Systematic bias patterns at different spatial scales are captured through three parallel pathways. Finally, a gated fusion unit generates comprehensive features. The specific structure is as follows:

[0087] (1) Planetary-scale pathway: The core structure consists of two 5×5×5 three-dimensional convolutional layers with 32 and 64 kernels respectively, a stride of 1, SAME padding, and ReLU activation function. A spectral attention mechanism is then applied, projecting the input feature map into 64-dimensional Q, K, and V matrices using a 1×1×1 convolution. Two-dimensional Fourier transforms are then performed on the Q, K, and V matrices to calculate the attention weights, expressed as follows:

[0088] ;

[0089] in, Represents a two-dimensional Fourier transform. Indicates dimension;

[0090] The features are then mapped back to the spatial domain through inverse Fourier transform, outputting a 64-channel planetary scale deviation feature map, which focuses on extracting circulation system deviations with wavelengths greater than 1000km.

[0091] (2) Weather-scale pathway: The core structure consists of two 3×3×3 deformable three-dimensional convolutional layers with 32 and 64 kernels respectively, a stride of 1, and the activation function is LeakyReLU with a negative slope of 0.01. It is equipped with a 1×1×1 auxiliary convolutional layer with 81 output channels, corresponding to offsets in the x, y, and z directions, with an offset of 27 sampling points in each direction. The weather-scale pathway outputs a 64-channel weather-scale deviation feature map, enabling the deformable convolution to adaptively adjust the sampling position and accurately capture the weather system deviation morphology at the 100-1000km scale.

[0092] (3) Convective scale pathway: The core structure consists of two 3×3×3 three-dimensional convolutional layers with 32 and 64 convolutional kernels, a stride of 1, a padding method of SAME, and an activation function of ReLU. It outputs a 64-channel convective scale deviation feature map to capture convective scale deviation features less than 100km.

[0093] (4) Gated fusion unit: For the 64-channel feature maps output from the three parallel paths, adaptive weight coefficients are calculated using the sigmoid function, as follows:

[0094] ;

[0095] ;

[0096] ;

[0097] in, , , A 1×1×1 convolution is used to unify the feature dimensions; , and These represent planetary-scale bias characteristic maps, synoptic-scale bias characteristic maps, and convective-scale bias characteristic maps, respectively.

[0098] Then, the feature maps output from the three parallel paths are fused and represented as follows:

[0099] ;

[0100] The final output is a 64-channel multi-scale bias feature. .

[0101] S3. Prediction of bias correction based on U-Net.

[0102] Multi-scale bias features are concatenated with the original input features and then fed into the U-Net network, where the core encoder and decoder are used for deep feature fusion and mapping.

[0103] The U-Net network encoder performs downsampling through four max pooling operations to gradually extract abstract features. Before each downsampling, the feature expression is enhanced by 3×3 convolution and ReLU activation function.

[0104] The U-Net network decoder performs upsampling through four transposed convolutions to restore spatial details; after each upsampling, it makes skip connections with the feature maps of the corresponding layers of the encoder to preserve the low-level spatial information.

[0105] The U-Net network decoder uses 1×1 convolutions to map features into deviation corrections for each meteorological element.

[0106] S4. Construction of physical constraint loss function and model training.

[0107] 1. The core of this embodiment lies in designing a composite loss function that integrates numerical accuracy and physical consistency, guiding model training through this loss function rather than adding complex differentiable physical layers within the network. The designed composite loss function integrating numerical accuracy and physical consistency is as follows:

[0108] (1) Numerical precision loss function Mean squared error is used to measure the amount of correction for the bias in model predictions. Deviation from reality The difference between them is expressed as:

[0109] ;

[0110] (2) Physical consistency loss function : Measures the degree to which the correction field violates key physical laws, and includes two core constraint terms.

[0111] (I) Geostrophic wind constraint loss The wind field and height field after constraint correction satisfy approximately geostrophic equilibrium.

[0112] The root mean square error between the actual wind and the geostrophic wind on the 500 hPa isobaric surface of the correction field is calculated and expressed as:

[0113] ;

[0114] In the formula For actual wind Quantity, For actual wind Quantity, For the earth to turn wind Quantity, For the earth to turn wind The formulas for calculating the two components of geostrophic wind are as follows:

[0115] ;

[0116] ;

[0117] In the formula For potential height, For Coriolis parameters.

[0118] (II) Vortex pseudoenergy constraint loss Unreasonable abrupt changes in atmospheric vortex intensity during constraint correction.

[0119] Calculate the relative vorticity before and after correction The rate of change of pseudo-energy is expressed as:

[0120] ;

[0121] in, For time step, Indicates the average of the regional space. This is the corrected relative vorticity. The relative vorticity before correction; the formula for calculating relative vorticity is:

[0122] .

[0123] (3) The final composite loss function that combines numerical accuracy and physical consistency , is represented as:

[0124] ;

[0125] in, These are the weighting coefficients.

[0126] 2. The multi-scale feature extraction network and the U-Net network are trained using a composite loss function that combines numerical accuracy and physical consistency. The model training process is as follows:

[0127] (1) Data set division: 2003-2019 is the training set, 2020-2022 is the validation set, and 2023 is the independent test set; the training set is input into the model in batches of batch_size=16.

[0128] (2) Initialization and optimization: All convolutional layers were initialized using the He normal distribution; the Adam optimizer was used with an initial learning rate of 0.001. Parameters are updated to optimize the target; Dropout layers (ratio 0.2) are enabled during training to prevent overfitting, and Dropout is disabled on the validation and test sets;

[0129] (3) Early stopping strategy: Monitor the change of validation set loss in real time. When the validation set loss no longer decreases for 10 consecutive epochs, stop training immediately and save the model parameters with the best performance on the validation set; the maximum number of iterations is set to 200 epochs.

[0130] S5, Business-oriented correction application.

[0131] The real-time BCC-S2S v2.0 model forecast product awaiting correction. The same preprocessing method as in step S1 is used, and the preprocessed bias correction model is input to output a physically consistent bias correction amount. Final revision of the forecast field The calculation formula is:

[0132] .

[0133] The method described in this embodiment was used for experimental verification.

[0134] First, we train the model:

[0135] (1) Data source: The training and validation data are historical back-calculation data of BCC-S2S v2.0 model from 2003 to 2022 and ERA5 reanalysis data from 2003 to 2023. The test data are real-time forecast data of BCC-S2S v2.0 model in 2023 and corresponding ERA5 reanalysis data.

[0136] (2) Evaluation indicators: Numerical accuracy indicators include root mean square error (RMSE), mean absolute error (MAE), and correlation coefficient (CC); physical consistency indicators include geostrophic balance deviation and vorticity pseudo-energy change rate; operational performance indicators include time taken for single batch correction and the reduction in forecast deviation for extreme weather (high temperature in northern China and heavy rain in southern China in 2023).

[0137] Then, the trained model is used to select a specific event for verification:

[0138] The 2020 plum rain season in the Yangtze-Huaihe River Basin lasted for 52 days, with cumulative rainfall 1.5 times higher than the average for the same period, the highest since 1961, providing an extreme case test for sub-seasonal forecasting methods. This embodiment uses this event as an example, employing the sub-seasonal forecasting method of this embodiment, with May 21, 2020 as the starting point, to continuously produce precipitation correction products for three forecast periods: 15-20 days (June 5-10), 21-30 days (June 11-20), and 31-60 days (June 21-July 20). Figures 2-4The forecast results for three time periods are presented respectively. The comparison results before and after the correction show that the method of the present invention effectively suppresses the bias of the original model forecast, accurately reproduces the spatial distribution pattern of precipitation in the Meiyu area of ​​the Yangtze River and Huai River basins and its evolution over time. In particular, it has stable forecasting skills for the maintenance of the main rain belt and the period of concentrated heavy precipitation from mid-June to early July. It provides reliable sub-seasonal forecasting support for early warning and disaster prevention and mitigation deployment in the Yangtze River and Huai River basins with the extended Meiyu season, and has significant operational application value and socio-economic benefits.

[0139] Verification has shown that the intelligent correction method for physical consistency of sub-seasonal model forecast products implemented in this paper, compared with existing solutions, demonstrates the following advantages:

[0140] More accurate multi-scale error capture and significantly improved correction accuracy: Through the three-scale parallel path + gating fusion design in step S2, the deviation characteristics of the circulation system at different scales are extracted in a targeted manner.

[0141] Strict physical consistency is ensured to eliminate unreasonable forecast results: The geostrophic balance and eddy energy constraints are embedded into the entire model training process through the composite physical constraint loss function in step S4. Experimental results show that the geostrophic balance deviation in the corrected forecast field is significantly reduced, and the results conform to atmospheric physical laws, fundamentally eliminating physically unreasonable forecast results. This is particularly suitable for long-term sub-seasonal forecasts of 30-60 days.

[0142] The model architecture is simple and easy to debug, reducing the threshold for business application: physical constraints are simplified from the internal network layer to loss function guidance, reducing the number of model parameters by more than 40% and improving the training convergence speed by 30%; at the same time, the weight parameters of the loss function have clear physical meaning, making the debugging direction clear, and solving the problem of the traditional physical embedding model architecture being bloated and difficult to deploy.

[0143] Strong multi-element synergy and preservation of physical correlation: Through the multi-element channel splicing design in step S1, the model can simultaneously learn the deviation correlation laws of wind field, altitude field, temperature field, and humidity field. Experiments show that after correction, the correlation coefficient between wind field and altitude field increases from 0.65 to 0.88, the synergy between humidity field and precipitation field increases by 30%, and the thermal wind balance satisfaction increases by 80%. The physical correlation between elements is significantly better than existing technologies.

[0144] Good business applicability and a balance between computational efficiency and accuracy: Thanks to the lightweight architecture and U-Net's efficient inference capabilities, this invention can process 100 global forecast fields with a lead time of 1-60 days in a single batch in less than 3 minutes, which is 50% faster than existing methods; after the model is trained, it can be directly adapted to all the start times and years of the BCC-S2S v2.0 mode without the need for retraining for different time periods, making it easy to promote and apply in business.

[0145] Example 2: This example provides a non-transitory computer-readable storage medium storing computer instructions that cause a computer to execute a method for intelligent correction of the physical consistency of a next-season model forecast product. The method includes the following steps:

[0146] S1. Obtain historical atmospheric reanalysis data and corresponding model forecast data, extract multiple meteorological elements, perform preprocessing, and calculate the bias field of each meteorological element on the two types of data.

[0147] S2. The deviation field and multiple meteorological elements extracted from the model forecast data are used as the original input features and input into a multi-scale feature extraction network. The multiple parallel paths of the multi-scale feature extraction network are used to capture the systematic deviation features at different spatial scales respectively. The feature maps output by each parallel path are fused through a gated fusion unit to generate multi-scale deviation features.

[0148] S3. The multi-scale deviation features are concatenated with the original input features and then input into the U-Net network. The core encoder and decoder are used to perform deep feature fusion and mapping, and the deviation correction amount of each meteorological element is output.

[0149] S4. Design a composite loss function that integrates numerical accuracy and physical consistency. Use this composite loss function to guide the training of the multi-scale feature extraction network and the U-Net network to obtain the bias correction model after training.

[0150] S5. Process the real-time model forecast data to be corrected using the same preprocessing method as in step S1, and input it into the trained bias correction model. Output a physically consistent bias correction amount, and use this bias correction amount to correct the real-time model forecast data to be corrected.

[0151] Example 3: This example provides an electronic device that may include a processor, a communication interface, a memory, and a communication bus. The processor, communication interface, and memory communicate with each other via the communication bus. The processor can call logical instructions from the memory to execute a smart correction method for the physical consistency of next-season model forecast products. This method includes the following steps:

[0152] S1. Obtain historical atmospheric reanalysis data and corresponding model forecast data, extract multiple meteorological elements, perform preprocessing, and calculate the bias field of each meteorological element on the two types of data.

[0153] S2. The deviation field and multiple meteorological elements extracted from the model forecast data are used as the original input features and input into a multi-scale feature extraction network. The multiple parallel paths of the multi-scale feature extraction network are used to capture the systematic deviation features at different spatial scales respectively. The feature maps output by each parallel path are fused through a gated fusion unit to generate multi-scale deviation features.

[0154] S3. The multi-scale deviation features are concatenated with the original input features and then input into the U-Net network. The core encoder and decoder are used to perform deep feature fusion and mapping, and the deviation correction amount of each meteorological element is output.

[0155] S4. Design a composite loss function that integrates numerical accuracy and physical consistency. Use this composite loss function to guide the training of the multi-scale feature extraction network and the U-Net network to obtain the bias correction model after training.

[0156] S5. Process the real-time model forecast data to be corrected using the same preprocessing method as in step S1, and input it into the trained bias correction model. Output a physically consistent bias correction amount, and use this bias correction amount to correct the real-time model forecast data to be corrected.

[0157] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and sold or used as independent products, and can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0158] Example 4: This example provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute a smart correction method for the physical consistency of a next-season model forecast product. This method includes the following steps:

[0159] S1. Obtain historical atmospheric reanalysis data and corresponding model forecast data, extract multiple meteorological elements, perform preprocessing, and calculate the bias field of each meteorological element on the two types of data.

[0160] S2. The deviation field and multiple meteorological elements extracted from the model forecast data are used as the original input features and input into a multi-scale feature extraction network. The multiple parallel paths of the multi-scale feature extraction network are used to capture the systematic deviation features at different spatial scales respectively. The feature maps output by each parallel path are fused through a gated fusion unit to generate multi-scale deviation features.

[0161] S3. The multi-scale deviation features are concatenated with the original input features and then input into the U-Net network. The core encoder and decoder are used to perform deep feature fusion and mapping, and the deviation correction amount of each meteorological element is output.

[0162] S4. Design a composite loss function that integrates numerical accuracy and physical consistency. Use this composite loss function to guide the training of the multi-scale feature extraction network and the U-Net network to obtain the bias correction model after training.

[0163] S5. Process the real-time model forecast data to be corrected using the same preprocessing method as in step S1, and input it into the trained bias correction model. Output a physically consistent bias correction amount, and use this bias correction amount to correct the real-time model forecast data to be corrected.

[0164] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0165] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0166] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for physically consistent intelligent correction of sub-seasonal forecast products, characterized in that, Includes the following steps: S1. Obtain historical atmospheric reanalysis data and corresponding model forecast data, extract multiple meteorological elements, perform preprocessing, and calculate the bias field of each meteorological element on the two types of data. S2. The deviation field and multiple meteorological elements extracted from the model forecast data are used as the original input features and input into a multi-scale feature extraction network. The multiple parallel paths of the multi-scale feature extraction network are used to capture the systematic deviation features at different spatial scales respectively. The feature maps output by each parallel path are fused through a gated fusion unit to generate multi-scale deviation features. S3. The multi-scale deviation features are concatenated with the original input features and then input into the U-Net network. The core encoder and decoder are used to perform deep feature fusion and mapping, and the deviation correction amount of each meteorological element is output. S4. Design a composite loss function that integrates numerical accuracy and physical consistency. Use this composite loss function to guide the training of the multi-scale feature extraction network and the U-Net network to obtain the bias correction model after training. S5. Process the real-time model forecast data to be corrected using the same preprocessing method as in step S1, and input it into the trained bias correction model. Output a physically consistent bias correction amount, and use this bias correction amount to correct the real-time model forecast data to be corrected.

2. The physically consistent intelligent correction method of sub-seasonal pattern forecast product according to claim 1, characterized in that, Step S1 involves acquiring historical atmospheric reanalysis data and corresponding period model forecast data, extracting multiple meteorological elements, preprocessing them, and calculating the bias fields of each meteorological element across the two types of data; specifically: Acquire historical atmospheric reanalysis data and corresponding period model forecast data, and extract key meteorological elements from the two types of data, including geopotential height, wind field u component, wind field v component, 2m temperature and near-surface specific humidity; Atmospheric reanalysis data are resampled using bilinear interpolation to a resolution consistent with the model forecast data, and only data with the same start time as the model forecast data are retained in time to ensure spatiotemporal matching between the two types of data. For each grid point, each forecast time, the bias field is computed is expressed as: ； wherein, are meteorological elements of the atmospheric reanalysis data, are meteorological elements of the model forecast data, each bias field Bt unit is consistent with the corresponding element; Each deviation field Meteorological elements corresponding to model forecast data The input features are merged using a channel splicing method to form 10 channels of original input features, which are represented as follows: , , , , , , , , and ; Then, the original input features are standardized.

3. The physically consistent intelligent correction method of sub-seasonal model forecast product according to claim 1, characterized in that, The multi-scale feature extraction network in step S2 includes three parallel pathways: a planetary-scale pathway, a weather-scale pathway, and a convective-scale pathway. The planetary-scale pathway includes two 5×5×5 three-dimensional convolutional layers with 32 and 64 kernels respectively, a stride of 1, SAME padding, and ReLU activation. A spectral attention mechanism is then applied, projecting the input feature map using a 1×1×1 convolution into 64-dimensional Q, K, and V matrices. Two-dimensional Fourier transforms are then performed on the Q, K, and V matrices to calculate the attention weights, expressed as follows: ； wherein denotes a two-dimensional Fourier transform, denotes a dimension; The features are then mapped back to the spatial domain through inverse Fourier transform, outputting a 64-channel planetary scale deviation feature map, focusing on extracting circulation system deviations with wavelengths greater than 1000km. The weather-scale pathway includes two 3×3×3 deformable 3D convolutional layers with 32 and 64 kernels respectively, a stride of 1, and the activation function LeakyReLU with a negative slope of 0.

01. It is accompanied by a 1×1×1 auxiliary convolutional layer with 81 output channels, corresponding to offsets in the x, y, and z directions, with 27 sampling points for each direction. The weather-scale pathway outputs a 64-channel weather-scale deviation feature map, enabling the deformable convolution to adaptively adjust the sampling position and accurately capture the weather system deviation morphology at the 100-1000km scale. The convective-scale pathway includes two 3×3×3 three-dimensional convolutional layers with 32 and 64 kernels respectively, a stride of 1, a SAME padding method, and a ReLU activation function. It outputs a 64-channel convective-scale deviation feature map to capture convective-scale deviation features less than 100 km.

4. The physically consistent intelligent correction method of sub-seasonal model forecast product according to claim 3, characterized in that, In step S2, after obtaining planetary-scale bias feature maps, weather-scale bias feature maps, and convective-scale bias feature maps using the multi-scale feature extraction network, the gating fusion unit calculates adaptive weight coefficients for the 64-channel feature maps output from the three parallel paths using the sigmoid function, as shown below: ； ； ； in, , , A 1×1×1 convolution is used to unify feature dimensions; , and These represent planetary-scale bias characteristic maps, synoptic-scale bias characteristic maps, and convective-scale bias characteristic maps, respectively. Then, the feature maps output from the three parallel paths are fused and represented as follows: ； The final output is a 64-channel multi-scale bias feature .

5. The method for physically consistent intelligent correction of sub-seasonal model forecast products according to claim 1, wherein, In step S3, the multi-scale bias features are concatenated with the original input features, then input into the U-Net network. The core encoder-decoder performs deep feature fusion and mapping, outputting the bias correction field for each meteorological element; specifically: The multi-scale bias features are concatenated with the original input features and then fed into the U-Net network; The U-Net network encoder performs downsampling through four max pooling operations to gradually extract abstract features. Before each downsampling, the feature expression is enhanced by 3×3 convolution and ReLU activation function. The U-Net network decoder performs upsampling through four transposed convolutions to restore spatial details; after each upsampling, it makes skip connections with the feature maps of the corresponding layers of the encoder to preserve the low-level spatial information. The U-Net network decoder uses 1×1 convolutions to map features into deviation corrections for each meteorological element.

6. The method for physically consistent intelligent correction of sub-seasonal model forecast products according to claim 1, wherein, The composite loss function designed in step S4, which integrates numerical accuracy and physical consistency, includes a numerical accuracy loss function and a physical consistency loss function. Wherein the numerical accuracy loss function adopts mean square error, for measuring the deviation correction quantity of model prediction The difference between the true deviation is expressed as: ； The physical consistency loss function measures the degree to which the corrected field violates key physical laws. It includes two core constraint terms: the geostrophic wind constraint loss, which ensures that the corrected wind field and the height field satisfy approximately geostrophic equilibrium. And the vorticity pseudoenergy constraint loss due to unreasonable abrupt changes in atmospheric vortex intensity during the constraint correction process. ; Among them, geostrophic wind constraint loss The root mean square error between the actual wind and the geostrophic wind on the 500 hPa isobaric surface of the correction field is calculated and expressed as: ； In the formula For actual wind Quantity, For actual wind Quantity, For the earth to turn wind Quantity, For the earth to turn wind The formulas for calculating the two components of geostrophic wind are as follows: ； ； wherein is the geopotential height, is the Coriolis parameter; vorticity pseudoenergy-constrained loss Calculate the relative vorticity before and after correction The rate of change of pseudo-energy is expressed as: ； where, is the time step, denotes the area space average, is the corrected relative vorticity, is the uncorrected relative vorticity; the relative vorticity is calculated as ； Fusing numerical precision and physical consistency of composite loss function is represented as: ； wherein are weight coefficients.

7. The physically consistent intelligent correction method of sub-seasonal pattern forecast product according to claim 6, characterized in that, The multi-scale feature extraction network and the U-Net network are trained using a composite loss function that combines numerical accuracy and physical consistency. During training, all convolutional layers are initialized using a He normal distribution; the Adam optimizer is used with an initial learning rate of 0.

001. Parameters are updated to optimize the target; Dropout layers are enabled during training to prevent overfitting, while Dropout is disabled on the validation and test sets; The training process employs an early stopping strategy, which monitors the changes in validation set loss in real time. When the validation set loss stops decreasing for 10 consecutive epochs, training is immediately stopped, and the model parameters with the best performance on the validation set are saved. The maximum number of iterations is set to 200 epochs.

8. The physically consistent intelligent correction method of sub-seasonal pattern forecast product according to claim 6, wherein, In step S5, the real-time model forecast data to be corrected is... The same preprocessing method as in step S1 is used, and the preprocessed bias correction model is input to output a physically consistent bias correction amount. Final revision of the forecast field The calculation formula is: 。 9. A non-transitory computer-readable storage medium, comprising: It stores computer instructions that cause the computer to execute the intelligent correction method for physical consistency of the sub-seasonal model forecast product as described in any one of claims 1-9.

10. An electronic device, comprising: include: The system includes a processor, a communication interface, a memory, and a communication bus. The processor, communication interface, and memory communicate with each other via the communication bus. The processor calls logical instructions from the memory to execute the intelligent correction method for physical consistency of the sub-seasonal model forecast product as described in any one of claims 1-9.

Images