A sub-seasonal prediction method and system based on the combination of dynamic mode downscaling and machine learning downscaling

Abstract

The application belongs to the field of sub-seasonal climate prediction, and provides a sub-seasonal climate prediction method and system based on the combination of dynamic model downscaling and machine learning downscaling, which comprises the following steps: S1, generating initial field and model boundary field information required for dynamic downscaling based on global climate model output data; S2, driving regional climate model to perform dynamic downscaling by using the initial field and model boundary field information; S3, generating input field required for machine learning downscaling based on the output circulation field information of the regional climate model; S4, performing machine learning downscaling correction optimization on the output circulation field of dynamic downscaling based on a convolution model; S5, performing machine learning super-resolution based on the output data of machine learning downscaling correction; and S6, generating sub-seasonal prediction information of double downscaling of dynamic model and machine learning. The application utilizes the complementary advantages of dynamic downscaling and deep learning downscaling, improves the sub-seasonal prediction skill, and copes with new challenges brought by climate change.

Description

Technical Field

[0001] This invention relates to the fields of meteorology and climate, and in particular to a method and system for subseasonal forecasting based on a combination of dynamic model downscaling and machine learning downscaling. Background Technology

[0002] Subseasonal forecasting refers to the systematic prediction of weather and climate on a timescale of two weeks to two months, filling the gap between weather forecasts (less than 10 days) and seasonal climate forecasts (more than 3 months). This timescale achieves a better balance between accuracy and timeliness, providing sufficiently detailed forecast information while leaving a crucial response window for early warning and scientific decision-making, thus being regarded as a "golden window of opportunity" in climate services. Subseasonal forecasting is of great value in disaster prevention and mitigation, providing early warning information for disasters such as wildfires, floods, and tropical cyclones; in resource management, subseasonal forecasting provides key decision support for sectors such as renewable energy, water resources, shipping, and agriculture; in addition, subseasonal forecasting is of great significance to agricultural planning, enabling the prediction of crop growth a month in advance, optimizing field management, agricultural arrangements, and agricultural product pricing. However, in this forecasting, the initial atmospheric signals have already decayed rapidly, while the boundary signals of the climate system such as sea surface temperature have not yet fully manifested, resulting in relatively low forecasting skill, and is considered a "predictability desert" (Vitart et al., 2018).

[0003] In the Coupled Model Intercomparison Project (CMIP6), approximately 80% of global climate models only accurately capture the transition to subseasonal timescales in a few years (Song et al., 2023). In the World Meteorological Organization's (WMO) Subseason-to-Season (S2S) program, mainstream international operational models have limited subseasonal forecasting skill for precipitation and extreme precipitation. For global land-based weekly average precipitation, most S2S models' forecasting skill is limited to the first week (de Andrade et al., 2019).

[0004] Currently, my country's subseasonal climate prediction is still in the exploratory and improvement stage (Zhu et al., 2019). Although numerical models have made some progress in operational climate prediction research, my country's location in the East Asian monsoon region, its large north-south span, complex topography, and variable influencing systems, coupled with uncertainties in initial values ​​and the models themselves, result in a generally low level of subseasonal climate prediction in my country. Furthermore, dynamical models represent atmospheric states through discretized grids and describe state transitions through numerical solutions of partial differential equations, involving high computational costs and low computational efficiency (Bi et al., 2023). In summary, besides high computational costs and low computational efficiency, the main shortcomings of dynamical models also include: limited accuracy in predicting local elements, imperfect parameterization of physical processes, complex sources of predictability, and insufficient ability to predict extreme events.

[0005] Deep learning is another approach for sub-seasonal forecasting. Although deep learning methods excel at modeling nonlinear systems, the models still suffer from inherent small-sample oversmoothing problems, and the forecast solution process lacks physical constraints. Therefore, data-driven weather / climate forecasting faces the numerical field ambiguity effect (Lam et al., 2023), which is the main reason why current deep learning sub-seasonal forecasting performance is slightly inferior to the (ECMWF) dynamical model ensemble forecast (Zhong and Wu, 2023). Its main drawbacks include high data requirements, poor interpretability, limited extrapolation capabilities, and insufficient ability to predict extreme events. Summary of the Invention

[0006] To overcome the aforementioned problems in existing technologies, this invention provides a method and system for subseasonal climate prediction based on a combination of dynamical model downscaling and machine learning downscaling, thereby addressing the problems present in existing technologies. The technical solution of this invention is as follows: This invention provides a sub-seasonal climate prediction method based on a combination of dynamical model downscaling and machine learning downscaling. The method includes the following steps: S1 generating initial field and model boundary field information required for dynamical downscaling based on global climate model output data; S2 using the initial field and model boundary field information to drive regional climate models for dynamical downscaling; S3 generating the input field required for machine learning downscaling based on the regional climate model output circulation field information; S4 performing machine learning downscaling correction optimization on the dynamically downscaled output circulation field based on a convolutional model; S5 performing machine learning super-resolution based on the machine learning downscaling correction output data; S6 generating sub-seasonal prediction information with dual downscaling from both dynamical models and machine learning. Further, S1 includes the following steps: S11 Global climate model output information collection, executing Secondary2NewPiYYYYMMDD.py to monitor and collect atmospheric, ocean, and land surface variable data for global model prediction integration in real time, ensuring the data input required for dynamic downscaling; S12 Global model data preprocessing, executing icbc_MMDD_yr.sh to process the real-time collected global model atmospheric, ocean, and land surface data to generate the regional model initial field and boundary field, obtaining the data format required for regional climate model dynamic downscaling prediction integration; Furthermore, the initial field and model boundary field include: initial field information including atmospheric initial field and surface initial field information, and model boundary field information including lateral boundary conditions and lower boundary conditions; Further, S2 includes the following steps: S21 Performing dynamic downscaling integration on the global climate initial field and model boundary field, executing make_exp_gcmsst.csh; S22 Calling the regional climate model executable file run_MMDD_yr-case.sh to start the model integration step; S23 Submitting the regional climate model run to the queue via submit_intel.sh to start dynamic downscaling, obtaining the integrated regional climate model dynamic downscaling prediction data product; Furthermore, the regional climate model mentioned in step S22 is a climate-scale extended version of a mesoscale weather forecasting model. This version is a regional climate model version with configuration optimization for long-term climate prediction integration, capable of stable integration operation for more than two months. Based on the mesoscale weather forecasting model, this version adopts a side-boundary relaxation approximation scheme and debugs and integrates the physical process parameterization scheme, which includes convection parameterization, microphysical process parameterization, radiative transfer parameterization, boundary layer parameterization, and land surface process parameterization schemes. Furthermore, step S3 involves systematically post-processing the dynamic downscaling prediction results of the regional climate model to output circulation field information. The data post-processing includes projection system conversion and data standardization. Furthermore, the circulation field information includes 19 key meteorological variable channels: geopotential height field and zonal wind field at four levels: 200 hPa, 500 hPa, 700 hPa and 850 hPa; relative humidity field, temperature field and meridional wind field at 500 hPa, 700 hPa and 850 hPa; and near-surface 2-meter temperature field and surface pressure field; Furthermore, in S4, the convolutional model structure SEDES-2D-Cw3m2s2-Pw3m2s1.onnx is designed using CRA-40 data for the nonlinear relationship between circulation and precipitation; Furthermore, the machine learning super-resolution in step S5 adopts the residual channel attention network machine learning super-resolution model w3.4xSR.onnx, which includes four modules: shallow feature extraction, deep feature extraction, feature fusion, and sampling reconstruction. This invention also provides a subseasonal climate prediction system based on a combination of dynamical model downscaling and machine learning downscaling. The system includes the following modules: an initial information generation module, which generates initial field and model boundary field information required for dynamical downscaling based on global climate model output data; a dynamical downscaling module, which uses the initial field and model boundary field information to drive a regional climate model for dynamical downscaling; an input information generation module, which generates the input field required for machine learning downscaling based on the regional climate model output circulation field information; a machine learning downscaling module, which performs machine learning downscaling correction optimization on the dynamically downscaled output circulation field based on a convolutional model; a machine learning super-resolution module, which performs machine learning super-resolution on the machine learning downscaling corrected output data; and a prediction information generation module, which generates subseasonal prediction information downscaled by both dynamical model and machine learning methods.

[0007] The beneficial effects of this invention are as follows: This invention combines dynamic downscaling with refined local influencing factors to construct a deep learning downscaling model based on physical mechanism constraints and local data-driven approaches. In line with the "accurate forecasting" requirements of modernizing meteorological science and technology capabilities and social services, it employs a dual downscaling method of dynamic downscaling and deep learning downscaling, leveraging the complementary advantages of these two approaches to improve sub-seasonal forecasting skills and address the new challenges brought about by climate change. Attached Figure Description

[0008] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.

[0009] Figure 1 This invention relates to a subseasonal climate prediction method based on a combination of dynamic model downscaling and machine learning downscaling. Figure 2 A schematic diagram of the corrected model structure for a convolutional neural network (CNN); Figure 3 Example of random sampling output for a Residual Channel Attention Network (RCAN) super-resolution model; Figure 4 Example chart of second-season forecast data downscaled using both dynamic model and machine learning; Figure 5 Spatial correlation coefficients of the percentage of monthly precipitation anomalies in eastern China in June 2024, calculated based on predictions from the global model, regional model, CNN correction model, and RCAN super-resolution model (reporting dates are May 8, 11, 15, 18, and 22, 2024, respectively).

[0010] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0011] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure as detailed in the appended claims.

[0012] The terms "first," "second," etc., used in the specification and claims of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented, for example, in orders other than those illustrated or described herein.

[0013] Furthermore, the terms “comprising” and “having”, and any variations thereof, are intended to cover non-exclusive inclusion, such that a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or units that are not explicitly listed or that are inherent to such process, method, product, or apparatus.

[0014] Multiple, including two or more. And / or, it should be understood that the term "and / or" as used in this disclosure is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, and B alone. It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0015] Global climate models (GCMs) and regional climate models (RCMs) are two different scale tools in climate simulation. Global climate models typically have a horizontal resolution of 100-300 kilometers, covering the entire Earth and capable of simulating global-scale climate systems and atmospheric circulation characteristics. Regional climate models, on the other hand, have a horizontal resolution of 10-50 kilometers or even higher, covering only specific regions and obtaining more refined local climate information through dynamic downscaling techniques. In terms of simulation capabilities, regional models can more accurately simulate local climate phenomena such as topographic forcing, sea-land breezes, urban heat islands, and valley winds, and are more capable of simulating extreme weather events.

[0016] The main limitations of current dynamical model forecasting techniques include the large uncertainty of initial atmospheric conditions due to limited observational data (e.g., in the Arctic region), and the insufficient accuracy of global models in characterizing the impact of boundary forcing (e.g., sea ice, sea surface temperature, stratospheric explosive warming) on ​​warm-cold transition processes. Compared to the large-scale information output by global climate models, dynamical downscaling regional climate models offer higher resolution and more locally adaptable physical parameter schemes, showing potential for predicting the sub-seasonal temporal scale characteristics of my country. China's complex topography makes it difficult to accurately characterize its topographic features using the spatial resolution of current mainstream operational global climate numerical models. Consequently, global models have certain deficiencies in describing local thermal circulation and are insufficient in describing boundary layer dynamics. Dynamical downscaling methods based on regional climate models can improve these problems to some extent, while also being more adaptable to regional physical process parameterization schemes. Sub-seasonal forecasting of climate elements is a complex task involving spatiotemporal dynamics, requiring the construction of accurate and stable spatiotemporal neural network forecasting models. Deep learning excels at constructing nonlinear models and predicting nonlinear processes. Dynamical climate models, as key tools for studying climate characteristics and mechanisms, demonstrate significant advantages in predicting large-scale ocean-atmosphere characteristics in their global models, while regional climate models can more accurately characterize complex local topography and boundary layer forcing. Deep learning technology, with its strong nonlinear modeling capabilities, can effectively capture complex spatiotemporal interactions among multiple time scales, multiple regions, and multiple influencing factors.

[0017] Based on the above advantages, this invention proposes to combine dynamic downscaling and deep learning downscaling in a dual downscaling prediction technique, and apply it to the national-level climate prediction operational platform to effectively promote the improvement of my country's sub-seasonal prediction capabilities.

[0018] See Figure 1 This invention relates to a subseasonal climate prediction method based on a combination of dynamical model downscaling and machine learning downscaling, the method comprising the following steps: S1 generates the initial field and model boundary field information required for dynamic downscaling based on global climate model output data; S2 uses initial field and model boundary field information to drive the dynamic downscaling of regional climate models; S3 generates the input field required for machine learning downscaling based on the circulation field information output from the regional climate model. S4, based on the convolutional model, performs machine learning downscaling correction optimization on the dynamic downscaled output circulation field; S5, based on machine learning downscaling correction of output data, performs machine learning super-resolution; S6 generates sub-seasonal precipitation forecasts using both dynamical models and machine learning downscaling.

[0019] Specifically: S1: The initial field and model boundary field required for dynamic downscaling based on global climate model output data include: 1) Global model output information collection part, whose main function is to monitor and collect data on atmospheric, ocean and land surface variables for global model prediction integration in real time. The data has a global horizontal resolution of about 45 kilometers, ensuring the data input required for dynamic downscaling; 2) Global model data preprocessing, whose main function is to process the real-time collected global model atmospheric, ocean and land surface data to generate regional model initial field and boundary field, and obtain the data format required for regional climate model dynamic downscaling prediction integration.

[0020] The initial and boundary field information for regional models, required for dynamic downscaling, is generated based on global climate output data. The initial field information mainly includes: 1. Atmospheric initial field: including three-dimensional meteorological elements such as atmospheric temperature, pressure, humidity, and wind field. These data are typically extracted from global climate models or reanalysis data and converted to the grid format required by the regional climate model through interpolation. 2. Surface initial field: including land surface process variables such as surface temperature, soil temperature and humidity, surface albedo, and vegetation cover. This information is crucial for the accuracy of regional climate simulations and needs to maintain physical consistency with the atmospheric initial field. The model boundary field information mainly includes: 1. Lateral boundary conditions: including time-varying information of variables such as atmospheric temperature, pressure, humidity, and wind field at the regional model boundary. This data is provided at a temporal resolution of 6 hours or less and is used to drive the regional model to maintain consistency with the global model throughout the simulation process. 2. Lower boundary conditions: mainly sea surface temperature (SST) and sea ice cover data. These data have a significant impact on regional climate simulations, especially the climate characteristics of coastal and marine areas, and usually require high spatiotemporal resolution SST data as a driving force.

[0021] The quality of these initial and boundary field information directly affects the accuracy and reliability of dynamic downscaling simulations, and serves as fundamental data support for regional climate research.

[0022] Dynamic downscaling is based on solving physical equations (such as atmospheric dynamics and thermodynamics) at a regional scale. By using nesting techniques, the output of the global model is used as the boundary condition to drive the regional climate model (RCM) to perform high-resolution simulations. This method can capture the impact of local forcings such as topography, land-sea distribution, and land use on climate, and provide more refined climate information than the global model.

[0023] S2: Using initial field and model boundary field information to drive regional climate model dynamic downscaling includes: (1) performing dynamic downscaling integration on global climate initial field and model boundary field; (2) calling the regional climate model executable file and starting the model integration step; (3) submitting the regional climate model run to the queue, starting dynamic downscaling, and obtaining the integrated regional climate model dynamic downscaling prediction data product.

[0024] The Regional Climate Model (RCM) used is a climate-scale extension of the World Weather Prediction Model (WRF). Based on the WRF model, a nudging approach is employed for the lateral boundaries. Since the RCM only covers a limited area, its lateral boundaries require the acquisition of driving fields from global models or reanalysis data. Nudging is a boundary processing technique that gradually "pulls" variables within the model (such as temperature and wind field) toward the driving field during model integration, avoiding systematic biases that are inconsistent with the driving field. This ensures that the RCM simulation results are consistent with large-scale global circulation. In addition, the physical process parameterization schemes were debugged and integrated, including convection, microphysical processes, radiative transfer, boundary layer, and land surface processes. These parameterization schemes need to be coordinated to avoid model bias caused by mismatches between different physical processes. The surface information construction of this extended climate-scale regional model consists of two parts: (1) constructing reliable underlying surface information from high-resolution observational data, including static topography and soil, and dynamic vegetation, lakes, glaciers, snow cover, and water temperature; (2) adjusting the underlying surface dynamic parameterization scheme (including surface albedo, roughness, etc.) at the corresponding scale according to the 15-kilometer horizontal resolution to improve the simulation accuracy of variables. The regional climate model outputs Lambert projection grid data with a horizontal resolution of 15 kilometers after integration, including 462 grid points in the east-west direction and 342 grid points in the north-south direction. The center of the data grid is located at 110.34°E, 35.45°N.

[0025] Dynamic downscaling integration refers to the total time elapsed from the initial moment to the end of a regional climate model's (RCM) dynamic downscaling simulation. This integration time is typically measured in hours, days, or years and is an important indicator of regional climate simulation capabilities. In the WRF model, the main executable files include: performing dynamic downscaling integration on the global climate initial field and model boundary field by executing `make_exp_gcmsst.csh`; calling the regional climate model executable file `run_MMDD_yr-case.sh` to initiate the model integration step; and submitting the regional climate model run to the queue via `submit_intel.sh` to start the regional climate simulation calculation.

[0026] S3: Generating the input field required for machine learning downscaling based on the circulation field information output by the regional climate model: This process mainly includes systematic data post-processing of the dynamic downscaling prediction results of the regional climate model, that is, outputting circulation field information, involving 19 key meteorological variable channels, namely: geopotential height field and zonal wind field at four levels of 200 hPa, 500 hPa, 700 hPa and 850 hPa; relative humidity field, temperature field and meridional wind field at 500 hPa, 700 hPa and 850 hPa; and near-surface 2-meter temperature field and surface pressure field.

[0027] The circulation field information output by the regional climate model refers to the three-dimensional meteorological element field describing the atmospheric circulation characteristics output during the simulation process of the regional climate model (RCM). It mainly includes the following key variables: (1) Basic circulation field wind field: including horizontal wind field (U and V components) and vertical velocity (W), used to describe the atmospheric motion state; (2) Geopotential height field: describing the distribution of atmospheric pressure field, reflecting the configuration and intensity of high-pressure and low-pressure systems; (3) Temperature field: including atmospheric temperature, potential temperature, pseudo-equivalent potential temperature, etc., reflecting the thermal state and stability of the atmosphere; (4) Humidity field: specific humidity / relative humidity: describing the atmospheric water vapor content, which directly affects the formation of clouds and precipitation.

[0028] The data post-processing workflow encompasses two key steps: 1) Projection system conversion: converting the raw data from the Lambert projection coordinate system commonly used in regional climate models to an isotropic grid coordinate system. This is the process of converting planar projected coordinates to geographic latitude and longitude coordinates, executed in the file circu.CWRF.step1.grid_transform.py; 2) Data standardization: considering annual cycles and spatial variations, calculating the mean and standard deviation for each date, each grid point, and each variable, executed in the file circu.CWRF.step2.make_standard.py for normalization to eliminate the influence of dimensions and improve the training efficiency and stability of machine learning models. Its main function is to process the generated data after dynamic downscaling integration of regional climate models, forming the data format required for convolutional correction models. S4: Machine learning-based downscaling correction optimization of the dynamic downscaled output circulation field based on a convolutional model: The convolutional model structure designed for the nonlinear relationship between circulation and precipitation using CRA-40 data is as follows: Figure 2 As shown, the model takes a multi-channel two-dimensional circulation field (dimensions 128×128×19) as input, including 19 channel variables: geopotential height fields and zonal wind fields at 200 hPa, 500 hPa, 700 hPa, and 850 hPa; relative humidity fields, temperature fields, and meridional wind fields at 500 hPa, 700 hPa, and 850 hPa; and temperature and surface pressure fields at 2 m. In addition, the model uses month and date as additional input information.

[0029] exist Figure 2 In this model, the encoder-decoder architecture supports multi-scale extraction and spatial mapping. The encoder consists of six downsampling modules: each block includes two 3×3 convolutional layers (ReLU activation), followed by 2×2 max pooling, with the number of channels gradually doubling (32→64→128→256→512→1024). This hierarchical structure captures features from local cyclic details to large-scale cyclic patterns. The decoder uses transposed convolutions for upsampling to restore spatial resolution. Features from the corresponding encoder layers are connected via skip connections, effectively integrating fine-scale circulation details (such as low-level wind structure) with deep semantic information (such as the Western Pacific subtropical high). The output layer uses 1×1 convolutions to generate single-channel precipitation predictions. The model employs mean squared error (MSE) as a loss function to constrain the spatial distribution of precipitation, thereby enhancing the modeling of the physical connection between circulation dynamics and precipitation patterns.

[0030] S5: Machine learning super-resolution is performed on the output data after downscaling correction. Based on the correction, the Residual Channel Attention Network (RCAN) machine learning super-resolution algorithm is adopted. RCAN is a deep convolutional neural network model for image super-resolution (SR). Through its core residual-in-residual (RIR) structure and channel attention (CA) mechanism, it can effectively learn the complex nonlinear mapping between low-resolution and high-resolution fields, significantly improving the performance of image super-resolution while maintaining the network depth. Figure 3 The image shows a random sampling output example of the Residual Channel Attention Network (RCAN) super-resolution model. The high-resolution input data for the model uses CLDAS, with a horizontal range of 0-65°N and 60-160°E, and a resolution of 0.0625° × 0.0625°. The Residual Channel Attention Network (RCAN) super-resolution model comprises four modules: shallow feature extraction, deep feature extraction, feature fusion, and sampling reconstruction.

[0031] 1) The shallow feature extraction module initially captures the spatial features of the precipitation field through convolutional layers, and the calculation formula is SF = f_sf(X_LR); where X_LR is the multivariate, multi-time circulation field, specifically including Batch (sample batch size), Time (time series length), Height / Lat (network latitude dimension), Width / Lon (network longitude dimension), and Channels / Features (number of feature / variable channels); f_sf is the shallow feature extraction function, which is a learnable function mapping defined by the neural network layer; SF is the output of the module.

[0032] 2) The deep feature extraction module consists of stacked residual groups, each containing a residual block (RCAB) with channel attention mechanism. The calculation formula is RCAB_out = RCAB_in + (Conv(σ(fc(GAP(Conv(RCAB_in)))) ⊙ Conv(RCAB_in)) Where RCAB_in is the input feature map, RCAB_out is the feature output map, GAP() is global average pooling to obtain channel statistics; fc() is a fully connected layer to generate channel attention weights; σ() is the Sigmoid activation function to normalize the weights to [0,1]; ⊙ is channel-wise multiplication to recalibrate the features.

[0033] 3) Shallow and deep features are fused using the formula DF = SF + f_skip(RGs_out). Here, DF represents the fused deep features, combining the global context of the shallow layer with the local details of the deep layer; RGs_out is the residual set output, the final output of multiple stacked deep feature extraction modules; and f_skip is the skip connection processing function, used to process and transform RGs_out so that it can be effectively fused with SF. By focusing on subtle differences through residual learning and utilizing an attention mechanism to assign weights to different feature channels (such as precipitation of different intensities), long-range dependencies are efficiently modeled, enhancing the ability to capture key meteorological information.

[0034] 4) The feature map size is enlarged by the upsampling module, and the final high-resolution result (Y_HR) is output by the reconstruction module. The formula is Y_HR = f_recon(DF), where f_recon is the reconstruction function, which is a learnable function mapping. The RCAN super-resolution model can interpolate the sub-seasonal machine learning correction data of regional climate models to a high-resolution grid of approximately 6.25 km.

[0035] S6: Generate dual downscaling sub-seasonal forecast information using dynamical models and machine learning: Outputs the final high-resolution meteorological element field with richer physical details, forming dual downscaling sub-seasonal forecast information using dynamical models and machine learning. Taking 30-day precipitation forecast data starting from July 27, 2024 as an example, the dual downscaling forecast data storage format is CWRF.tp.anomperc.HR.s20240727.w3.nc, with a single data file size of approximately 65.22MB. The data format is as follows... Figure 4 As shown.

[0036] This invention fully leverages the advantages of dynamic models and deep learning, effectively integrating the two to improve the accuracy of sub-seasonal forecasts, particularly significantly enhancing the forecasting performance of heavy rainfall in the Jiangnan region. Figure 5 As shown, it can be effectively applied to the sub-seasonal forecasting of meteorological elements.

[0037] This invention fully leverages the advantages of dynamical models and deep learning, effectively integrating the two. Deep learning has demonstrated efficient characterization capabilities for nonlinear climate systems in climate prediction research. However, given the limited availability of sub-seasonal climate data, it is necessary to prioritize the in-depth mining and integration of prior human knowledge to reduce uncertainty in our understanding of the climate system and ensure that model predictions locally conform to physical mechanisms. Secondly, data augmentation is required to enhance the model's generalization performance. Dynamical downscaling techniques are employed, refining local information through regional climate models and combining it with large-scale modal information from global climate models to provide a more localized dataset for the deep learning model. By utilizing the complementary advantages of dynamical downscaling and deep learning downscaling, sub-seasonal prediction techniques are improved to address the new challenges posed by climate change.

[0038] This invention combines dynamic downscaling with refined local influencing factors to construct a deep learning downscaling model based on physical mechanism constraints and local data-driven approaches. To meet the "accurate forecasting" requirements of modernizing meteorological science and technology capabilities and social services, a dual downscaling method of dynamic downscaling and deep learning downscaling is employed to overcome the bottleneck in sub-seasonal climate prediction.

[0039] Dynamic downscaling based on regional climate models has become an effective way to provide gridded high-resolution meteorological data, which can better describe the detailed underlying surface features and small- and medium-scale physical processes in my country. While emphasizing high resolution, it is also necessary to consider various physical process parameterization schemes and interactions, which are increasingly important for the sub-grid processes in high-resolution models. Failure to improve parameterization schemes can lead to significant simulation errors. This study adopts a climate-scale extension of the internationally widely used mesoscale weather forecasting model (WRF). Based on the WRF model, adjustments and integrations were made to the lateral boundary relaxation approximation scheme, physical process parameterization schemes, etc., including convection parameterization schemes, microphysical process parameterization schemes, radiative transfer parameterization schemes, boundary layer parameterization schemes, and land surface process parameterization schemes. Based on dynamic downscaling, and by integrating local refined forecast information from regional models, a convolutional neural network deep learning downscaling correction model is driven and super-resolution is applied. A dual downscaling prediction method is used to correct climate model forecast information, achieving synergistic optimization of physical process constraints and deep learning corrections. The study focuses on addressing the problem of insufficient representation of nonlinear interactions of influencing factors in dynamic models, leveraging the advantages of the dual downscaling method in sub-seasonal scale predictability.

[0040] The sequence numbers of the above embodiments of the present invention are merely for descriptive purposes and do not represent the superiority or inferiority of the embodiments. Through the above description of the embodiments, those skilled in the art can clearly understand that the above implementation methods can be implemented using software plus necessary general-purpose hardware platforms. Of course, they can also be implemented using hardware, but in many cases, the former is a better implementation method. Based on this understanding, the technical solution of the present invention, 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 is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes several instructions to cause a terminal (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0041] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A subseasonal climate prediction method based on a combination of dynamic model downscaling and machine learning downscaling, characterized in that, The method includes the following steps: S1 generates the initial field and model boundary field information required for dynamic downscaling based on global climate model output data; S2 uses the initial field and model boundary field information to drive the regional climate model for dynamic downscaling. S3 generates the input field required for machine learning downscaling based on the circulation field information output by the regional climate model; S4 performs machine learning downscaling correction optimization on the dynamic downscaled output circulation field based on a convolutional model; S5 performs machine learning super-resolution based on downscaling correction of output data using machine learning; S6 generates sub-seasonal precipitation forecasts using a combination of dynamical models and machine learning to downscale the precipitation data.

2. The sub-seasonal climate prediction method according to claim 1, characterized in that, S1 includes the following steps: S11 Global Climate Model Output Information Collection: Real-time monitoring and collection of atmospheric, oceanic, and land surface variable data from global model prediction integrals, ensuring the data input required for dynamic downscaling. S12 global model data preprocessing processes real-time collected global model atmospheric, oceanic, and land surface data to generate regional model initial and boundary fields, obtaining the data format required for regional climate model dynamic downscaling prediction integration.

3. The sub-seasonal climate prediction method according to claim 2, characterized in that, The initial field and model boundary field include: initial field information including atmospheric initial field and surface initial field information, and model boundary field information including lateral boundary conditions and lower boundary conditions.

4. The sub-seasonal climate prediction method according to claim 1, characterized in that, S2 includes the following steps: S21 performs dynamic downscaling integration on the global climate initial field and model boundary field; S22 calls the regional climate model executable file and initiates the model integration step; S23 submits the regional climate model operation to the queue, begins dynamic downscaling, and obtains the integrated regional climate model dynamic downscaling prediction data product.

5. The sub-seasonal climate prediction method according to claim 4, characterized in that, The regional climate model mentioned in step S22 is a climate-scale extension of the mesoscale weather forecast model. It adopts a side-boundary relaxation approximation scheme based on the mesoscale weather forecast model and adjusts and integrates the physical process parameterization scheme. The physical process parameterization scheme includes convection parameterization, microphysical process parameterization, radiation transfer parameterization, boundary layer parameterization, and land surface process parameterization scheme.

6. The sub-seasonal climate prediction method according to claim 1, characterized in that, Step S3 involves systematically post-processing the dynamic downscaling prediction results of regional climate models to output circulation field information. The post-processing includes projection system conversion and data standardization.

7. The sub-seasonal climate prediction method according to claim 6, characterized in that, The circulation field information includes 19 key meteorological variable channels: geopotential height field and zonal wind field at four levels: 200 hPa, 500 hPa, 700 hPa and 850 hPa; relative humidity field, temperature field and meridional wind field at 500 hPa, 700 hPa and 850 hPa; and near-surface temperature field and surface pressure field at 2 meters.

8. The sub-seasonal climate prediction method according to claim 1, characterized in that, The convolutional model structure in S4 is designed using CRA-40 data to represent the nonlinear relationship between circulation and precipitation.

9. The sub-seasonal climate prediction method according to claim 1, characterized in that, The machine learning super-resolution in step S5 adopts the residual channel attention network machine learning super-resolution model, which includes four modules: shallow feature extraction, deep feature extraction, feature fusion, and sampling reconstruction.

10. A subseasonal climate prediction system based on a combination of dynamic model downscaling and machine learning downscaling, characterized in that, The system includes the following modules: An initial information generation module is used to generate initial field and model boundary field information required for dynamic downscaling based on global climate model output data. A dynamic downscaling module, which uses the initial field and model boundary field information to drive the regional climate model to perform dynamic downscaling; An input information generation module generates the input field required for machine learning downscaling based on the circulation field information output by the regional climate model. A machine learning downscaling module, which performs machine learning downscaling correction and optimization on the dynamic downscaling output circulation field based on a convolutional model; A machine learning super-resolution module, which performs machine learning super-resolution based on machine learning downscaling correction output data; A prediction information generation module generates sub-seasonal prediction information that combines dynamic patterns and machine learning for dual downscaling.

Images