Multi-scale cloud system dynamic evolution simulation modeling method and system based on digital twin

Abstract

The present invention provides a multi-scale cloud system dynamic evolution simulation modeling method and system based on digital twin, which relates to the field of digital twin technology. It includes obtaining multi-dimensional data of the cloud system through a multi-source collaborative perception network, extracting the primitive features of the cloud system and dividing the cloud cluster regions by using a recursive segmentation algorithm after data registration and quality assessment; then, constructing a deep digital twin mapping network between the primitive features of the cloud system and the real cloud system, and performing feature verification and parameter optimization; finally, performing adaptive multi-grid discretization on the cloud system, establishing cross-region coupling constraints including conservation laws, characterizing the dynamic evolution process of the cloud system as a partial differential equation set, and integrating and solving it with the deep digital twin mapping network, and dynamically correcting the evolution prediction results by using real-time digital twin data.

Description

Technical Field

[0001] The present invention relates to digital twin technology, and particularly to a multi-scale cloud system dynamic evolution simulation modeling method and system based on digital twin. Background Art

[0002] Existing cloud system evolution prediction technologies mainly rely on single sensors to obtain cloud system feature information, making it difficult to comprehensively reflect the complex structure and dynamic change characteristics of cloud systems. In the process of multi-source data fusion, there are problems such as insufficient spatio-temporal registration accuracy and lack of data quality assessment mechanisms, which affect the accuracy of cloud system feature extraction.

[0003] Currently, cloud system boundary division methods lack clear physical bases and are difficult to accurately distinguish the differential characteristics between the core area and the transition area. At the same time, traditional feature extraction algorithms cannot effectively adapt to the dynamic change process of cloud system morphology and physical properties, resulting in inaccurate description of cloud system evolution characteristics. In numerical simulation, existing methods adopt fixed grid division strategies, making it difficult to specifically depict the evolution characteristics of different regions, and insufficient consideration of the coupling effect between regions, making it difficult to balance calculation efficiency and accuracy.

[0004] In addition, existing cloud system evolution prediction methods generally lack a feedback mechanism for real-time observation data, and there are large deviations between prediction results and actual situations. Physical constraint conditions are not fully considered in the prediction correction process, making it difficult to ensure the physical rationality of correction results. To solve the above problems, there is an urgent need to establish a cloud system evolution prediction method that can comprehensively utilize multi-source data, accurately extract cloud system features, efficiently perform numerical simulation, and achieve dynamic correction. Summary of the Invention

[0005] Embodiments of the present invention provide a multi-scale cloud system dynamic evolution simulation modeling method and system based on digital twin, which can solve the problems in the prior art.

[0006] In the first aspect of the embodiments of the present invention,

[0007] A multi-scale cloud system dynamic evolution simulation modeling method based on digital twin is provided, including:

[0008] Construct a multi-source collaborative perception network using a multi-spectral camera, lidar, and millimeter-wave radar to obtain near-ground data of the cloud system, and combine satellite remote sensors to obtain top data of the cloud system and ground meteorological stations to obtain meteorological element data; perform phase consistency analysis on the obtained multi-source data, establish a spatial transformation matrix between sensor coordinate systems to obtain registered data, and perform quality assessment on the registered data according to signal-to-noise ratio, spatial resolution, and time consistency, and screen out valid data that meet the assessment threshold;

[0009] Based on the valid data, a recursive segmentation algorithm is used to extract the primitive features of cloud systems. The extreme points of the reflectivity gradient and temperature gradient are calculated to divide the core region of the cloud cluster, and the parameter attenuation curve around the core region is calculated to divide the transition region. Based on the region division results, the contour curvature, gray distribution, texture features, and multi-scale edge features of the cloud system are extracted, and a deep digital twin mapping network between the primitive features of the cloud system and the real cloud system is constructed. The morphological feature parameters output by the deep digital twin mapping network are verified with meteorological measured data, and the parameters of the deep digital twin mapping network are dynamically optimized based on the verification results.

[0010] The core region is refined and discretized using an adaptive multi-grid method, and the grid scale is automatically adjusted according to the cloud cluster turbulence intensity and thermal parameter gradient. A non-uniform grid considering the internal vortex evolution of the cloud system is established for the transition region. Based on the grid division results, a simulation model of the dynamic evolution of the cloud system is established, and cross-region coupling constraints including mass conservation, momentum conservation, and energy conservation are established. The dynamic evolution process of the cloud system is characterized as a partial differential equation system including turbulent diffusion, thermodynamic parameters, and regional coupling. The partial differential equation system is integrated with the deep digital twin mapping network, and a domain decomposition method is used to solve the overall evolution characteristics of the cloud system, and the real-time digital twin data is used to dynamically correct the evolution prediction results.

[0011] In an optional implementation manner,

[0012] The steps of using a recursive segmentation algorithm to extract the primitive features of cloud systems based on the valid data, calculating the extreme points of the reflectivity gradient and temperature gradient to divide the core region of the cloud cluster, and calculating the parameter attenuation curve around the core region to divide the transition region include:

[0013] A recursive segmentation algorithm is used to extract the primitive features of cloud systems. The reflectivity mean, reflectivity standard deviation, temperature mean, and temperature standard deviation are calculated for each sub-region. When the reflectivity standard deviation or temperature standard deviation of a sub-region is greater than the preset segmentation threshold, the region is continuously segmented until the reflectivity standard deviation and temperature standard deviation are less than the preset segmentation threshold or reach the minimum segmentation scale. The feature vectors of the primitive regions are recorded to form a primitive feature set.

[0014] Based on the primitive feature set, the reflectivity gradient and temperature gradient between adjacent primitives are calculated. The dynamic weight coefficients of reflectivity and temperature are calculated according to the variances of the reflectivity gradient and temperature gradient. The reflectivity gradient, temperature gradient, and their cross terms are weighted and fused to obtain a comprehensive gradient field.

[0015] Perform multi-scale decomposition on the comprehensive gradient field using the Gaussian function, extract local maximum points for the gradient field at each scale, establish a correspondence between the local maximum points at different scales to form an extreme point chain; use the continuous length of the extreme point chain in the scale space as a stability index, and select the extreme point with the highest stability index as a candidate point for the core region; perform spatial connectivity analysis on the candidate points, and define the set of candidate points with an Euclidean space distance less than the set minimum distance threshold as the core region;

[0016] Calculate the horizontal wind field gradient outside the core region, calculate the local turbulence intensity based on the horizontal wind field gradient, and combine the local turbulence intensity with the reflectivity and temperature to obtain a corrected parameter value; divide the radial distance from the core region to the outside into multiple intervals, use a double exponential function to fit the parameter decay curve in each interval according to the corrected parameter value, and impose a derivative continuity constraint at the interval connection; calculate the second derivative and the third derivative of the parameter decay curve, and use the weighted sum of the absolute value of the second derivative and the integral of the third derivative as the boundary recognition objective function, and solve for the position that minimizes the boundary recognition objective function as the boundary of the transition region.

[0017] In an optional implementation manner,

[0018] The steps of constructing a deep digital twin mapping network for the cloud element characteristics and the real cloud system, verifying the morphological feature parameters output by the deep digital twin mapping network with meteorological measured data, and dynamically optimizing the parameters of the deep digital twin mapping network based on the verification results include:

[0019] Input the contour curvature, gray level distribution, texture features and multi-scale edge features of the cloud system into the deep digital twin mapping network for feature processing. The deep digital twin mapping network includes a feature encoder, a hierarchical dynamic memory module for storing cloud system morphological patterns and a multi-task decoder; the feature encoder encodes the features based on an adaptive sampling strategy for cloud system morphological complexity to obtain encoded features, and the cloud system morphological complexity is calculated by the weighted combination of the contour curvature variance and the texture entropy;

[0020] The hierarchical dynamic memory module updates memory units through an attention mechanism and a gated update method that introduces physical constraints such as temperature gradients and pressure fields, dynamically fusing memory states and encoding features to obtain fused features. Based on the fused features, a multi-granularity spatiotemporal attention mechanism is used to process cloud morphological dependencies, capturing spatial dependency features through adaptive receptive field adjustment based on cloud center distance, and introducing recursive attention units to model temporal dependencies, resulting in a feature representation with spatiotemporal dependencies. The feature representation is input into a multi-task decoder for morphological parameter prediction and reconstruction, and a joint loss function comprising task loss, physical constraint loss, memory update loss, and temporal smoothing loss is constructed to optimize the prediction results. The weight coefficient of the joint loss function is dynamically adjusted according to the prediction error.

[0021] The prediction results output by the multi-task decoder are accurately evaluated based on the shape similarity of the Hausdorff distance, and are physically verified through area-perimeter ratio constraints and neighborhood consistency constraints, wherein the area-perimeter ratio constraint limits the value range of the shape coefficient, and the neighborhood consistency constraint ensures that the changes in the morphological parameters at adjacent moments conform to the laws of meteorological development, and obtains verification performance indicators; based on the verification performance indicators, a dynamic learning strategy is adopted to optimize the deep digital twin mapping network, wherein the learning rate is dynamically adjusted based on a periodic attenuation function, and the sample weight coefficient is calculated based on the prediction error, and the sample weight coefficient is calculated by the exponential function of the prediction error and the temperature parameter, and the network parameters are updated after the sample weight coefficient is multiplied by the task loss, physical constraint loss, memory update loss and temporal smoothing loss respectively.

[0022] In an optional embodiment,

[0023] The core region is discretized using an adaptive multigrid method. The grid size is automatically adjusted according to the cloud turbulence intensity and thermal parameter gradient. The steps to establish a non-uniform grid in the transition region that takes into account the evolution of vortices within the cloud system include:

[0024] The turbulent kinetic energy spectrum and vorticity field distribution in the core area are calculated based on meteorological radar reflectivity and Doppler velocity field data. The turbulence intensity distribution at different scales is obtained by analyzing the turbulent kinetic energy spectrum. The vortex structure in the core area is identified using the vortex threshold method, and the vortex center position and vortex intensity are extracted.

[0025] The temperature field and pressure field of the cloud area are constructed using meteorological sounding data, the temperature gradient and pressure gradient are calculated, and the absolute value of the temperature gradient and the absolute value of the pressure gradient are weightedly combined to obtain a thermal parameter gradient intensity index; the turbulence intensity distribution at different scales and the thermal parameter gradient intensity index are input into an exponential function, and a grid scale mapping function is obtained through weighted combination;

[0026] A quadtree structure is used for multi-grid partitioning. The grid cells are judged for subdivision according to the ratio of the output value of the grid scale mapping function to the initial grid scale. The grid cells that meet the subdivision threshold condition are divided into four sub-grids, and the difference in adjacent grid levels is restricted.

[0027] Based on the vortex center position and vortex intensity, an evolution relationship of the vortex intensity is constructed. The grid density attenuation coefficient of the transition region is calculated through the hyperbolic tangent function according to the distance from the vortex center. The grid density attenuation coefficient increases with the increase of the distance to the vortex center.

[0028] The grid shape quality, orthogonality index, and feature capture ability are weighted and combined to obtain a grid quality evaluation index. When the grid quality evaluation index is lower than the preset threshold, the grid cells to be optimized are determined. The position information of the grid cells to be optimized is input into the grid scale mapping function to recalculate the local grid scale. The grid topology structure is updated based on the recalculated grid scale, and the physical field is interpolated and reconstructed.

[0029] In an optional implementation manner,

[0030] Based on the grid partitioning result, a simulation model of the dynamic evolution of the cloud system is established. Cross-region coupling constraints including mass conservation, momentum conservation, and energy conservation are established. The dynamic evolution process of the cloud system is characterized as a partial differential equation system including turbulent diffusion, thermodynamic parameters, and region coupling. Integrating the partial differential equation system with a deep digital twin mapping network, using the domain decomposition method to solve the overall evolution characteristics of the cloud system, and dynamically correcting the evolution prediction results using real-time digital twin data includes the following steps:

[0031] Cross-region coupling constraints of mass conservation, momentum conservation, and energy conservation are established on the interface between the core region and the transition region. The mass conservation constraint is based on the flux continuity conditions of the density field and velocity field. The momentum conservation constraint is based on pressure and viscous stress tensors. The energy conservation constraint is based on specific internal energy, thermal conductivity, and temperature field.

[0032] A partial differential equation system for the evolution of the cloud system including turbulent diffusion terms, thermodynamic parameter terms, and region coupling terms is constructed. The turbulent diffusion terms are characterized by turbulent kinetic energy, turbulent diffusion coefficient, and dissipation rate. The thermodynamic parameter terms are characterized by temperature field, pressure field, and density field. The region coupling terms are determined by interface flux balance. The partial differential equation system correlates the cross-region coupling constraints with the turbulent diffusion terms, thermodynamic parameter terms, and region coupling terms.

[0033] Input the partial differential equation system into the deep digital twin mapping network, establish a mapping loss function including a physical constraint loss term and a conservation constraint loss term, and optimize the parameters of the deep digital twin mapping network by minimizing the mapping loss function;

[0034] Use the domain decomposition method to solve the output result of the deep digital twin mapping network. Based on the distribution characteristics of the turbulent diffusion term and the thermodynamic parameter term, fix the physical parameters in the transition region to solve the core region, and use the updated physical parameters in the core region as boundary conditions to solve the physical parameters in the transition region. Iterate cyclically until the physical parameters in the core region and the transition region reach the preset convergence conditions. The physical parameters include velocity field, density field, turbulent kinetic energy, turbulent diffusion coefficient, dissipation rate, temperature field, pressure field, specific internal energy, and thermal conductivity;

[0035] Receive the near-surface data and the top data of the cloud system as real-time digital twin data. Construct a Kalman gain matrix based on the prediction error covariance of the physical parameters in the core region and the transition region and the observation error covariance of the real-time digital twin data. Conduct an error analysis on the difference between the real-time digital twin data and the solution result of the domain decomposition method, and use the Kalman gain matrix to dynamically correct the cloud system evolution prediction result.

[0036] In an optional implementation manner,

[0037] The steps of inputting the partial differential equation system into the deep digital twin mapping network, establishing a mapping loss function including a physical constraint loss term and a conservation constraint loss term, and optimizing the parameters of the deep digital twin mapping network by minimizing the mapping loss function include:

[0038] The physical constraint loss term includes a turbulent dynamics loss and a thermodynamic coupling loss. The turbulent dynamics loss is constructed based on the spatio-temporal derivatives of the turbulent kinetic energy and the dissipation rate and the turbulent production term. The weight coefficient of the turbulent dynamics loss is adaptively adjusted by the local turbulent intensity and the turbulent Reynolds number. The thermodynamic coupling loss is constructed based on the convection-diffusion equations of the temperature field and the pressure field. The weight coefficient of the thermodynamic coupling loss is dynamically determined by a non-linear function of the temperature gradient and the pressure gradient;

[0039] Construct the conservation constraint loss term based on the cross-region coupling constraint. The conservation constraint loss term includes a mass conservation constraint, a momentum conservation constraint, and an energy conservation constraint, and introduce a stress tensor correction term based on vorticity correlation;

[0040] Construct a feature-driven combined loss function, which combines the physical constraint loss term and the conservation constraint loss term with weights. The weight coefficients of the combined loss function are dynamically adjusted by the cloud system morphology complexity, temperature gradient, and pressure gradient. The cloud system morphology complexity is calculated through the weighted combination of the contour curvature variance and texture entropy;

[0041] Introduce a multi-scale residual correction mechanism, which includes local residuals and global residuals. The local residuals are calculated based on turbulence characteristic quantities, and the global residuals are calculated based on thermodynamic quantities. The weight coefficients of the local residuals and global residuals are dynamically determined by the ratio of the turbulence characteristic scale to the cloud system characteristic scale; Optimize and update the parameters of the deep digital twin mapping network by minimizing the combined loss function and combining the multi-scale residual correction results. The learning rate is adaptively adjusted based on the physical quantity gradients and prediction errors. The physical quantities include velocity field, vorticity field, pressure field, density field, temperature field, turbulent kinetic energy, and dissipation rate. At the same time, introduce a periodic decay mechanism. At the end of each training cycle, the learning rate is adjusted stage by stage according to the satisfaction degrees of the physical constraint loss term and the conservation constraint loss term.

[0042] In an optional implementation manner,

[0043] The steps of receiving the near-surface data and the top data of the cloud system as real-time digital twin data, constructing a Kalman gain matrix based on the prediction error covariance of the physical parameters in the core area and the transition area and the observation error covariance of the real-time digital twin data, performing error analysis on the real-time digital twin data and the solution results of the regional decomposition method, and dynamically correcting the cloud system evolution prediction results using the Kalman gain matrix include:

[0044] Obtain the near-surface data of the cloud system including reflectivity, radial velocity, and spectral width, fuse the multi-site radar data using the three-dimensional variational assimilation method, construct the distribution of the three-dimensional wind field, temperature field, and water vapor field in the near-surface layer, and determine the ground-based observation error covariance matrix based on historical observation statistics; Obtain the top data of the cloud system including cloud top height, cloud top temperature, and cloud top radiation flux, extract the cloud system characteristic parameters using the infrared-visible two-channel inversion algorithm, and obtain the cloud system motion information through multi-temporal sequence registration. Determine the satellite observation error covariance matrix through channel noise analysis;

[0045] Construct a prediction error covariance matrix based on the difference prediction error between the prediction results and the corresponding true values of the physical parameters in the core area and the transition area, and introduce a local correlation constraint based on the physical parameter spatial distance to correct the prediction error covariance matrix. The local correlation constraint is determined by the correlation length scale;

[0046] Construct a Kalman gain matrix based on the predicted error covariance matrix, the ground-based observation error covariance matrix, and the satellite observation error covariance matrix, and introduce an adaptive adjustment mechanism based on the cloud system morphology complexity to correct the Kalman gain matrix. The cloud system morphology complexity is calculated through the weighted combination of the contour curvature variance and the texture entropy;

[0047] Map the near-ground data and the top data of the cloud system to the observation space, calculate the observation residual vector with the predicted value, and update the predicted state based on the observation residual vector and the corrected Kalman gain matrix; introduce the turbulence feature constraint, the mass conservation constraint, and the energy conservation constraint to the updated predicted state. The turbulence feature constraint is constructed based on the non-negativity of the turbulent kinetic energy and the dissipation rate; use the constrained predicted state as the initial condition for the next moment prediction to form a dynamically corrected closed-loop iterative process.

[0048] In the second aspect of the embodiments of the present invention,

[0049] Provide a multi-scale cloud system dynamic evolution simulation modeling system based on digital twin, including:

[0050] The first unit is used to construct a multi-source collaborative perception network using a multi-spectral camera, lidar, and millimeter-wave radar to obtain near-ground data of the cloud system, and combine satellite remote sensors to obtain top data of the cloud system and ground meteorological stations to obtain meteorological element data; perform phase consistency analysis on the obtained multi-source data, establish a spatial transformation matrix between sensor coordinate systems to obtain registered data, and perform quality assessment on the registered data according to signal-to-noise ratio, spatial resolution, and time consistency, and screen out valid data that meet the assessment threshold;

[0051] The second unit is used to extract the cloud system primitive features based on the valid data, calculate the extreme points of the reflectivity gradient and the temperature gradient to divide the core area of the cloud cluster, and calculate the parameter attenuation curve around the core area to divide the transition area; based on the area division result, extract the contour curvature, gray distribution, texture features, and multi-scale edge features of the cloud system, construct a deep digital twin mapping network between the cloud system primitive features and the real cloud system, verify the morphological feature parameters output by the deep digital twin mapping network with meteorological measured data, and dynamically optimize the parameters of the deep digital twin mapping network based on the verification results;

[0052] The third unit is used to perform refined discretization on the core area using an adaptive multi-grid method. The grid scale is automatically adjusted according to the cloud cluster turbulence intensity and the gradient of thermodynamic parameters. A non-uniform grid considering the evolution of internal vortices in the cloud system is established for the transition area. Based on the grid division results, a simulation model of the dynamic evolution of the cloud system is built, and cross-region coupling constraints including mass conservation, momentum conservation, and energy conservation are established. The dynamic evolution process of the cloud system is characterized as a partial differential equation system including turbulent diffusion, thermodynamic parameters, and regional coupling. The partial differential equation system is integrated with a deep digital twin mapping network, and a domain decomposition method is used to solve the overall evolution characteristics of the cloud system, and real-time digital twin data is used to dynamically correct the evolution prediction results.

[0053] In the third aspect of the embodiments of the present invention,

[0054] a kind of electronic device is provided, including:

[0055] a processor;

[0056] a memory for storing instructions executable by the processor;

[0057] wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0058] In the fourth aspect of the embodiments of the present invention,

[0059] a computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0060] The present invention uses a multi-source collaborative sensing network to obtain comprehensive cloud system data, and extracts the basic characteristics of the cloud system through a deep digital twin mapping network, realizing the effective fusion of the morphological characteristics of the cloud system and meteorological measured data, thereby improving the accuracy of cloud system simulation modeling. The refined grid division strategy and cross-region coupling constraints further enhance the simulation ability of the dynamic evolution of the cloud system.

[0061] The present invention characterizes the dynamic evolution process of the cloud system as a partial differential equation system, and combines a deep digital twin mapping network and a domain decomposition method for solution, realizing the prediction of the overall evolution characteristics of the cloud system. Using real-time digital twin data to dynamically correct the prediction results further improves the accuracy and reliability of the prediction.

[0062] The present invention uses an adaptive multi-grid method to perform refined discretization on the core area, automatically adjusts the grid scale according to the cloud cluster turbulence intensity and the gradient of thermodynamic parameters, avoiding unnecessary waste of computing resources. The application of the domain decomposition method also improves the efficiency of solving the overall evolution characteristics of the cloud system. Description of the Drawings

[0063] Figure 1 This is a schematic flowchart of the method for dynamic evolution simulation modeling of multi-scale cloud systems based on digital twin in an embodiment of the present invention;

[0064] Figure 2 This is a schematic structural diagram of the system for dynamic evolution simulation modeling of multi-scale cloud systems based on digital twin in an embodiment of the present invention. Specific embodiments

[0065] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0066] The technical solutions of the present invention will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0067] Figure 1 This is a schematic flowchart of the method for dynamic evolution simulation modeling of multi-scale cloud systems based on digital twin in an embodiment of the present invention, as Figure 1 shown, the method includes:

[0068] S1. Construct a multi-source collaborative perception network using a multi-spectral camera, lidar, and millimeter-wave radar to obtain near-ground data of cloud systems, and combine satellite remote sensors to obtain top data of cloud systems and ground meteorological stations to obtain meteorological element data; perform phase consistency analysis on the obtained multi-source data, establish a spatial transformation matrix between sensor coordinate systems to obtain registered data, and perform quality assessment on the registered data according to signal-to-noise ratio, spatial resolution, and time consistency, and screen out valid data that meet the assessment threshold;

[0069] S2. Based on the valid data, use a recursive segmentation algorithm to extract the primitive features of cloud systems, calculate the extreme points of reflectivity gradient and temperature gradient to divide the core area of cloud clusters, and calculate the parameter attenuation curve around the core area to divide the transition area; based on the area division results, extract the contour curvature, gray distribution, texture features, and multi-scale edge features of cloud systems, construct a deep digital twin mapping network between the primitive features of cloud systems and real cloud systems, verify the morphological feature parameters output by the deep digital twin mapping network with meteorological measured data, and dynamically optimize the parameters of the deep digital twin mapping network based on the verification results;

[0070] S3. The core region is discretized in a refined manner using the adaptive multi-grid method, where the grid scale is automatically adjusted according to the cloud cluster turbulence intensity and the gradient of thermodynamic parameters. For the transition region, a non-uniform grid considering the internal vortex evolution of the cloud system is established. Based on the grid division results, a simulation model of the dynamic evolution of the cloud system is built, and cross-region coupling constraints including mass conservation, momentum conservation, and energy conservation are established. The dynamic evolution process of the cloud system is characterized as a partial differential equation system including turbulent diffusion, thermodynamic parameters, and regional coupling. The partial differential equation system is integrated with the deep digital twin mapping network, and the regional decomposition method is used to solve the overall evolution characteristics of the cloud system, and the real-time digital twin data is used to dynamically correct the evolution prediction results.

[0071] In an optional implementation manner,

[0072] Based on the valid data, the steps of extracting the primitive features of the cloud system using the recursive partitioning algorithm, calculating the extreme points of the reflectivity gradient and the temperature gradient to divide the core region of the cloud cluster, and calculating the parameter attenuation curve around the core region to divide the transition region include:

[0073] Use the recursive partitioning algorithm to extract the primitive features of the cloud system. Calculate the mean reflectivity, standard deviation of reflectivity, mean temperature, and standard deviation of temperature for each sub-region. When the standard deviation of reflectivity or the standard deviation of temperature in a sub-region is greater than the preset partitioning threshold, the region is continuously partitioned until the standard deviation of reflectivity and the standard deviation of temperature are less than the preset partitioning threshold or reach the minimum partitioning scale. Record the feature vectors of the primitive regions to form a primitive feature set;

[0074] Calculate the reflectivity gradient and temperature gradient between adjacent primitives based on the primitive feature set. Calculate the dynamic weight coefficients of reflectivity and temperature according to the variances of the reflectivity gradient and the temperature gradient. Weight and fuse the reflectivity gradient, temperature gradient, and their cross terms to obtain a comprehensive gradient field;

[0075] Perform multi-scale decomposition on the comprehensive gradient field using the Gaussian function. Extract the local maximum points for the gradient field at each scale. Establish a correspondence between the local maximum points at different scales to form an extreme point chain. Use the continuous length of the extreme point chain in the scale space as a stability index, and select the extreme point with the highest stability index as the candidate point for the core region. Conduct a spatial connectivity analysis on the candidate points, and define the set of candidate points with an Euclidean space distance less than the set minimum distance threshold as the core region;

[0076] Calculate the horizontal wind field gradient around the core region, calculate the local turbulence intensity based on the horizontal wind field gradient, and combine the local turbulence intensity with reflectivity and temperature to obtain a corrected parameter value; divide the radial distance from the core region to the periphery into multiple intervals, and use a double exponential function to fit the parameter attenuation curve in each interval, and apply a derivative continuity constraint at the connection of intervals; calculate the second derivative and the third derivative of the parameter attenuation curve, and use the weighted sum of the absolute value of the second derivative and the integral of the third derivative as the boundary recognition objective function, and solve for the position that minimizes the boundary recognition objective function as the transition region boundary.

[0077] Exemplarily, first, extract the primitive features of the cloud system. Input the cloud system data to be processed, such as satellite cloud map data. Set the segmentation thresholds for reflectivity and temperature, as well as the minimum segmentation scale of the region. Recursively segment the cloud system data: calculate the mean reflectivity, reflectivity standard deviation, mean temperature, and temperature standard deviation of each sub-region. If the reflectivity standard deviation or temperature standard deviation of the sub-region is greater than the preset segmentation threshold, and the scale of the sub-region is greater than the minimum segmentation scale, then continue to segment the sub-region. Otherwise, stop segmentation and record the feature vector of the sub-region (such as mean reflectivity, reflectivity standard deviation, mean temperature, temperature standard deviation, region area, region position, etc.) into the primitive feature set. For example, for a piece of cloud system data, the initial reflectivity standard deviation is 20, the temperature standard deviation is 5, the set segmentation thresholds are 10 and 3, and the minimum segmentation scale is 1 square kilometer. After recursive segmentation, several sub-regions are finally obtained, and their reflectivity standard deviation and temperature standard deviation are both less than the set thresholds or the scale is less than the minimum segmentation scale, and the features of these sub-regions form the primitive feature set.

[0078] Then, calculate the comprehensive gradient field. Traverse all adjacent primitives in the primitive feature set and calculate the reflectivity gradient and temperature gradient between them. For example, if the reflectivities of two adjacent primitives are 60 and 70 respectively, and the distance is 2 kilometers, then the reflectivity gradient is (70 - 60) / 2 = 5 / km. Calculate the dynamic weight coefficients of reflectivity and temperature according to the variances of the reflectivity gradients and temperature gradients of all adjacent primitive pairs. For example, if the variance of the reflectivity gradient is larger, then assign a larger weight coefficient to the reflectivity gradient. Weightedly fuse the reflectivity gradient, temperature gradient, and their cross-term (the product of the reflectivity gradient and the temperature gradient) to obtain the comprehensive gradient field.

[0079] Next, extract the core region. Perform multi-scale decomposition on the comprehensive gradient field using the Gaussian function, for example, convolve the gradient field with Gaussian kernels of different standard deviations. Extract the local maximum points of the gradient field at each scale. Establish correspondence relationships between these local maximum points across different scales to form an extreme point chain. For example, if a point is a local maximum point at three scales, then these three points form an extreme point chain. Take the continuous length of the extreme point chain in the scale space as the stability index. For example, if an extreme point chain exists at five scales, its stability index is 5. Select the extreme point with the highest stability index as a candidate point for the core region. Conduct spatial connectivity analysis on the candidate points, and define the set of candidate points with an Euclidean space distance less than the set minimum distance threshold as the core region. For example, if the distance between two candidate points is 1 km, which is less than the set minimum distance threshold of 2 km, then these two points belong to the same core region.

[0080] Finally, divide the transition region. Calculate the horizontal wind field gradient outside the core region. Calculate the local turbulence intensity based on the horizontal wind field gradient. Combine the local turbulence intensity with the reflectivity and temperature to obtain the corrected parameter value. Divide the radial distance from the core region to the periphery into multiple intervals. For example, divide the distance into three intervals: 0 - 5 km, 5 - 10 km, and 10 - 15 km. In each interval, fit the parameter decay curve using a double exponential function according to the corrected parameter value, and impose derivative continuity constraints at the interval connection points to ensure the smoothness of the curve. For example, fit a double exponential function in the 0 - 5 km interval and another double exponential function in the 5 - 10 km interval, and ensure that the derivatives of the two functions are equal at 5 km. Calculate the second derivative and the third derivative of the parameter decay curve. Take the weighted sum of the absolute value of the second derivative and the integral of the third derivative as the boundary recognition objective function. Solve for the position that minimizes the boundary recognition objective function as the transition region boundary.

[0081] Through the recursive segmentation algorithm, the present invention can more precisely depict the internal structure of the cloud system, extract more accurate primitive features. The multi-scale analysis and the selection of the stability index can effectively suppress the influence of noise, improve the accuracy and stability of core region recognition, consider the influence of the horizontal wind field gradient and the local turbulence intensity, and use a double exponential function to fit the parameter decay curve, which can more accurately describe the change trend of cloud system parameters, and thus more precisely divide the transition region.

[0082] In an optional implementation manner,

[0083] The steps of constructing a deep digital twin mapping network between the primitive features of the cloud system and the real cloud system, verifying the morphological feature parameters output by the deep digital twin mapping network with meteorological measured data, and dynamically optimizing the parameters of the deep digital twin mapping network based on the verification results include:

[0084] Input the contour curvature, gray-scale distribution, texture features, and multi-scale edge features of the cloud system into a deep digital twin mapping network for feature processing. The deep digital twin mapping network includes a feature encoder, a hierarchical dynamic memory module for storing cloud system morphological patterns, and a multi-task decoder. The feature encoder encodes the features based on an adaptive sampling strategy for cloud system morphological complexity to obtain encoded features, where the cloud system morphological complexity is calculated through a weighted combination of contour curvature variance and texture entropy.

[0085] The hierarchical dynamic memory module updates the memory unit through an attention mechanism and a gated update method that introduces temperature gradient and pressure field physical constraints, dynamically fusing the memory state and encoded features to obtain fused features. Based on the fused features, a multi-granularity spatio-temporal attention mechanism is used to process the dependence relationship of the cloud system morphology, capturing spatial dependence features by adjusting the adaptive receptive field based on the distance from the cloud center, and introducing a recursive attention unit to model the time dependence relationship, obtaining a feature representation with spatio-temporal dependence. Input the feature representation into the multi-task decoder for morphological parameter prediction and reconstruction, and construct a joint loss function including task loss, physical constraint loss, memory update loss, and temporal smoothing loss to optimize the prediction result. The weight coefficients of the joint loss function are dynamically adjusted according to the prediction error.

[0086] Evaluate the accuracy of the prediction result output by the multi-task decoder based on the shape similarity of the Hausdorff distance, and conduct physical verification through area perimeter ratio constraint and neighborhood consistency constraint. The area perimeter ratio constraint limits the value range of the shape coefficient, and the neighborhood consistency constraint ensures that the change of morphological parameters at adjacent times conforms to the meteorological development law, obtaining a verification performance index. Based on the verification performance index, adopt a dynamic learning strategy to optimize the deep digital twin mapping network, where the learning rate is dynamically adjusted based on a periodic decay function, and the sample weight coefficient is calculated based on the prediction error. The sample weight coefficient is calculated through an exponential function of the prediction error and the temperature parameter. Multiply the sample weight coefficient by the task loss, physical constraint loss, memory update loss, and temporal smoothing loss respectively and then update the network parameters.

[0087] Exemplarily, first, cloud system feature extraction is performed. The acquired cloud system image data is preprocessed, including operations such as denoising and enhancing contrast to improve the image quality. Then, various morphological features of the cloud system are extracted, including contour curvature, gray-scale distribution, texture features, and multi-scale edge features. The contour curvature describes the shape complexity of the cloud by calculating the curvature change of the cloud edge. The gray-scale distribution reflects the density and thickness information of the cloud by counting the number of pixels with different gray-scale values. The texture features are extracted using methods such as the gray-level co-occurrence matrix to obtain the texture information of the cloud system, such as roughness and smoothness. The multi-scale edge features are obtained by performing edge detection on the cloud system image using convolution kernels of different sizes to capture edge information at different scales. For example, convolution kernels of sizes 3x3, 5x5, 7x7, etc. are used for edge detection. Suppose the gray-scale value range of a cloud system image is 0 - 255, the contour curvature change range is 0 - 1, the texture feature value range is 0 - 10, and the multi-scale edge intensity range is 0 - 1.

[0088] Next, a deep digital twin mapping network is constructed. This network consists of a feature encoder, a hierarchical dynamic memory module, and a multi-task decoder. The feature encoder receives the extracted cloud system features and performs adaptive sampling according to the cloud system morphological complexity. The cloud system morphological complexity is calculated through a weighted combination of the contour curvature variance and texture entropy. For example, the contour curvature variance weight is 0.6, and the texture entropy weight is 0.4. The encoder encodes the features into a compact feature vector. The hierarchical dynamic memory module is used to store the cloud system morphological patterns. This module updates the memory cells through an attention mechanism and a gated update method that introduces physical constraints such as temperature gradient and pressure field. The memory cells store the morphological features of different types of cloud systems. For example, the memory cells store the typical morphological features of different types of cloud systems such as cumulus clouds, stratus clouds, and cirrus clouds. The multi-task decoder receives the fused features and performs morphological parameter prediction and reconstruction.

[0089] Then, cloud system morphological parameter prediction and reconstruction are carried out. The encoded features and the output of the memory module are fused to obtain the fused features. A multi-granularity spatio-temporal attention mechanism is used to process the dependency relationships of the cloud system morphology. The adaptive receptive field adjustment based on the distance from the cloud system center captures spatial dependency features. For example, the closer the area is to the center, the smaller the receptive field, and the farther the area is from the center, the larger the receptive field. And a recursive attention unit is introduced to model the temporal dependency relationships. For example, the feature information of the previous moment is used to predict the cloud system morphology at the current moment. The obtained feature representation with spatio-temporal dependency relationships is input into the multi-task decoder for morphological parameter prediction and reconstruction. For example, parameters such as the area, perimeter, and height of the cloud system are predicted, and the three-dimensional morphology of the cloud system is reconstructed.

[0090] Next, construct and optimize the loss function. A joint loss function that includes task loss, physical constraint loss, memory update loss, and temporal smoothing loss is constructed to optimize the prediction results. The task loss is used to measure the difference between the predicted parameters and the true parameters. The physical constraint loss is used to ensure that the prediction results conform to physical laws. For example, the area-perimeter ratio constraint and the neighborhood consistency constraint. The memory update loss is used to constrain the update process of the memory unit. The temporal smoothing loss is used to ensure the smoothness of the prediction results at adjacent times. The weight coefficients of the joint loss function are dynamically adjusted according to the prediction error. For example, if the physical constraint loss is large, its weight coefficient is increased.

[0091] Finally, perform dynamic optimization of the network parameters. The accuracy of the prediction results output by the multi-task decoder is evaluated based on the shape similarity of the Hausdorff distance, and physical verification is performed through the area-perimeter ratio constraint and the neighborhood consistency constraint. The area-perimeter ratio constraint limits the value range of the shape coefficient. For example, the value range of the shape coefficient is 0.5 - 1.5. The neighborhood consistency constraint ensures that the change of the morphological parameters at adjacent times conforms to the meteorological development law. For example, the change of the cloud system area at adjacent times does not exceed 10%. After obtaining the verification performance index, a dynamic learning strategy is adopted to optimize the deep digital twin mapping network. The learning rate is dynamically adjusted based on a periodic decay function. For example, the initial value of the learning rate is 0.01, and it decays to 0.9 times the original value every 10 epochs. And the sample weight coefficient is calculated based on the prediction error. The sample weight coefficient is calculated through the exponential function of the prediction error and the temperature parameter. The network parameters are updated after multiplying the sample weight coefficient by the task loss, physical constraint loss, memory update loss, and temporal smoothing loss respectively.

[0092] By constructing a deep digital twin mapping network and combining multi-granularity spatio-temporal attention mechanism and physical constraints, the present invention can more accurately capture the morphological characteristics and evolution laws of cloud systems, thereby improving the prediction accuracy of cloud system morphological parameters; the multi-task decoder can reconstruct the three-dimensional morphology of cloud systems, providing more intuitive and comprehensive cloud system information, which helps to better understand the formation and development process of clouds; the dynamic memory module and dynamic learning strategy can enable the network to adapt to different types of cloud systems and different meteorological conditions, improving the generalization ability and robustness of the network.

[0093] In an optional embodiment,

[0094] The steps of performing refined discretization on the core area using the adaptive multi-grid method, where the grid scale is automatically adjusted according to the cloud cluster turbulence intensity and the thermal parameter gradient, and establishing a non-uniform grid considering the internal vortex evolution of the cloud system in the transition area include:

[0095] Calculate the turbulent kinetic energy spectrum and vorticity field distribution in the core area based on meteorological radar reflectivity and Doppler velocity field data. Obtain the turbulent intensity distribution at different scales through the analysis of the turbulent kinetic energy spectrum, and identify the vortex structure in the core area by the vorticity threshold method. Extract the vortex center position and vortex intensity.

[0096] Use meteorological sounding data to construct the temperature field and pressure field in the cloud cluster area, calculate the temperature gradient and pressure gradient, and obtain the thermal parameter gradient intensity index by weighted combination of the absolute value of the temperature gradient and the absolute value of the pressure gradient. Input the turbulent intensity distribution at different scales and the thermal parameter gradient intensity index into an exponential function, and obtain the grid scale mapping function through weighted combination.

[0097] Adopt a quadtree structure for multi-grid division. Make a subdivision judgment on grid cells according to the ratio of the output value of the grid scale mapping function to the initial grid scale. Divide the grid cells that meet the subdivision threshold conditions into four sub-grids, and limit the difference in adjacent grid levels.

[0098] Construct the vortex intensity evolution relationship based on the vortex center position and vortex intensity. Calculate the grid density attenuation coefficient in the transition area through the hyperbolic tangent function according to the vortex center distance. The grid density attenuation coefficient increases with the increase of the distance to the vortex center.

[0099] Obtain the grid quality evaluation index by weighted combination of the grid shape quality, orthogonality index and feature capture ability. When the grid quality evaluation index is lower than the preset threshold, determine the grid cells to be optimized. Input the position information of the grid cells to be optimized into the grid scale mapping function to recalculate the local grid scale. Update the grid topology structure based on the recalculated grid scale and interpolate and reconstruct the physical field.

[0100] Exemplarily, first, obtain meteorological radar reflectivity and Doppler velocity field data. Take a severe convective weather process as an example. Use the volume scan data obtained by a Doppler meteorological radar with a time resolution of 6 minutes and a spatial resolution of 1 km × 1°. Perform quality control on the obtained radar base data to remove noise and clutter interference.

[0101] Then, calculate the turbulent intensity and vortex characteristics in the core area. Calculate the turbulent kinetic energy spectrum using radar reflectivity data and analyze the turbulent intensity distribution at different scales. Assume that the turbulent kinetic energy spectrum at a certain grid point is E(k), where k is the wave number. By integrating E(k) over different wave number segments, the turbulent energy at different scales can be obtained. At the same time, calculate the vorticity field distribution using Doppler velocity field data, and adopt the vorticity threshold method to identify the vortex structure in the core area. Extract the vortex center position and vortex intensity. Record the coordinate position of the vortex center and the maximum vorticity value in the vortex area as the vortex intensity.

[0102] Next, construct the thermal parameter gradient intensity index for the cloud cluster region. Use meteorological sounding data to construct the temperature field and pressure field of the cloud cluster region. Assume that the time resolution of the sounding data is 12 hours and the spatial resolution is 500 meters. Through interpolation processing of the sounding data, a temperature field and a pressure field with the same spatial resolution as the radar data are obtained. Then, calculate the temperature gradient and pressure gradient, and perform a weighted combination of the absolute values of the temperature gradient and the absolute value of the pressure gradient to obtain the thermal parameter gradient intensity index. For example, set the temperature gradient weight to 0.7 and the pressure gradient weight to 0.3, and perform a weighted sum of the two to obtain the thermal parameter gradient intensity index. Subsequently, construct the grid scale mapping function. Input the turbulent intensity distribution at different scales and the thermal parameter gradient intensity index into an exponential function, and obtain the grid scale mapping function through a weighted combination.

[0103] Then, perform adaptive multi-grid partitioning. Use a quadtree structure for multi-grid partitioning, and judge the subdivision of grid cells according to the ratio of the output value of the grid scale mapping function to the initial grid scale. Assume that the initial grid scale is 1 km and the subdivision threshold is set to 2. If the output value of the grid scale mapping function of a certain grid cell is 0.5 km, then this grid cell meets the subdivision condition and is divided into four sub-grids, and the scale of each sub-grid is 0.5 km. At the same time, limit the difference in adjacent grid levels to not exceed 1 to ensure the smooth transition of the grid.

[0104] Next, construct a non-uniform grid for the transition region. Based on the vortex center position and vortex intensity, construct the vortex intensity evolution relationship, and calculate the grid density attenuation coefficient in the transition region through the hyperbolic tangent function according to the distance from the vortex center. Assume that the grid density at the vortex center is 1, and as the distance from the vortex center increases, the grid density gradually decreases.

[0105] Finally, perform grid quality optimization. Combine the grid shape quality, orthogonality index, and feature capture ability through weighting to obtain the grid quality evaluation index. Assume that the grid shape quality weight is 0.4, the orthogonality index weight is 0.3, and the feature capture ability weight is 0.3, and perform a weighted sum of the three to obtain the grid quality evaluation index. When the grid quality evaluation index is lower than the preset threshold, determine the grid cells that need to be optimized, input the position information of the grid cells that need to be optimized into the grid scale mapping function to recalculate the local grid scale, update the grid topology structure based on the recalculated grid scale, and perform interpolation reconstruction on the physical field.

[0106] The present invention can more accurately capture the fine structure and evolution process inside the cloud cluster. Through the adaptive grid adjustment strategy, it realizes the reasonable allocation of computing resources, improves the computing efficiency, enhances the forecasting accuracy of severe convective weather processes, and provides a more reliable scientific basis for disaster prevention and mitigation.

[0107] In an alternative embodiment,

[0108] Based on the grid division results, perform simulation modeling on the dynamic evolution of the cloud system, establish cross-region coupling constraints including mass conservation, momentum conservation, and energy conservation, characterize the dynamic evolution process of the cloud system as a partial differential equation system including turbulent diffusion, thermodynamic parameters, and region coupling, integrate the partial differential equation system with the deep digital twin mapping network, solve the overall evolution characteristics of the cloud system using the region decomposition method, and dynamically correct the evolution prediction results using real-time digital twin data. The steps include:

[0109] Establish cross-region coupling constraints of mass conservation, momentum conservation, and energy conservation at the interface between the core region and the transition region. The mass conservation constraint is based on the flux continuity conditions of the density field and the velocity field, the momentum conservation constraint is based on pressure and the viscous stress tensor, and the energy conservation constraint is based on specific internal energy, thermal conductivity, and the temperature field;

[0110] Construct a partial differential equation system for the evolution of the cloud system including a turbulent diffusion term, a thermodynamic parameter term, and a region coupling term. The turbulent diffusion term is characterized by turbulent kinetic energy, turbulent diffusion coefficient, and dissipation rate. The thermodynamic parameter term is characterized by the temperature field, pressure field, and density field. The region coupling term is determined by the interface flux balance. The partial differential equation system correlates the cross-region coupling constraints with the turbulent diffusion term, the thermodynamic parameter term, and the region coupling term;

[0111] Input the partial differential equation system into the deep digital twin mapping network, establish a mapping loss function including a physical constraint loss term and a conservation constraint loss term, and optimize the parameters of the deep digital twin mapping network by minimizing the mapping loss function;

[0112] Use the region decomposition method to solve the output results of the deep digital twin mapping network. Based on the distribution characteristics of the turbulent diffusion term and the thermodynamic parameter term, fix the physical parameters of the transition region to solve the physical parameters of the core region, and use the updated physical parameters of the core region as boundary conditions to solve the physical parameters of the transition region. Iterate until the physical parameters of the core region and the transition region reach the preset convergence conditions. The physical parameters include the velocity field, density field, turbulent kinetic energy, turbulent diffusion coefficient, dissipation rate, temperature field, pressure field, specific internal energy, and thermal conductivity;

[0113] Receive the near-surface data and the top data of the cloud system as real-time digital twin data, construct a Kalman gain matrix based on the prediction error covariance of the physical parameters of the core region and the transition region and the observation error covariance of the real-time digital twin data, perform error analysis on the real-time digital twin data and the solution results of the region decomposition method, and dynamically correct the evolution prediction results of the cloud system using the Kalman gain matrix.

[0114] Exemplarily, first, spatial discretization is performed. The three-dimensional space where the cloud system is located is divided into a series of non-overlapping grid cells. In order to capture the details of the core region (such as the severe convection region), a finer grid division is adopted for the core region, while a coarser grid is used for the surrounding transition region. For example, the grid resolution of the core region can be set to 1 km in the horizontal direction and 0.5 km in the vertical direction, and the grid resolution of the transition region can be set to 5 km in the horizontal direction and 2 km in the vertical direction.

[0115] Then, coupling constraints between the core region and the transition region are established. On the interface between the core region and the transition region, the conservation of physical quantities needs to be ensured. The mass conservation constraint requires that the mass flowing into the interface is equal to the mass flowing out of the interface, which can be achieved through the flux balance of density and velocity on the interface. The momentum conservation constraint requires that the pressure and viscous forces on both sides of the interface are balanced with each other. The energy conservation constraint requires that the energy exchange on both sides of the interface remains balanced, which can be achieved through the flux balance of specific internal energy, thermal conductivity, and temperature on the interface.

[0116] Next, a partial differential equation system for the evolution of the cloud system is constructed. This equation system describes the spatio-temporal evolution laws of various physical quantities in the cloud system and mainly includes three parts: the turbulent diffusion term, the thermodynamic parameter term, and the regional coupling term. The turbulent diffusion term describes the influence of turbulent motion in the cloud system on physical quantities, such as the diffusion of momentum, heat, and water vapor. It can be expressed by turbulent kinetic energy, turbulent diffusion coefficient, and dissipation rate. The thermodynamic parameter term describes the influence of thermodynamic parameters such as temperature, pressure, and density on the evolution of the cloud system. The regional coupling term describes the interaction between the core region and the transition region and is determined by the flux balance conditions on the interface. This partial differential equation system correlates factors such as conservation constraints, turbulent diffusion, thermodynamic parameters, and regional coupling to form a complete cloud system evolution model.

[0117] Subsequently, a deep digital twin mapping network is constructed. The above partial differential equation system is input into a deep neural network, and this network learns the complex non-linear relationships of the cloud system evolution. In order to train this network, a loss function needs to be constructed, and this loss function includes two parts: the physical constraint loss term and the conservation constraint loss term. The physical constraint loss term is used to ensure that the output of the network satisfies physical laws, such as the conservation of mass, momentum, and energy. The conservation constraint loss term is used to ensure that the conservation constraints between the core region and the transition region are satisfied. By minimizing the loss function, the parameters of the network can be optimized to enable it to accurately predict the cloud system evolution. For example, a neural network containing convolutional layers and recurrent layers can be used, and the mean square error can be used as the loss function.

[0118] Next, the domain decomposition method is adopted for solution. The output result of the deep digital twin mapping network is used as the input of the domain decomposition method. First, according to the distribution characteristics of turbulent diffusion and thermodynamic parameters, the physical parameters (such as velocity, temperature, pressure, etc.) in the transition region are fixed, and the physical parameters in the core region are solved. Then, the updated physical parameters in the core region are used as boundary conditions to solve the physical parameters in the transition region. Such cyclic iteration is carried out until the physical parameters in both the core region and the transition region reach the preset convergence conditions, such as the difference between the results of two iterations is less than a certain threshold.

[0119] Finally, dynamic correction is performed using real-time digital twin data. The near-earth data and top data of the cloud system are obtained in real time, such as satellite observation data, radar data, and ground meteorological station data. The prediction error covariance of the physical parameters in the core region and the transition region, as well as the error covariance of the real-time observation data, are calculated. A Kalman gain matrix is constructed based on these two covariance matrices. The real-time observation data is fused with the solution result of the domain decomposition method using the Kalman gain matrix to dynamically correct the prediction result of the cloud system evolution and improve the prediction accuracy. For example, assuming the observation data includes temperature and wind speed, the Kalman filtering algorithm can be used to fuse the observation data with the model prediction result to correct the predicted values of temperature and wind speed.

[0120] The present invention can more accurately capture the complex dynamic evolution process of the cloud system by combining conservation laws, deep learning, and real-time data, thereby improving the prediction accuracy; by introducing conservation laws into the model, the physical consistency of the model is ensured, and physical unreasonable phenomena that may occur in traditional numerical models are avoided; through the domain decomposition method, complex computational tasks are decomposed into multiple subtasks and solved in parallel, thereby improving the computational efficiency and making real-time prediction of the cloud system possible.

[0121] In an optional implementation manner,

[0122] The steps of inputting the partial differential equation system into the deep digital twin mapping network, establishing a mapping loss function including a physical constraint loss term and a conservation constraint loss term, and optimizing the parameters of the deep digital twin mapping network by minimizing the mapping loss function include:

[0123] The physical constraint loss term includes a turbulent dynamics loss and a thermodynamic coupling loss. The turbulent dynamics loss is constructed based on the spatio-temporal derivatives of turbulent kinetic energy and dissipation rate and the turbulent production term. The weight coefficient of the turbulent dynamics loss is adaptively adjusted by local turbulent intensity and turbulent Reynolds number. The thermodynamic coupling loss is constructed based on the convection-diffusion equations of the temperature field and the pressure field. The weight coefficient of the thermodynamic coupling loss is dynamically determined by a non-linear function of the temperature gradient and the pressure gradient;

[0124] Construct the conservation constraint loss term based on the cross-region coupling constraint. The conservation constraint loss term includes mass conservation constraint, momentum conservation constraint and energy conservation constraint, and introduce a stress tensor correction term based on vorticity correlation;

[0125] Construct a feature-driven combined loss function, which combines the physical constraint loss term and the conservation constraint loss term with weights. The weight coefficients of the combined loss function are dynamically adjusted by the cloud system morphology complexity, temperature gradient and pressure gradient. The cloud system morphology complexity is calculated by the weighted combination of contour curvature variance and texture entropy;

[0126] Introduce a multi-scale residual correction mechanism, which includes local residual and global residual. The local residual is calculated based on turbulent characteristic quantities, and the global residual is calculated based on thermodynamic quantities. The weight coefficients of the local residual and global residual are dynamically determined by the ratio of the turbulent characteristic scale to the cloud system characteristic scale; Optimize and update the parameters of the deep digital twin mapping network by minimizing the combined loss function and combining the multi-scale residual correction results. The learning rate is adaptively adjusted based on the physical quantity gradient and prediction error. The physical quantities include velocity field, vorticity field, pressure field, density field, temperature field, turbulent kinetic energy and dissipation rate. At the same time, introduce a periodic decay mechanism. At the end of each training cycle, the learning rate is adjusted stage by stage according to the satisfaction degrees of the physical constraint loss term and the conservation constraint loss term.

[0127] Exemplarily, first, a deep digital twin mapping network needs to be constructed. This network can adopt a deep learning architecture, such as a convolutional neural network (CNN) or a recurrent neural network (RNN). The input data includes the initial state and boundary conditions of cloud system evolution, such as the numerical values of physical quantities such as velocity field, temperature field, and pressure field at the initial moment and boundaries. The output is the cloud system state at a future moment, such as the predicted values of velocity field, temperature field, and pressure field at a future moment. The specific structure of the network can be adjusted according to the actual situation, such as the number of convolutional layers, the size of convolutional kernels, and the depth of the network.

[0128] Next, construct a mapping loss function to train the network. This loss function includes multiple parts, which are used to constrain the prediction results of the network to satisfy physical laws and conservation laws respectively.

[0129] The physical constraint loss term is used to constrain the prediction results of the network to conform to the laws of hydrodynamics and thermodynamics. Among them, the turbulent kinetic energy loss term is used to constrain the spatio-temporal variation laws of turbulent kinetic energy and dissipation rate. The weight coefficient of the turbulent kinetic energy loss is adaptively adjusted according to the local turbulence intensity and turbulent Reynolds number. For example, the greater the turbulence intensity, the greater the weight coefficient. The thermodynamics coupling loss term is used to constrain the convection-diffusion laws of the temperature field and pressure field, and its weight coefficient is dynamically determined according to the non-linear function of the temperature gradient and pressure gradient. For example, the greater the temperature gradient, the greater the weight coefficient.

[0130] The conservation constraint loss term is used to constrain the prediction results of the network to satisfy the laws of mass, momentum and energy conservation. For example, the mass conservation loss can be constructed by calculating the difference between the change of the predicted density field within the control volume and the mass flux through the surface of the control volume. The momentum conservation and energy conservation losses can be constructed in a similar way. In addition, by introducing a stress tensor correction term based on vorticity correlation, the accuracy of the conservation constraint can be improved.

[0131] The physical constraint loss term and the conservation constraint loss term are weighted and combined to construct a feature-driven combined loss function. The weight coefficient of the combined loss function is dynamically adjusted according to the cloud system morphology complexity, temperature gradient and pressure gradient. The cloud system morphology complexity can be calculated by the weighted combination of the contour curvature variance and texture entropy. For example, the more complex the cloud system morphology, the greater the contour curvature variance and texture entropy, and the greater the corresponding weight coefficient.

[0132] To further improve the prediction accuracy, a multi-scale residual correction mechanism is introduced. This mechanism includes local residuals and global residuals. The local residuals are calculated based on turbulent characteristic quantities, such as based on turbulent fluctuating velocity. The global residuals are calculated based on thermodynamic quantities, such as based on the average values of temperature and pressure. The weight coefficients of the local residuals and global residuals are dynamically determined according to the ratio of the turbulent characteristic scale to the cloud system characteristic scale. For example, the smaller the turbulent characteristic scale relative to the cloud system characteristic scale, the greater the weight coefficient of the local residuals.

[0133] Finally, the network parameters are optimized and updated by minimizing the combined loss function and combining the multi-scale residual correction results. The learning rate is adaptively adjusted based on the physical quantity gradient and prediction error. For example, the greater the physical quantity gradient and the greater the prediction error, the greater the learning rate. At the same time, a periodic decay mechanism is introduced. At the end of each training cycle, the learning rate is adjusted stage by stage according to the satisfaction degrees of the physical constraint loss term and the conservation constraint loss term. For example, the smaller the value of the loss term, the higher the model prediction accuracy, and the learning rate can be appropriately reduced.

[0134] By combining physical constraints and conservation constraints and introducing a multi-scale residual correction mechanism, the present invention can effectively improve the prediction accuracy of cloud system evolution and more accurately capture the evolution process of cloud systems; by adaptively adjusting the weight coefficients and learning rates of the loss function, the network can better adapt to different cloud system morphologies and evolution processes, enhancing the generalization ability of the network; through deep learning technology, the computational cost of cloud system evolution simulation can be effectively reduced, the computational efficiency can be improved, and rapid prediction of cloud system evolution can be achieved.

[0135] In an optional implementation manner,

[0136] The steps of receiving the cloud system near-surface data and the cloud system top data as real-time digital twin data, constructing a Kalman gain matrix based on the prediction error covariance of the physical parameters in the core region and the transition region and the observation error covariance of the real-time digital twin data, performing error analysis on the real-time digital twin data and the solution result of the regional decomposition method, and dynamically correcting the cloud system evolution prediction result by using the Kalman gain matrix include:

[0137] Obtain cloud system near-surface data including reflectivity, radial velocity, and spectral width, fuse multi-site radar data by using a three-dimensional variational assimilation method, construct the distribution of the three-dimensional wind field, temperature field, and water vapor field in the near-surface layer, and determine the ground-based observation error covariance matrix based on historical observation statistics; obtain cloud system top data including cloud top height, cloud top temperature, and cloud top radiation flux, extract cloud system characteristic parameters by using an infrared-visible two-channel inversion algorithm, and obtain cloud system motion information through multi-temporal sequence registration, and determine the satellite observation error covariance matrix through channel noise analysis;

[0138] Construct a prediction error covariance matrix based on the difference prediction error between the prediction result of the physical parameters in the core region and the transition region and the corresponding true value, and introduce a local correlation constraint based on the spatial distance of the physical parameters to correct the prediction error covariance matrix, and the local correlation constraint is determined by the correlation length scale;

[0139] Construct a Kalman gain matrix based on the prediction error covariance matrix, the ground-based observation error covariance matrix, and the satellite observation error covariance matrix, and introduce an adaptive adjustment mechanism based on the cloud system morphology complexity to correct the Kalman gain matrix, and the cloud system morphology complexity is calculated by a weighted combination of the contour curvature variance and the texture entropy;

[0140] Map the near - surface data and the top - layer data of the cloud system to the observation space, calculate the observation residual vector with respect to the predicted value, and update the predicted state based on the observation residual vector and the corrected Kalman gain matrix. Introduce the turbulence feature constraint, the mass conservation constraint, and the energy conservation constraint to the updated predicted state. The turbulence feature constraint is constructed based on the non - negativity of the turbulent kinetic energy and dissipation rate. Use the constrained predicted state as the initial condition for the prediction at the next moment to form a dynamically corrected closed - loop iterative process.

[0141] Exemplarily, first, obtain the near - surface data of the cloud system. Use the Doppler weather radar network deployed on the ground to scan and collect radar echo data containing information such as reflectivity, radial velocity, and spectral width. Then, adopt the three - dimensional variational assimilation method to fuse the observation data from multiple radar stations, thereby constructing the three - dimensional wind field, temperature field, and water vapor field distributions near the ground. At the same time, according to the statistical results of past observation data, determine the error covariance matrix of the ground - based radar observation for subsequent data fusion and error correction.

[0142] Next, obtain the top - layer data of the cloud system. Use the infrared and visible light sensors carried by meteorological satellites to obtain data such as cloud top height, cloud top temperature, and cloud top radiation flux. Adopt the infrared - visible light dual - channel inversion algorithm to extract cloud system characteristic parameters, such as cloud optical thickness and effective particle radius. Obtain the motion information of the cloud system, such as cloud movement speed and direction, by registering the multi - temporal satellite image sequences. In addition, determine the error covariance matrix of the satellite observation by analyzing the channel noise of the sensor.

[0143] Then, construct and correct the prediction error covariance matrix. Compare the predicted results of the physical parameters in the core region and the transition region obtained by the regional decomposition method with the corresponding true values, calculate their differences, that is, the prediction errors. Use these prediction errors to construct the prediction error covariance matrix. To improve the prediction accuracy, introduce a local correlation constraint based on the spatial distance of physical parameters to correct the prediction error covariance matrix. Specifically, define the local correlation constraint through the correlation length scale. The closer the grid points are, the higher the correlation of their physical parameters.

[0144] Next, construct the Kalman gain matrix and perform correction. Use the previously obtained prediction error covariance matrix, ground - based observation error covariance matrix, and satellite observation error covariance matrix to construct the Kalman gain matrix. To adapt to the complexity of different cloud system morphologies, introduce an adaptive adjustment mechanism based on the cloud system morphology complexity to correct the Kalman gain matrix. The cloud system morphology complexity is calculated through the weighted combination of the contour curvature variance and texture entropy. For example, for a cloud system with a complex contour and rich texture, assign a higher weight, thereby increasing the Kalman gain and paying more attention to the influence of the observation data.

[0145] Subsequently, the prediction state is updated. The near-ground data and top data of the cloud system are mapped to the observation space, and the observation residual vector with respect to the predicted value is calculated. Using the calculated observation residual vector and the corrected Kalman gain matrix, the prediction state is updated to obtain a more accurate state estimate of the cloud system.

[0146] Finally, constraints are introduced to the updated prediction state and iteration is performed. Turbulence feature constraints, mass conservation constraints, and energy conservation constraints are introduced to the updated prediction state to ensure the physical consistency of the prediction results. The turbulence feature constraints are constructed based on the non-negativity of the turbulent kinetic energy and dissipation rate. The constrained prediction state is used as the initial condition for the next moment's prediction, forming a dynamically corrected closed-loop iterative process to achieve continuous and dynamic correction of the cloud system evolution.

[0147] Suppose at a certain moment, the radar observes that the wind speed at the bottom of the cloud system is 10 m / s, and the satellite observes that the cloud top height is 5000 m. The wind speed predicted by the domain decomposition method is 8 m / s, and the cloud top height is 4500 m. After error analysis and Kalman gain correction, the finally predicted wind speed is corrected to 9.5 m / s, and the cloud top height is corrected to 4800 m, which is closer to the actual observed values.

[0148] The present invention can effectively reduce the prediction error and improve the accuracy of cloud system evolution prediction by integrating ground-based radar and satellite observation data and combining with the prediction results of the domain decomposition method; introducing an adaptive adjustment mechanism based on the morphological complexity of the cloud system can dynamically adjust the Kalman gain matrix according to the characteristics of different cloud systems, so as to better adapt to different types of cloud systems and improve the robustness of the prediction; introducing turbulence feature constraints, mass conservation constraints, and energy conservation constraints ensures the physical consistency of the prediction results and makes the prediction results more in line with the actual situation.

[0149] Figure 2 This is a schematic structural diagram of the multi-scale cloud system dynamic evolution simulation modeling system based on digital twin according to an embodiment of the present invention, as Figure 2 shown, the system includes:

[0150] The first unit is used to construct a multi-source collaborative perception network using a multi-spectral camera, lidar, and millimeter-wave radar to obtain near-ground data of the cloud system, and combine satellite remote sensors to obtain top data of the cloud system and ground meteorological stations to obtain meteorological element data; perform phase consistency analysis on the obtained multi-source data, establish a spatial transformation matrix between sensor coordinate systems to obtain registered data, and perform quality assessment on the registered data according to signal-to-noise ratio, spatial resolution, and time consistency, and screen out valid data that meet the assessment threshold;

[0151] A second unit, which is configured to extract the cloud system primitive features based on the valid data by using a recursive partitioning algorithm, calculate the extreme points of the reflectivity gradient and the temperature gradient to divide the core region of the cloud cluster, and calculate the parameter attenuation curve around the core region to divide the transition region; based on the region division result, extract the contour curvature, gray distribution, texture features and multi-scale edge features of the cloud system, construct a deep digital twin mapping network between the cloud system primitive features and the real cloud system, verify the morphological feature parameters output by the deep digital twin mapping network with meteorological measured data, and dynamically optimize the parameters of the deep digital twin mapping network based on the verification result;

[0152] A third unit, which is configured to perform refined discretization on the core region by using an adaptive multi-grid method, automatically adjust the grid scale according to the cloud cluster turbulence intensity and the thermal parameter gradient, and establish a non-uniform grid considering the internal vortex evolution of the cloud system for the transition region; based on the grid division result, perform simulation modeling on the dynamic evolution of the cloud system, establish cross-region coupling constraints including mass conservation, momentum conservation and energy conservation, characterize the dynamic evolution process of the cloud system as a partial differential equation system including turbulent diffusion, thermodynamic parameters and region coupling, integrate the partial differential equation system with the deep digital twin mapping network, use the domain decomposition method to solve the overall evolution characteristics of the cloud system, and dynamically correct the evolution prediction result by using real-time digital twin data.

[0153] In the third aspect of the embodiments of the present invention,

[0154] There is provided an electronic device, including:

[0155] A processor;

[0156] A memory for storing processor-executable instructions;

[0157] Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0158] In the fourth aspect of the embodiments of the present invention,

[0159] There is provided a computer-readable storage medium, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0160] The present invention may be a method, an apparatus, a system, and / or a computer program product. The computer program product may include a computer-readable storage medium, on which computer-readable program instructions for executing various aspects of the present invention are uploaded.

[0161] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than limiting them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-scale cloud system dynamic evolution simulation modeling method based on digital twins, characterized by: include: A multi-source collaborative sensing network is constructed using multispectral cameras, lidar, and millimeter-wave radar to acquire cloud near-ground data. This network is combined with satellite remote sensors to acquire cloud top data and ground-based meteorological stations to acquire meteorological element data. Phase consistency analysis is performed on the acquired multi-source data, and a spatial transformation matrix is established between the sensor coordinate systems to obtain registered data. The registered data is then quality-assessed based on signal-to-noise ratio, spatial resolution, and temporal consistency, and valid data that meets the assessment threshold is selected. Based on the valid data, a recursive segmentation algorithm is used to extract the primitive features of the cloud system, the extreme points of the reflectivity gradient and the temperature gradient are calculated to divide the core area of the cloud cluster, and the parameter attenuation curve around the core area is calculated to divide the transition area; based on the regional division result, the contour curvature, grayscale distribution, texture features and multi-scale edge features of the cloud system are extracted, and a deep digital twin mapping network of the primitive features of the cloud system and the real cloud system is constructed. The morphological feature parameters output by the deep digital twin mapping network are verified with the meteorological measured data, and the parameters of the deep digital twin mapping network are dynamically optimized based on the verification results. The deep digital twin mapping network includes a feature encoder, a hierarchical dynamic memory module for storing cloud morphological patterns, and a multi-task decoder; the feature encoder encodes the features based on the adaptive sampling strategy of the cloud morphological complexity to obtain encoded features. The cloud morphological complexity is calculated through a weighted combination of contour curvature variance and texture entropy. The hierarchical dynamic memory module updates the memory units through an attention mechanism and a gated update method that introduces temperature gradient and pressure field physical constraints, dynamically fusing memory states and encoding features to obtain fused features. Based on the fused features, a multi-granularity spatiotemporal attention mechanism is used to process the cloud morphological dependencies. Spatial dependency features are captured through adaptive receptive field adjustment based on cloud center distance, and temporal dependencies are modeled through the introduction of recursive attention units to obtain feature representations with spatiotemporal dependencies. The feature representations are input into a multi-task decoder for morphological parameter prediction and reconstruction, and a joint loss function consisting of task loss, physical constraint loss, memory update loss, and temporal smoothing loss is constructed to optimize the prediction results. The weight coefficients of the joint loss function are dynamically adjusted according to the prediction error. An adaptive multi-grid method is used to finely discretize the core area, and the grid scale is automatically adjusted with the cloud turbulence intensity and thermal parameter gradient. A non-uniform grid is established for the transition area to consider the evolution of vortices within the cloud system. Based on the grid division results, the dynamic evolution of the cloud system is simulated and modeled, and cross-regional coupling constraints including mass conservation, momentum conservation and energy conservation are established. The dynamic evolution process of the cloud system is characterized as a set of partial differential equations including turbulent diffusion, thermodynamic parameters and regional coupling. The set of partial differential equations is integrated with the deep digital twin mapping network, and the regional decomposition method is used to solve the overall evolution characteristics of the cloud system. The evolution prediction results are dynamically corrected using real-time digital twin data.

2. The method according to claim 1, characterized in that Based on the valid data, a recursive segmentation algorithm is used to extract cloud primitive features. The steps of calculating the extreme points of reflectivity gradient and temperature gradient to divide the core area of the cloud cluster and calculating the parameter attenuation curve around the core area to divide the transition area include: A recursive segmentation algorithm is used to extract cloud primitive features. The mean and standard deviation of reflectivity, as well as the mean and standard deviation of temperature, are calculated for each sub-region. When the standard deviation of reflectivity or the standard deviation of temperature in a sub-region is greater than a preset segmentation threshold, the region is further segmented until the standard deviation of reflectivity and the standard deviation of temperature are less than the preset segmentation threshold or the minimum segmentation scale is reached. The feature vectors of the primitive regions are recorded to form a primitive feature set. Calculating the reflectivity gradient and temperature gradient between adjacent primitives based on the primitive feature set, calculating the dynamic weight coefficients of the reflectivity and temperature according to the variance of the reflectivity gradient and the temperature gradient, and performing weighted fusion on the reflectivity gradient, the temperature gradient and their cross terms to obtain a comprehensive gradient field; The integrated gradient field is decomposed into multiple scales using a Gaussian function, and local maximum points are extracted from the gradient field at each scale. Correspondence between the local maximum points at different scales is established to form an extreme point chain. The persistence length of the extreme point chain in scale space is used as a stability indicator, and the extreme point with the highest stability indicator is selected as a candidate point for the core area. Spatial connectivity analysis is performed on the candidate points, and the set of candidate points whose Euclidean space distance is less than a set minimum distance threshold is defined as the core area. Calculate the horizontal wind field gradient around the core area, calculate the local turbulence intensity based on the horizontal wind field gradient, and combine the local turbulence intensity with reflectivity and temperature to obtain a corrected parameter value; divide the radial distance from the core area to the periphery into multiple intervals, fit the parameter attenuation curve using a double exponential function in each interval based on the corrected parameter value, and impose derivative continuity constraints at the interval connections; calculate the second-order derivative and the third-order derivative of the parameter attenuation curve, use the weighted sum of the absolute value of the second-order derivative and the integral of the third-order derivative as the boundary identification objective function, and solve the position that minimizes the boundary identification objective function as the transition area boundary.

3. The method according to claim 1, characterized in that The steps of constructing a deep digital twin mapping network between the cloud primitive features and the real cloud system, verifying the morphological feature parameters output by the deep digital twin mapping network with meteorological measured data, and dynamically optimizing the parameters of the deep digital twin mapping network based on the verification results include: The cloud system’s contour curvature, grayscale distribution, texture features, and multi-scale edge features are input into the deep digital twin mapping network for feature processing; The prediction results output by the multi-task decoder are accurately evaluated based on the shape similarity of the Hausdorff distance, and are physically verified through area-perimeter ratio constraints and neighborhood consistency constraints, wherein the area-perimeter ratio constraint limits the value range of the shape coefficient, and the neighborhood consistency constraint ensures that the changes in the morphological parameters at adjacent moments conform to the laws of meteorological development, and obtains verification performance indicators; based on the verification performance indicators, a dynamic learning strategy is adopted to optimize the deep digital twin mapping network, wherein the learning rate is dynamically adjusted based on a periodic attenuation function, and the sample weight coefficient is calculated based on the prediction error, and the sample weight coefficient is calculated by the exponential function of the prediction error and the temperature parameter, and the network parameters are updated after the sample weight coefficient is multiplied by the task loss, physical constraint loss, memory update loss and temporal smoothing loss respectively.

4. The method according to claim 1, wherein The core region is discretized using an adaptive multigrid method. The grid size is automatically adjusted according to the cloud turbulence intensity and thermal parameter gradient. The steps to establish a non-uniform grid in the transition region that takes into account the evolution of vortices within the cloud system include: The turbulent kinetic energy spectrum and vorticity field distribution in the core area are calculated based on meteorological radar reflectivity and Doppler velocity field data. The turbulence intensity distribution at different scales is obtained by analyzing the turbulent kinetic energy spectrum. The vortex structure in the core area is identified using the vortex threshold method, and the vortex center position and vortex intensity are extracted. The temperature field and pressure field of the cloud area are constructed using meteorological sounding data, the temperature gradient and pressure gradient are calculated, and the absolute value of the temperature gradient and the absolute value of the pressure gradient are weightedly combined to obtain a thermal parameter gradient intensity index; the turbulence intensity distribution at different scales and the thermal parameter gradient intensity index are input into an exponential function, and a grid scale mapping function is obtained through weighted combination; A quadtree structure is used for multiple grid division, and the grid cells are subdivided according to the ratio of the output value of the grid scale mapping function to the initial grid scale. The grid cells that meet the subdivision threshold condition are divided into four subgrids, and the difference between adjacent grid levels is limited. A vortex intensity evolution relationship is established based on the vortex center position and the vortex intensity, and a grid density attenuation coefficient of the transition region is calculated using a hyperbolic tangent function according to the distance from the vortex center, wherein the grid density attenuation coefficient increases with increasing distance from the vortex center; A grid quality evaluation index is obtained by weighted combination of grid shape quality, orthogonality index and feature capture capability. When the grid quality evaluation index is lower than a preset threshold, the grid cells that need to be optimized are determined, and the position information of the grid cells that need to be optimized is input into the grid scale mapping function to recalculate the local grid scale. The grid topology structure is updated based on the recalculated grid scale, and the physical field is interpolated and reconstructed.

5. The method according to claim 1, characterized in that Based on the grid division results, the dynamic evolution of the cloud system is simulated and modeled. Cross-region coupling constraints including mass conservation, momentum conservation, and energy conservation are established. The dynamic evolution process of the cloud system is characterized as a set of partial differential equations including turbulent diffusion, thermodynamic parameters, and regional coupling. The partial differential equations are integrated with the deep digital twin mapping network, and the regional decomposition method is used to solve the overall evolution characteristics of the cloud system. The evolution prediction results are dynamically corrected using real-time digital twin data. The steps include: Cross-region coupling constraints of mass conservation, momentum conservation, and energy conservation are established at the interface between the core region and the transition region. The mass conservation constraint is based on the flux continuity conditions of the density field and velocity field, the momentum conservation constraint is based on the pressure and viscous stress tensors, and the energy conservation constraint is based on the specific internal energy, thermal conductivity, and temperature field. Constructing a system of partial differential equations for cloud system evolution including a turbulent diffusion term, a thermodynamic parameter term, and a regional coupling term, wherein the turbulent diffusion term is characterized by turbulent kinetic energy, turbulent diffusion coefficient, and dissipation rate, the thermodynamic parameter term is characterized by temperature field, pressure field, and density field, and the regional coupling term is determined by interface flux balance, wherein the system of partial differential equations relates the cross-region coupling constraint to the turbulent diffusion term, the thermodynamic parameter term, and the regional coupling term; Inputting the partial differential equations into the deep digital twin mapping network, establishing a mapping loss function including a physical constraint loss term and a conservation constraint loss term, and optimizing the parameters of the deep digital twin mapping network by minimizing the mapping loss function; A domain decomposition method is used to solve the output results of the deep digital twin mapping network. The physical parameters of the transition region are fixed based on the distribution characteristics of the turbulent diffusion term and the thermodynamic parameter term to solve the core region. The updated physical parameters of the core region are used as boundary conditions to solve the physical parameters of the transition region. The process is iterated until the physical parameters of the core region and the transition region meet the preset convergence conditions. The physical parameters include velocity field, density field, turbulent kinetic energy, turbulent diffusion coefficient, dissipation rate, temperature field, pressure field, specific internal energy, and thermal conductivity. The cloud system near-ground data and the cloud system top data are received as real-time digital twin data, a Kalman gain matrix is constructed based on the prediction error covariance of the physical parameters of the core area and the transition area and the observation error covariance of the real-time digital twin data, error analysis is performed between the real-time digital twin data and the solution result of the regional decomposition method, and the Kalman gain matrix is used to dynamically correct the cloud system evolution prediction result.

6. The method according to claim 5, characterized in that Inputting the partial differential equations into the deep digital twin mapping network, establishing a mapping loss function including a physical constraint loss term and a conservation constraint loss term, and optimizing the parameters of the deep digital twin mapping network by minimizing the mapping loss function includes the following steps: The physical constraint loss term includes turbulent dynamics loss and thermodynamic coupling loss. The turbulent dynamics loss is constructed based on the spatiotemporal derivatives of turbulent kinetic energy and dissipation rate and the turbulence generation term. The weight coefficient of the turbulent dynamics loss is adaptively adjusted by the local turbulence intensity and turbulent Reynolds number. The thermodynamic coupling loss is constructed based on the convection-diffusion equation of the temperature field and pressure field. The weight coefficient of the thermodynamic coupling loss is dynamically determined by the nonlinear function of the temperature gradient and the pressure gradient. Constructing the conservation constraint loss term based on the cross-region coupling constraint, wherein the conservation constraint loss term includes a mass conservation constraint, a momentum conservation constraint, and an energy conservation constraint, and introducing a stress tensor correction term based on vorticity correlation; Constructing a feature-driven combined loss function, wherein the combined loss function performs a weighted combination of the physical constraint loss term and the conservation constraint loss term, wherein the weight coefficient of the combined loss function is dynamically adjusted according to the cloud morphological complexity, temperature gradient, and pressure gradient. The cloud morphological complexity is calculated by a weighted combination of profile curvature variance and texture entropy. A multi-scale residual correction mechanism is introduced, which includes local residuals and global residuals. The local residuals are calculated based on turbulence characteristic quantities, and the global residuals are calculated based on thermodynamic quantities. The weight coefficients of the local residuals and the global residuals are dynamically determined by the ratio of the turbulence characteristic scale to the cloud characteristic scale. The parameters of the deep digital twin mapping network are optimized and updated by minimizing the combined loss function and combining the multi-scale residual correction results, wherein the learning rate is adaptively adjusted based on the gradient of physical quantities and the prediction error. The physical quantities include velocity field, vorticity field, pressure field, density field, temperature field, turbulent kinetic energy and dissipation rate. At the same time, a periodic attenuation mechanism is introduced. At the end of each training cycle, the learning rate is adjusted in stages according to the degree of satisfaction of the physical constraint loss term and the conservation constraint loss term.

7. The method according to claim 5, characterized in that The steps of receiving the cloud system near-ground data and the cloud system top data as real-time digital twin data, constructing a Kalman gain matrix based on the prediction error covariance of physical parameters of the core area and the transition area and the observation error covariance of the real-time digital twin data, performing error analysis on the real-time digital twin data and the solution result of the regional decomposition method, and dynamically correcting the cloud system evolution prediction result using the Kalman gain matrix include: Acquire cloud near-surface data including reflectivity, radial velocity, and spectral width, fuse multi-site radar data using a three-dimensional variational assimilation method, construct the three-dimensional wind, temperature, and water vapor field distributions near the surface, and determine the ground-based observation error covariance matrix based on historical observation statistics; acquire cloud top data including cloud top height, cloud top temperature, and cloud top radiation flux, extract cloud characteristic parameters using an infrared-visible dual-channel inversion algorithm, and obtain cloud motion information through multi-time sequence registration. Determine the satellite observation error covariance matrix through channel noise analysis; Based on the difference prediction errors between the predicted results of the physical parameters in the core area and the transition area and the corresponding true values, a prediction error covariance matrix is constructed, and a local correlation constraint based on the spatial distance of the physical parameters is introduced to correct the prediction error covariance matrix. The local correlation constraint is determined by the correlation length scale; A Kalman gain matrix is constructed based on the prediction error covariance matrix, the ground-based observation error covariance matrix, and the satellite observation error covariance matrix, and the Kalman gain matrix is corrected by introducing an adaptive adjustment mechanism based on cloud morphological complexity, where the cloud morphological complexity is calculated by a weighted combination of profile curvature variance and texture entropy; The cloud system near-ground data and the cloud system top data are mapped to the observation space, and the observation residual vector with the predicted value is calculated. The predicted state is updated based on the observation residual vector and the modified Kalman gain matrix; the turbulence characteristic constraint, the mass conservation constraint and the energy conservation constraint are introduced into the updated predicted state, and the turbulence characteristic constraint is constructed based on the non-negativity of turbulent kinetic energy and dissipation rate; the constrained predicted state is used as the initial condition for the prediction at the next moment, forming a closed-loop iterative process of dynamic correction.

8. A multi-scale cloud system dynamic evolution simulation modeling system based on digital twins, used to implement the method according to any one of claims 1 to 7, characterized in that: include: The first unit is used to build a multi-source collaborative sensing network using multispectral cameras, lidar, and millimeter-wave radar to obtain cloud near-ground data, combined with satellite remote sensors to obtain cloud top data and ground-based meteorological stations to obtain meteorological element data. Phase consistency analysis is performed on the acquired multi-source data, and a spatial transformation matrix is established between the sensor coordinate systems to obtain registered data. The registered data is then quality-assessed based on signal-to-noise ratio, spatial resolution, and temporal consistency to select valid data that meets the assessment threshold. The second unit is used to extract cloud primitive features based on the valid data using a recursive segmentation algorithm, calculate the extreme points of the reflectivity gradient and the temperature gradient to divide the core area of the cloud cluster, and calculate the parameter attenuation curve around the core area to divide the transition area; based on the regional division results, extract the contour curvature, grayscale distribution, texture characteristics and multi-scale edge features of the cloud system, construct a deep digital twin mapping network between the cloud primitive features and the real cloud system, perform feature verification on the morphological feature parameters output by the deep digital twin mapping network and the measured meteorological data, and dynamically optimize the parameters of the deep digital twin mapping network based on the verification results; The third unit is used to finely discretize the core area using an adaptive multi-grid method. The grid scale is automatically adjusted with the cloud turbulence intensity and thermal parameter gradient, and a non-uniform grid is established for the transition area to consider the evolution of vortices within the cloud system. Based on the grid division results, the dynamic evolution of the cloud system is simulated and modeled, and cross-regional coupling constraints including mass conservation, momentum conservation, and energy conservation are established. The dynamic evolution process of the cloud system is characterized as a set of partial differential equations including turbulent diffusion, thermodynamic parameters, and regional coupling. The set of partial differential equations is integrated with the deep digital twin mapping network, and the regional decomposition method is used to solve the overall evolution characteristics of the cloud system. The evolution prediction results are dynamically corrected using real-time digital twin data.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images