A cloud microphysical parameter inversion method based on large model to accelerate optimal estimation
By combining the libRadtran and TabPFN models with a large-scale accelerated optimal estimation method, the problems of low efficiency and insufficient accuracy in acquiring cloud microphysical parameters in existing technologies are solved, achieving efficient and accurate cloud parameter inversion, which is applicable to fields such as cloud-radiation interaction research and solar energy resource assessment.
Patent Information
- Application Number
- CN202611219231.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-12
- Publication Date
- 2026-11-17
AI Technical Summary
Existing cloud microphysical parameter acquisition technologies suffer from low spatial resolution, high computational complexity, and difficulty in meeting the requirements for high temporal resolution and local cloud feature inversion.
A method based on large models to accelerate optimal estimation is adopted. Combining the libRadtran radiative transfer model and the TabPFN model, cloud optical thickness and effective cloud particle size are jointly inverted using surface solar radiation observations, MERRA-2 atmospheric prior information, and Himawari-9 cloud products, thereby reducing redundant radiative transfer calculations and iterative solutions.
It improves the efficiency and accuracy of cloud microphysical parameter inversion, can accurately capture local cloud radiation changes, and provides high temporal resolution cloud parameter datasets, suitable for multiple application scenarios.
Smart Images

Figure CN123116565A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of cloud microphysical parameter prediction technology, specifically involving a cloud microphysical parameter inversion method based on large model to accelerate optimal estimation. Background Technology
[0002] Existing techniques for acquiring cloud microphysical parameters mainly rely on methods such as satellite remote sensing inversion, atmospheric reanalysis data, and ground-based solar radiation observation physical inversion. Common technical solutions include: using meteorological satellite observation data to invert and obtain cloud microphysical parameters such as cloud optical thickness and effective cloud particle size; directly obtaining cloud parameters and atmospheric state information using atmospheric reanalysis data such as MERRA-2; and inverting cloud microphysical parameters based on ground-based solar radiation observation data, combined with optimal estimation (OE) and radiative transfer models. These methods typically establish inversion models based on the physical relationship between cloud microphysical characteristics and solar radiation, and combine radiative transfer calculations and iterative optimization to solve for cloud microphysical parameters. They have been widely used in atmospheric radiative transfer simulation, solar energy resource assessment, climate change research, and remote sensing applications.
[0003] While existing cloud microphysical parameter acquisition technologies can obtain cloud microphysical parameters such as cloud optical thickness and effective cloud particle size to a certain extent, they still have several technical defects and limitations. First, while existing satellite remote sensing-based inversion methods can acquire large-scale cloud microphysical parameter information, they are easily affected by factors such as spatial resolution, satellite observation geometry, and three-dimensional radiation effects. A scale mismatch exists between satellite observations and ground-based observations, making it difficult to accurately reflect local, high-temporal-resolution cloud microphysical characteristics. Second, while existing methods based on atmospheric reanalysis data such as MERRA-2 can provide continuous atmospheric state information and cloud parameter products, their cloud products only reflect the grid-scale average state, resulting in low spatial resolution. This makes it difficult to describe the rapid changes in local cloud fields and cannot meet the needs of high-temporal-resolution solar radiation simulation and cloud microphysical parameter inversion. Furthermore, while existing physical inversion methods based on ground-based solar radiation observation and optimal estimation (OE) have good physical consistency and inversion accuracy, their parameter solution process requires repeated calculations of radiative transfer and Jacobian matrix, resulting in high computational complexity and low computational efficiency. This makes it difficult to meet the practical needs of rapid processing of long-term, high-temporal-resolution ground-based solar radiation observation data and inversion of cloud microphysical parameters. Summary of the Invention
[0004] The problem this invention aims to solve is to improve the efficiency of cloud microphysical parameter inversion, and proposes a cloud microphysical parameter inversion method based on large model to accelerate optimal estimation.
[0005] To achieve the above objectives, the present invention provides the following technical solution:
[0006] A cloud microphysical parameter inversion method based on large model to accelerate optimal estimation includes the following steps:
[0007] S1. Obtain one-minute surface irradiance observations from the Baseline Surface Radiation Observation Network (BSRN), MERRA-2 atmospheric prior information that is spatiotemporally matched with ground-based observations, and Himawari-9 secondary cloud products that are spatiotemporally matched with ground-based observations. After preprocessing, obtain the dataset.
[0008] S2. A forward radiative transfer model is established using libRadtran. The surface solar radiation observation vector is set as the output of the model. The state vector to be inverted includes cloud optical thickness and cloud effective particle size. The background parameter vector includes solar zenith angle, cloud phase, aerosol optical thickness, Ångström index, water vapor, ozone, surface air pressure and surface albedo.
[0009] S3. Using the optimal estimation method, under the joint constraints of BSRN surface irradiance observations and MERRA-2 atmospheric prior information, cloud optical thickness and effective cloud particle size are jointly inverted to obtain the optimal inversion results;
[0010] S4. The TabPFN model is used to learn the nonlinear mapping relationship between the observed surface solar radiation, MERRA-2 atmospheric prior information, background parameter vector and the optimal inversion result, and the trained TabPFN model is used for cloud microphysical parameter inversion.
[0011] Furthermore, in step S1, the one-minute surface irradiance observations provided by the Baseline Surface Radiation Observation Network (BSRN) include global horizontal irradiance (GHI), direct normal irradiance (BNI), and diffuse horizontal irradiance (DHI). First, the raw observation data undergoes quality control to remove outliers and missing samples. Then, the Bright-Sun clear sky identification method is used to exclude clear sky samples and retain the observation data under cloud conditions.
[0012] Furthermore, in step S1, the prior atmospheric information of MERRA-2 includes total cloud cover, liquid water path, ice water path, aerosol optical thickness, Ångström index, water vapor, ozone, surface pressure, and surface albedo; samples with a total cloud cover of not less than 0.8 are retained, and liquid clouds and ice clouds are classified according to the relative proportions of liquid water and ice water.
[0013] Furthermore, in step S1, the Himawari-9 secondary cloud product includes cloud optical thickness, cloud effective particle size, and cloud phase.
[0014] Furthermore, in step S2, the expression for the forward radiative transfer model established using libRadtran is as follows:
[0015]
[0016] in, This is the observed vector of solar radiation on the Earth's surface. Let be the state vector to be inverted. For the background parameter vector, For the error term, This is a forward radiative transfer model built on libRadtran, used to invert the state vector. and background parameter vector Calculate the corresponding surface solar irradiance;
[0017]
[0018] in, For cloud optical thickness, Let T be the effective particle size of the cloud, and T be the transpose function.
[0019] Furthermore, the specific implementation method of step S3 includes the following steps:
[0020] S3.1. An optimal estimation method is used to jointly invert cloud optical thickness and effective cloud particle size. The optimal state vector is obtained by minimizing the objective function. The expression is:
[0021]
[0022] in, For the MERRA-2 prior state vector, The observation error covariance matrix, The prior error covariance matrix;
[0023] Represented as:
[0024]
[0025] in, For the prior values of MERRA-2 cloud optical thickness to match the spatiotemporal characteristics of ground-based observations, The prior value of the effective particle size of MERRA-2 clouds is used to match the spatiotemporal characteristics of ground-based observations;
[0026] S3.2. Use the MERRA-2 prior state vector as the initial state vector for iteration. :
[0027]
[0028] In the In the iteration, the th The state vector of the next iteration and background parameter vector Input the forward radiative transfer model to obtain the corresponding surface solar irradiance, and calculate and compare it with the BSRN measured irradiance. The residual between them is expressed as:
[0029]
[0030] in, For the first The irradiance residual vector of the next iteration;
[0031] S3.3. Calculate the Jacobian matrix of the forward radiative transfer model output relative to the state vector, and use the Jacobian matrix for iterative solution:
[0032]
[0033] in, Indicates the first The state vector of the next iteration The Jacobian matrix is calculated at point 1; based on the Jacobian matrix, irradiance residuals, and prior constraints, the Levenberg–Marquardt method is used to calculate the 1st... The state vector correction in the next iteration :
[0034]
[0035] in, For the first Damping factor for the next iteration. It is the identity matrix;
[0036] Then, update the cloud microphysics parameters based on the state vector correction:
[0037]
[0038] in, For the first The state vector of +1 iterations;
[0039] S3.4. After each state vector update, the updated cloud optical thickness and cloud effective particle size are re-input into the forward radiative transfer model to calculate the corresponding surface solar irradiance, and the irradiance residual, Jacobian matrix and objective function value are recalculated until the preset convergence condition is met.
[0040] During the iteration process, the following boundary constraints are set for cloud optical thickness and effective cloud particle size:
[0041]
[0042] ;
[0043] Iteration stops when the objective function changes between two consecutive iterations satisfy the following equation:
[0044]
[0045] in, The preset convergence threshold for the objective function. For the first The state vector of the next iteration The corresponding objective function value, For the first The state vector of the next iteration The corresponding objective function value;
[0046] After the iteration terminates, the state vector at this point is taken as the optimal inversion result for the corresponding observed sample:
[0047]
[0048] in, This represents the state vector of the optimal inversion result. The cloud optical thickness is the optimal inversion result. The effective particle size of the cloud is the optimal inversion result.
[0049] Furthermore, the specific implementation method of step S4 includes the following steps:
[0050] S4.1. Combine the surface solar radiation observation vector, the MERRA-2 prior state vector, and the background parameter vector to construct the TabPFN input feature vector z:
[0051] ;
[0052] The optimal state vector obtained by the optimal estimation method is used as the target variable of TabPFN, enabling TabPFN to establish the input feature vector. With the optimal state vector The mapping relationship between them;
[0053] S4.2. Train the TabPFN model constructed in step S4.1;
[0054] Input feature vectors corresponding to multiple observation samples and the optimal state vector obtained through the optimal estimation method A supervised sample set is constructed, and this supervised sample set is input into the pre-trained TabPFN model to form the input feature vector. As input, with the optimal state vector Using the supervised samples as the target variable, TabPFN establishes a nonlinear mapping relationship between the input feature vector and the optimal state vector, thereby obtaining the TabPFN model for cloud microphysical parameter prediction.
[0055] S4.3. After completing the TabPFN model training, for the new input feature vector The TabPFN model enables direct acquisition of corresponding cloud microphysical parameter prediction results:
[0056]
[0057] in, The input feature vector for the new observation sample, These are the cloud microphysical parameters predicted by TabPFN.
[0058] The beneficial effects of this invention are:
[0059] This invention presents a cloud microphysical parameter inversion method based on large model-accelerated optimal estimation, overcoming the limitations of traditional physical inversion methods. This invention utilizes TabPFN to learn the nonlinear mapping relationship between input features and the optimal estimation inversion results, reducing repeated radiative transfer calculations and iterative solution processes, thereby improving the inversion efficiency of cloud microphysical parameters.
[0060] This invention presents a cloud microphysical parameter inversion method based on a large model to accelerate optimal estimation while maintaining the physical consistency of the inversion results. Instead of employing a simple empirical data fitting method, this invention uses the libRadtran radiative transfer model and the optimal estimation results as physical constraints, enabling TabPFN to learn a mapping relationship jointly established by surface solar radiation observations, atmospheric prior information, and cloud microphysical parameters. The inverted cloud optical thickness and effective cloud particle size can re-drive the radiative transfer model and reconstruct surface solar irradiance, ensuring that the inversion results conform to actual cloud-radiation interactions.
[0061] This invention presents a cloud microphysical parameter inversion method based on large-scale model-accelerated optimal estimation, improving the characterization capability of local cloud microphysical parameters. A scale mismatch exists between satellite cloud products and ground-based observations; MERRA-2 cloud products primarily reflect the average cloud state at a grid scale, making it difficult to accurately describe rapidly changing local cloud fields. This invention directly utilizes one-minute surface solar irradiance observations to constrain cloud microphysical states, enabling more accurate capture of local, instantaneous cloud radiation changes and compensating for the shortcomings of satellite products and reanalysis data in characterizing local cloud features.
[0062] This invention presents a cloud microphysical parameter inversion method based on large-scale model-accelerated optimal estimation, improving the accuracy and stability of radiation reconstruction. The invention is validated through radiation closure experiments under liquid cloud and ice cloud conditions. Under liquid cloud conditions, the coefficient of determination for radiation closure is 0.916, with a normalized mean deviation of −1.8%; under ice cloud conditions, the coefficient of determination is 0.798, with a normalized mean deviation of −1.3%. The radiation closure results obtained by this invention are superior to those obtained using MERRA-2 and Himawari-9 level-two cloud products, demonstrating better accuracy and stability.
[0063] This invention presents a cloud microphysical parameter inversion method based on large-scale model-accelerated optimal estimation, possessing long-term data processing and extended application capabilities. This invention can be used to process long-term, one-minute temporal resolution ground-based solar radiation observation data, constructing continuous, local, high temporal resolution datasets of cloud optical thickness and effective cloud particle size. Furthermore, this invention can be further applied to various scenarios such as cloud-radiation interaction research, solar energy resource assessment, satellite cloud product verification, regional climate research, and solar energy development and utilization.
[0064] This invention presents a cloud microphysical parameter inversion method based on accelerated optimal estimation using a large model. This method can be widely applied in fields such as cloud microphysical parameter inversion, solar radiation observation and analysis, atmospheric radiative transfer simulation, solar energy resource assessment, climate change research, remote sensing inversion, and satellite product verification. By fusing ground-based solar radiation observation data, atmospheric reanalysis data, and radiative transfer models, and combining this with the TabPFN accelerated optimal estimation inversion method, this invention can achieve rapid and high-precision inversion of cloud microphysical parameters such as cloud optical thickness and effective cloud particle size, providing reliable technical support for the construction of long-term, high-temporal-resolution cloud microphysical parameter datasets. In the field of solar energy resource assessment, this invention can improve the accuracy of surface solar radiation resource assessment; in the field of climate change research, it can provide fundamental data for cloud-radiation interaction and regional climate analysis; furthermore, this invention can also be applied to satellite remote sensing product verification, multi-source remote sensing data analysis, and new energy development and utilization, providing a scientific basis for atmospheric environmental monitoring and efficient solar energy utilization. Attached Figure Description
[0065] Figure 1 This is a flowchart of a cloud microphysical parameter inversion method based on large model-accelerated optimal estimation, as described in this invention.
[0066] Figure 2 For cloud optical thickness under liquid cloud and ice cloud conditions Heyun effective particle size The results of radiation closure and multi-source comparison are shown in the figure, where (a) is the radiation closure result of the inversion result of the present invention under liquid cloud conditions, (b) is the radiation calculation result of MERRA-2 cloud parameters under liquid cloud conditions, (c) is the radiation calculation result of Himawari-9 secondary cloud product cloud parameters under liquid cloud conditions, (d) is the radiation closure result of the inversion result of the present invention under ice cloud conditions, (e) is the radiation calculation result of MERRA-2 cloud parameters under ice cloud conditions, and (f) is the radiation calculation result of Himawari-9 secondary cloud product cloud parameters under ice cloud conditions. Detailed Implementation
[0067] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only for explaining the invention and are not intended to limit the invention; that is, the described specific embodiments are merely a part of the embodiments of the invention, and not all of them. The components of the specific embodiments of the invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations, and the invention may also have other embodiments.
[0068] Therefore, the following detailed description of specific embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected specific embodiments of the invention. All other specific embodiments obtained by those skilled in the art based on these specific embodiments without inventive effort are within the scope of protection of this invention.
[0069] To further understand the invention's content, features, and effects, the following specific embodiments are provided, along with accompanying drawings. Figure 1 -Appendix Figure 2 Detailed explanation is as follows:
[0070] Example 1:
[0071] A cloud microphysical parameter inversion method based on large model to accelerate optimal estimation includes the following steps:
[0072] S1. Obtain one-minute surface irradiance observations from the Baseline Surface Radiation Observation Network (BSRN), MERRA-2 atmospheric prior information that is spatiotemporally matched with ground-based observations, and Himawari-9 secondary cloud products that are spatiotemporally matched with ground-based observations. After preprocessing, obtain the dataset.
[0073] Furthermore, in step S1, the one-minute surface irradiance observations provided by the Baseline Surface Radiation Observation Network (BSRN) include global horizontal irradiance (GHI), direct normal irradiance (BNI), and diffuse horizontal irradiance (DHI). First, the raw observation data undergoes quality control to remove outliers and missing samples. Then, the Bright-Sun clear sky identification method is used to exclude clear sky samples and retain the observation data under cloud conditions.
[0074] Furthermore, in step S1, the prior atmospheric information of MERRA-2 includes total cloud cover, liquid water path, ice water path, aerosol optical thickness, Ångström index, water vapor, ozone, surface pressure, and surface albedo; samples with a total cloud cover of not less than 0.8 are retained, and liquid clouds and ice clouds are classified according to the relative proportions of liquid water and ice water.
[0075] Furthermore, in step S1, the Himawari-9 secondary cloud product includes cloud optical thickness, cloud effective particle size, and cloud phase.
[0076] S2. A forward radiative transfer model is established using libRadtran. The surface solar radiation observation vector is set as the output of the model. The state vector to be inverted includes cloud optical thickness and cloud effective particle size. The background parameter vector includes solar zenith angle, cloud phase, aerosol optical thickness, Ångström index, water vapor, ozone, surface air pressure and surface albedo.
[0077] Furthermore, in step S2, the expression for the forward radiative transfer model established using libRadtran is as follows:
[0078]
[0079] in, This is the observed vector of solar radiation on the Earth's surface. Let be the state vector to be inverted. For the background parameter vector, For the error term, This is a forward radiative transfer model built on libRadtran, used to invert the state vector. and background parameter vector Calculate the corresponding surface solar irradiance;
[0080]
[0081] in, For cloud optical thickness, Let T be the effective particle size of the cloud, and T be the transpose function.
[0082] Furthermore, the expression for the surface solar radiation observation vector is:
[0083]
[0084] In the formula, , and These represent global horizontal irradiance, direct normal irradiance, and diffuse horizontal irradiance, respectively. The background parameter vector consists of known parameters from the first part that match the temporal and spatial characteristics of BSRN observations, including solar zenith angle, cloud facies, aerosol optical thickness, Ångström index, water vapor, ozone, surface pressure, and surface albedo. These background parameters remain unchanged during a single inversion iteration. This is the error term, which includes errors in surface solar radiation observations and forward radiative transfer modeling.
[0085] S3. Using the optimal estimation method, under the joint constraints of BSRN surface irradiance observations and MERRA-2 atmospheric prior information, cloud optical thickness and effective cloud particle size are jointly inverted to obtain the optimal inversion results;
[0086] Furthermore, the specific implementation method of step S3 includes the following steps:
[0087] S3.1. An optimal estimation method is used to jointly invert cloud optical thickness and effective cloud particle size. The optimal state vector is obtained by minimizing the objective function. The expression is:
[0088]
[0089] in, For the MERRA-2 prior state vector, The observation error covariance matrix, is the prior error covariance matrix; the first term of the objective function is the observation mismatch term, used to constrain the difference between the irradiance calculated by libRadtran and the measured irradiance of BSRN; the second term is the prior constraint term, used to limit the inversion state vector from deviating too much from the MERRA-2 prior state vector, thereby improving the stability and convergence of the inversion process.
[0090] Represented as:
[0091]
[0092] in, For the prior values of MERRA-2 cloud optical thickness to match the spatiotemporal characteristics of ground-based observations, The prior value of the effective particle size of MERRA-2 clouds is used to match the spatiotemporal characteristics of ground-based observations;
[0093] S3.2. Use the MERRA-2 prior state vector as the initial state vector for iteration. :
[0094]
[0095] In the In the iteration, the th The state vector of the next iteration and background parameter vector Input the forward radiative transfer model to obtain the corresponding surface solar irradiance, and calculate and compare it with the BSRN measured irradiance. The residual between them is expressed as:
[0096]
[0097] in, For the first The irradiance residual vector of the next iteration;
[0098] S3.3. Calculate the Jacobian matrix of the forward radiative transfer model output relative to the state vector, and use the Jacobian matrix for iterative solution:
[0099]
[0100] in, Indicates the first The state vector of the next iteration The Jacobian matrix is calculated at point 1; based on the Jacobian matrix, irradiance residuals, and prior constraints, the Levenberg–Marquardt method is used to calculate the 1st... The state vector correction in the next iteration This matrix is used to quantify the sensitivity of global horizontal irradiance, direct normal irradiance, and diffuse horizontal irradiance calculated by libRadtran to changes in cloud optical thickness and effective cloud particle size, and to provide gradient information required for iterative updates of the Levenberg–Marquardt numerical optimization method.
[0101]
[0102] in, For the first Damping factor for the next iteration. It is the identity matrix;
[0103] Then, update the cloud microphysics parameters based on the state vector correction:
[0104]
[0105] in, For the first The state vector of +1 iterations;
[0106] S3.4. After each state vector update, the updated cloud optical thickness and cloud effective particle size are re-input into the forward radiative transfer model to calculate the corresponding surface solar irradiance, and the irradiance residual, Jacobian matrix and objective function value are recalculated until the preset convergence condition is met.
[0107] During the iteration process, the following boundary constraints are set for cloud optical thickness and effective cloud particle size:
[0108]
[0109] ;
[0110] Iteration stops when the objective function changes between two consecutive iterations satisfy the following equation:
[0111]
[0112] in, The preset convergence threshold for the objective function. For the first The state vector of the next iteration The corresponding objective function value, For the first The state vector of the next iteration The corresponding objective function value;
[0113] After the iteration terminates, the state vector at this point is taken as the optimal inversion result for the corresponding observed sample:
[0114]
[0115] in, This represents the state vector of the optimal inversion result. The cloud optical thickness is the optimal inversion result. The effective particle size of the cloud is the optimal inversion result.
[0116] S4. The TabPFN model is used to learn the nonlinear mapping relationship between the observed surface solar radiation, MERRA-2 atmospheric prior information, background parameter vector and the optimal inversion result, and the trained TabPFN model is used for cloud microphysical parameter inversion.
[0117] Furthermore, the specific implementation method of step S4 includes the following steps:
[0118] S4.1. Combine the surface solar radiation observation vector, the MERRA-2 prior state vector, and the background parameter vector to construct the TabPFN input feature vector z:
[0119] ;
[0120] The optimal state vector obtained by the optimal estimation method is used as the target variable of TabPFN, enabling TabPFN to establish the input feature vector. With the optimal state vector The mapping relationship between them;
[0121] S4.2. Train the TabPFN model constructed in step S4.1;
[0122] Input feature vectors corresponding to multiple observation samples and the optimal state vector obtained through the optimal estimation method A supervised sample set is constructed, and this supervised sample set is input into the pre-trained TabPFN model to form the input feature vector. As input, with the optimal state vector Using the supervised samples as the target variable, TabPFN establishes a nonlinear mapping relationship between the input feature vector and the optimal state vector, thereby obtaining the TabPFN model for cloud microphysical parameter prediction.
[0123] S4.3. After completing the TabPFN model training, for the new input feature vector The TabPFN model enables direct acquisition of corresponding cloud microphysical parameter prediction results:
[0124]
[0125] in, The input feature vector for the new observation sample, These are the cloud microphysical parameters predicted by TabPFN.
[0126] The above method allows for the rapid acquisition of cloud microphysical parameters without repeatedly performing the complete libRadtran forward simulation and optimal estimation iteration for each new observation sample. This improves the processing efficiency of high temporal resolution observation data while maintaining the physical constraints of the optimal estimation inversion.
[0127] The cloud optical thickness obtained by inversion in this embodiment Heyun effective particle size The MERRA-2 prior cloud parameters and the cloud parameters from the Himawari-9 secondary cloud product were input into the libRadtran radiative transfer model to calculate the global horizontal irradiance (GHI) and diffuse horizontal irradiance (DHI), and then compared with the BSRN measured irradiance. Under liquid cloud conditions, the coefficient of determination of the radiative closure result in this embodiment was determined. The coefficient of determination is 0.916, the normalized mean deviation is -1.8%, and the normalized root mean square error is 23.5%. Under ice cloud conditions, the coefficient of determination is... The value is 0.798, the normalized mean deviation is −1.3%, and the normalized root mean square error is 34.1%. Figure 2 It can be seen that, compared with the irradiance calculated using MERRA-2 prior cloud parameters and Himawari-9 secondary cloud products, the GHI and DHI calculated in this embodiment have higher consistency with the BSRN measured results, indicating that the cloud optical thickness and cloud effective particle size obtained by inversion in this embodiment can accurately characterize the local cloud-radiation state.
[0128] It should be noted that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0129] Although this application has been described above with reference to specific embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of this application. In particular, as long as there is no structural conflict, the features in the specific embodiments disclosed in this application can be combined with each other in any way. The lack of an exhaustive description of these combinations in this specification is merely for the sake of brevity and resource conservation. Therefore, this application is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.
Claims
1. A cloud microphysical parameter inversion method based on large model-accelerated optimal estimation, characterized in that, Includes the following steps: S1. Obtain one-minute surface irradiance observations from the Baseline Surface Radiation Observation Network (BSRN), MERRA-2 atmospheric prior information that is spatiotemporally matched with ground-based observations, and Himawari-9 secondary cloud products that are spatiotemporally matched with ground-based observations. After preprocessing, obtain the dataset. S2. A forward radiative transfer model is established using libRadtran. The surface solar radiation observation vector is set as the output of the model. The state vector to be inverted includes cloud optical thickness and cloud effective particle size. The background parameter vector includes solar zenith angle, cloud phase, aerosol optical thickness, Ångström index, water vapor, ozone, surface air pressure and surface albedo. S3. Using the optimal estimation method, under the joint constraints of BSRN surface irradiance observations and MERRA-2 atmospheric prior information, cloud optical thickness and effective cloud particle size are jointly inverted to obtain the optimal inversion results; S4. The TabPFN model is used to learn the nonlinear mapping relationship between the observed surface solar radiation, MERRA-2 atmospheric prior information, background parameter vector and the optimal inversion result, and the trained TabPFN model is used for cloud microphysical parameter inversion.
2. The cloud microphysical parameter inversion method based on large model accelerated optimal estimation according to claim 1, characterized in that, In step S1, the one-minute surface irradiance observations provided by the Baseline Surface Radiation Observation Network (BSRN) include global horizontal irradiance (GHI), direct normal irradiance (BNI), and diffuse horizontal irradiance (DHI). First, the raw observation data undergoes quality control to remove outliers and missing samples. Then, the Bright-Sun clear sky identification method is used to exclude clear sky samples and retain the observation data under cloud conditions.
3. The cloud microphysical parameter inversion method based on large model accelerated optimal estimation according to claim 2, characterized in that, In step S1, the prior atmospheric information of MERRA-2 includes total cloud cover, liquid water path, ice water path, aerosol optical thickness, Ångström index, water vapor, ozone, surface pressure, and surface albedo; samples with a total cloud cover of not less than 0.8 are retained, and liquid clouds and ice clouds are classified according to the relative proportions of liquid water and ice water.
4. The cloud microphysical parameter inversion method based on large model accelerated optimal estimation according to claim 3, characterized in that, The Himawari-9 secondary cloud products in step S1 include cloud optical thickness, cloud effective particle size, and cloud phase.
5. The cloud microphysical parameter inversion method based on large model accelerated optimal estimation according to claim 4, characterized in that, Step S2 uses libRadtran to establish the expression for the forward radiative transfer model, which is: in, This is the observed vector of solar radiation on the Earth's surface. Let be the state vector to be inverted. For the background parameter vector, For error terms, This is a forward radiative transfer model built on libRadtran, used to invert the state vector. and background parameter vector Calculate the corresponding surface solar irradiance; in, For cloud optical thickness, Let T be the effective particle size of the cloud, and T be the transpose function.
6. The cloud microphysical parameter inversion method based on large model accelerated optimal estimation according to claim 5, characterized in that, The specific implementation method of step S3 includes the following steps: S3.
1. An optimal estimation method is used to jointly invert cloud optical thickness and effective cloud particle size. The optimal state vector is obtained by minimizing the objective function. The expression is: in, For the MERRA-2 prior state vector, The observation error covariance matrix, The prior error covariance matrix; Represented as: in, For the prior values of MERRA-2 cloud optical thickness to match the spatiotemporal characteristics of ground-based observations, The prior value of the effective particle size of MERRA-2 clouds is used to match the spatiotemporal characteristics of ground-based observations; S3.
2. Use the MERRA-2 prior state vector as the initial state vector for iteration. : In the In the next iteration, the first The state vector of the next iteration and background parameter vector Input the forward radiative transfer model to obtain the corresponding surface solar irradiance, and calculate and compare it with the BSRN measured irradiance. The residual between them is expressed as: in, For the first The irradiance residual vector of the next iteration; S3.
3. Calculate the output of the forward radiative transfer model relative to the state vector. The Jacobian matrix is obtained by iterative solution using the Jacobian matrix: in, Indicates the first The state vector of the next iteration The Jacobian matrix is calculated at point 1; based on the Jacobian matrix, irradiance residuals, and prior constraints, the Levenberg–Marquardt method is used to calculate the 1st... The state vector correction in the next iteration : in, For the first Damping factor for the next iteration. It is the identity matrix; Then, update the cloud microphysics parameters based on the state vector correction: in, For the first The state vector of +1 iterations; S3.
4. After each state vector update, the updated cloud optical thickness and cloud effective particle size are re-input into the forward radiative transfer model to calculate the corresponding surface solar irradiance, and the irradiance residual, Jacobian matrix and objective function value are recalculated until the preset convergence condition is met. During the iteration process, the following boundary constraints are set for cloud optical thickness and effective cloud particle size: ; Iteration stops when the objective function changes between two consecutive iterations satisfy the following equation: in, The preset convergence threshold for the objective function. For the first The state vector of the next iteration The corresponding objective function value, For the first The state vector of the next iteration The corresponding objective function value; After the iteration terminates, the state vector at this point is taken as the optimal inversion result for the corresponding observed sample: in, This represents the state vector of the optimal inversion result. The cloud optical thickness is the optimal inversion result. The effective particle size of the cloud is the optimal inversion result.
7. The cloud microphysical parameter inversion method based on large model accelerated optimal estimation according to claim 6, characterized in that, The specific implementation method of step S4 includes the following steps: S4.
1. Combine the surface solar radiation observation vector, the MERRA-2 prior state vector, and the background parameter vector to construct the TabPFN input feature vector z: ; The optimal state vector obtained by the optimal estimation method is used as the target variable of TabPFN, enabling TabPFN to establish the input feature vector. With the optimal state vector The mapping relationship between them; S4.
2. Train the TabPFN model constructed in step S4.1; Input feature vectors corresponding to multiple observation samples and the optimal state vector obtained through the optimal estimation method A supervised sample set is constructed, and this supervised sample set is input into the pre-trained TabPFN model to form the input feature vector. As input, with the optimal state vector Using the supervised samples as the target variable, TabPFN establishes a nonlinear mapping relationship between the input feature vector and the optimal state vector, thereby obtaining the TabPFN model for cloud microphysical parameter prediction. S4.
3. After completing the TabPFN model training, for the new input feature vector The TabPFN model enables direct acquisition of corresponding cloud microphysical parameter prediction results: in, The input feature vector for the new observation sample, These are the cloud microphysical parameters predicted by TabPFN.
