A method for estimating downlink shortwave radiation from cloud-to-ground surface

Abstract

The present invention provides a method for estimating downward shortwave radiation from a cloud-space surface, comprising: S1. collecting data and constructing a sample data set; S2. constructing a multi-task learning model with physical constraints to obtain estimated values ​​of two key parameters in the cloud-space radiation estimation model; S3. based on the cloud-space radiation estimation method in the physical model FARMS, combined with the key parameters estimated by the deep learning model, obtaining the estimated value of downward shortwave radiation from the cloud-space surface; S4. based on the estimated value of downward shortwave radiation from the cloud-space surface and the ground observation value, calculating the loss function, adjusting the parameters of the multi-task learning model with physical constraints by back propagation of the loss value, and training to obtain the optimal cloud-space surface downward shortwave radiation estimation model; and S5. based on the optimal cloud-space surface downward shortwave radiation estimation model obtained by training, estimating the estimated value of downward shortwave radiation from the cloud-space surface in the target area. The method of the present invention can improve the estimation accuracy of downward shortwave radiation from the surface under the cloud-space, and has good stability and interpretability.

Description

Technical Field

[0001] The present invention relates to the field of cloud-space surface downward shortwave radiation estimation, and in particular to a cloud-space surface downward shortwave radiation estimation method that couples deep learning and physical methods. Background Art

[0002] The downward shortwave radiation from the surface is a key parameter in the surface energy balance and a driving force for surface ecological and hydrological processes. In addition, it also plays an important role in atmospheric movement and water and heat exchange between the ground and the atmosphere. Therefore, the estimation of downward shortwave radiation from the surface has always been the focus of relevant research. The estimation method of downward shortwave radiation under clear skies is relatively mature, and the calculation speed and accuracy are satisfactory. However, under cloudy weather conditions, the transmission process of radiation in the atmosphere is more complicated, so the estimation method of downward shortwave radiation from the surface under cloudy skies still has a lot of room for improvement.

[0003] The estimation methods of downlink shortwave radiation from the surface can be roughly divided into physical methods and statistical methods. The former is based on the radiation transfer equation and has the advantages of high stability and strong interpretability, but it is often more complicated and difficult to implement. The deep learning method in the statistical method has high accuracy and is easy to implement when the sample size is sufficient, but the generalization and transferability are poor. Therefore, the combination of physical and deep learning models is expected to solve the problems of existing models and generate new methods that integrate the advantages of each model. However, different models have their own unique advantages and disadvantages. When coupling physical and deep learning models, it is necessary to retain the unique advantages of the current model and make targeted improvements to the unique problems of the model. However, the existing coupling model has not improved and combined the problems existing in the two types of models themselves, and cannot fully exert the advantages of the two types of models. The FARMS algorithm is an efficient and accurate method for estimating downlink shortwave radiation from the surface of the whole sky. For cloud sky, it uses a parameterized method to calculate the variables in the radiation transfer process. However, there are certain limitations in the ability to express certain physical processes with mathematical formulas, so there is a certain uncertainty in the calculation of certain parameters. The deep learning model has strong nonlinear fitting ability and can better fit the complex relationship between variables based on a large amount of data. The reasonable coupling of the two is expected to overcome the defects of their own models and build a cloud-to-ground downlink shortwave radiation estimation model that has the advantages of both models. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a method for estimating the downward shortwave radiation from the cloud-to-ground surface based on multi-task learning and the physical algorithm FARMS, so as to solve the problems that the existing estimation methods for the downward shortwave radiation from the cloud-to-ground surface are of low accuracy and poor interpretability, and the coupling model has limited performance improvement due to the lack of improvements on the defects of the model itself.

[0005] In order to solve the above technical problems, the technical solution of the present invention is:

[0006] The cloud-to-ground surface downward shortwave radiation estimation method of the present invention comprises the following steps: S1. collecting data and constructing a sample data set; S2. constructing a physically constrained multi-task learning model to obtain estimated values ​​of two key parameters in the cloud-to-ground surface downward shortwave radiation estimation model; S3. obtaining the cloud-to-ground surface downward shortwave radiation estimation value based on the cloud-to-ground radiation estimation method in the physical model FARMS and the key parameter estimation values ​​obtained by the physically constrained multi-task learning model; S4. calculating the loss function based on the cloud-to-ground surface downward shortwave radiation estimation value and the cloud-to-ground surface downward shortwave radiation ground observation value, adjusting the parameters of the physically constrained multi-task learning model by back propagation of the loss value, and training to obtain the optimal cloud-to-ground surface downward shortwave radiation estimation model; and S5. estimating the cloud-to-ground surface downward shortwave radiation value of the target area based on the trained optimal cloud-to-ground surface downward shortwave radiation estimation model.

[0007] Optionally, in the above-mentioned cloud-to-ground surface downlink shortwave radiation estimation method, step S1 includes: collecting input variable data and collecting ground observation data; and based on the ground observation data, extracting pixel values ​​of corresponding time and place from each input variable image, matching them with the cloud-to-ground surface downlink radiation data provided by the ground observation data, and constructing a sample data set.

[0008] Optionally, in the above-mentioned cloud-space surface downlink shortwave radiation estimation method, step S2 includes: S2-1. constructing a multi-task learning model based on a fully connected neural network, the multi-task learning model first extracts common features contained in the input variables through two shared layers, and then constructs an independent task-specific layer for the target variable, respectively learns the unique features related to the two target parameters in the input variables, and finally outputs two key parameters of cloud optical thickness and adjustment scale factor when the cloud scattering transmittance reaches the maximum value; S2-2. adding physical constraints to the multi-task learning model by constructing a physical loss function based on physical laws, and the definition of the physical loss function is as follows:

[0009]

[0010] is the physical loss function, is the cloud effective particle radius, is the cosine of the solar zenith angle, is the cloud optical thickness corresponding to the maximum value of cloud scattering transmittance, For each Under different values The number of values, is the linear correction unit function, for a certain Under the value, two adjacent The difference is defined as:

[0011]

[0012] The function formula is as follows:

[0013] .

[0014] Optionally, in the above-mentioned cloud-space surface downlink shortwave radiation estimation method, step S3 includes:

[0015] S3-1. Calculate cloud scattering transmittance based on the cloud parameterization method in the FARMS model: Input the two key parameter estimates output by the multi-task learning model of physical constraints into the FARMS model, and calculate the cloud scattering transmittance based on the following formula:

[0016]

[0017] is the cloud scattering transmittance, is the cloud effective particle radius, is the cloud optical thickness, is the cosine of the solar zenith angle, It is adjustment The scaling factor, yes The cloud optical thickness corresponding to the maximum value is, is the cloud diffuse transmittance for each given solar zenith angle The maximum value of is related to the cloud scattering transmittance A function related to width, and The calculation formula is as follows:

[0018] For Water Cloud:

[0019]

[0020]

[0021] For Ice Cloud:

[0022]

[0023] ;

[0024] S3-2. Based on the cloud parameterization method in the FARMS model, the cloud direct transmittance is calculated using the following formula:

[0025]

[0026] is the cloud direct transmittance, is the cloud optical thickness, is the cosine of the solar zenith angle;

[0027] S3-3, based on the cloud parameterization method in the FARMS model, calculate the cloud reflectivity, the formula is as follows:

[0028] For Water Cloud:

[0029]

[0030] For Ice Cloud:

[0031]

[0032] is the cloud reflectivity, is the cloud optical thickness;

[0033] S3-4. Based on the cloud parameterization method in the FARMS model, the scattering transmittance of scattering through the cloud layer in the atmosphere is calculated as follows:

[0034]

[0035] is the scattering transmittance in the atmosphere through scattering through clouds, is the cosine of the zenith angle of a single sun beam, is the direct transmittance under clear sky;

[0036] S3-5. Based on the cloud parameterization method in the FARMS model and the variables obtained from steps S3-1 to S3-4, calculate the downlink shortwave radiation from the cloud-to-surface space using the following formula:

[0037]

[0038] is the downward shortwave radiation from the cloud surface, is the cosine of the solar zenith angle, is the radiation flux at the top of the atmosphere, is the cloud direct transmittance, is the cloud scattering transmittance, is the scattering transmittance in the atmosphere due to scattering through clouds, is the cloud reflectivity, is the clear sky direct transmittance, is the clear sky diffuse transmittance, is the albedo of the large balloon under clear sky.

[0039] Optionally, in the above-mentioned cloud-space surface downlink shortwave radiation estimation method, step S4 includes:

[0040] S4-1. Construct the loss function of the cloud-to-ground surface downward shortwave radiation estimation model: Based on the error of the cloud-to-ground surface downward shortwave radiation estimation value and the physical loss function in step S2-2, the loss function of the cloud-to-ground surface downward shortwave radiation estimation method is constructed, which is defined as follows:

[0041]

[0042] is the model loss function, Represents ground observations With model estimates The mean square error between is the general regularization term, are the solution parameters of the model, The physical loss function constructed for S2-2, and is a hyperparameter used to adjust the weights, is defined as follows:

[0043] ;

[0044] S4-2. The loss value of each training of the cloud-to-ground surface downward shortwave radiation estimation method is calculated based on the loss function, and is transferred back to the multi-task learning model with physical constraints to adjust its parameters. By minimizing the loss function, the optimal cloud-to-ground surface downward shortwave radiation estimation model is obtained.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] The method for estimating downward shortwave radiation from the cloud-space surface of the present invention uses a deep learning model to replace the original calculation method of some key parameters in the physical model, thereby improving the accuracy of intermediate variables and further improving the overall accuracy of the model, that is, improving the estimation accuracy of downward shortwave radiation from the cloud-space surface. At the same time, since the framework of the physical model is retained, the model also has high stability and interpretability. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the specific implementation of the present invention or the technical solution in the prior art, the drawings required for use in the specific implementation or the description of the prior art are briefly introduced below.

[0048] Figure 1 It is a flow chart of the steps of the method for estimating the downward shortwave radiation from the cloud-sky surface of the present invention;

[0049] Figure 2It is a flowchart of the method for estimating the downward shortwave radiation of the cloud-sky surface of the present invention;

[0050] Figure 3 It is a flow chart of the present invention coupling deep learning and physical methods to estimate the downlink shortwave radiation of the cloud-space surface;

[0051] Figure 4 is the accuracy of the model on different test datasets;

[0052] Figure 5 is the estimated result of the model in the test area. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical solution and advantages of the embodiments of the present invention more clear, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0054] The cloud-space surface downward shortwave radiation estimation method of the present invention comprises the following steps: S1. collecting data and constructing a sample data set; S2. constructing a multi-task learning model with physical constraints to obtain the estimated values ​​of two key parameters in the cloud-space radiation estimation model; S3. obtaining the cloud-space surface downward shortwave radiation estimation value based on the cloud-space radiation estimation method in the physical model FARMS, combined with the key parameter estimation values ​​obtained by the multi-task learning model with physical constraints (i.e., the deep learning model); S4. calculating the loss function based on the cloud-space surface downward shortwave radiation estimation value and the ground observation value, adjusting the parameters of the multi-task learning model with physical constraints by back propagation of the loss value, and training to obtain the optimal cloud-space surface downward shortwave radiation estimation model; and S5. estimating the cloud-space surface downward shortwave radiation value of the target area based on the optimal cloud-space surface downward shortwave radiation estimation model obtained by training. The present invention uses a deep learning model to replace the original calculation method of some key parameters in the physical model, thereby improving the accuracy of the intermediate variables and further improving the overall accuracy of the model. At the same time, since the framework of the physical model is retained, the model also has high stability and interpretability.

[0055] like Figure 1 , Figure 2 and Figure 3 As shown, the method for estimating downward shortwave radiation from the cloud-sky surface of the present invention specifically comprises the following steps:

[0056] S1. Collect data and build sample data sets. Specifically include:

[0057] S1-1. Collect input variable data, including aerosol optical depth and aerosol single scattering albedo data in MOD04, cloud optical depth, cloud effective particle radius, cloud phase, cloud fraction data in MOD06, solar zenith angle, surface pressure, ozone thickness, precipitable water data in MOD07, and surface albedo data in MCD43A3.

[0058] S1-2. Collect ground observation data, including shortwave downlink radiation observation data from 105 stations of the FLUXNET observation network in 2013, and shortwave downlink radiation observation data from 49 stations of the BSRN observation network in 2013. The former is used for model training, and the latter is used for model accuracy verification.

[0059] S1-3. Construct a sample data set. Based on the ground observation data, extract the pixel values ​​of the corresponding time and location from each input variable image, match them with the cloud-space surface downlink radiation data provided by the station data (ground observation data), and construct a sample data set.

[0060] S2. Build a multi-task learning model with physical constraints to obtain the estimated values ​​of two key parameters in the cloud-to-ground downlink shortwave radiation estimation model. Specifically include:

[0061] S2-1. Construct a multi-task learning model based on a fully connected neural network. There are two key parameters in the cloud-sky radiation estimation method in the physical model FARMS: the cloud optical thickness corresponding to the maximum value of the cloud scattering transmittance and the scaling factor used to adjust the cloud scattering transmittance. There is a certain uncertainty in the calculation method of these two parameters in FARMS, so a multi-task learning model is constructed based on a fully connected neural network to replace the original estimation method. The multi-task learning model first extracts the common features contained in the input variables through two shared layers, and then constructs an independent task-specific layer for the target variable, respectively learning the unique features related to the two target parameters in the input variables, and finally outputs the two key parameters of the cloud optical thickness and the adjustment scaling factor when the cloud scattering transmittance reaches the maximum value.

[0062] S2-2. Based on the laws of physics, physical constraints are added to the multi-task learning model by constructing a physical loss function, which reduces the search space for solving model parameters and accelerates the convergence of the model. One of the target parameters of the multi-task learning model is the cloud optical thickness corresponding to the maximum value of the cloud scattering transmittance. The following rule applies: when the cloud effective particle radius remains unchanged, it increases monotonically with the increase of the cosine value of the solar zenith angle. Therefore, based on the above physical laws, physical constraints are added to the multi-task learning model by constructing a physical loss function. The definition of the physical loss function is as follows:

[0063]

[0064] is the physical loss function, is the cloud effective particle radius, is the cosine of the solar zenith angle, is the cloud optical thickness corresponding to the maximum value of cloud scattering transmittance. For each Under different values The number of values. is the linear correction unit function. for a certain Under the value, two adjacent The difference is defined as:

[0065]

[0066] The function formula is as follows:

[0067] .

[0068] S3. Based on the cloud-to-air radiation estimation method in the physical model FARMS, combined with the key parameter estimation values ​​obtained by the multi-task learning model (i.e., deep learning model) with physical constraints, the estimated value of the downlink shortwave radiation from the cloud-to-air surface is obtained, such as Figure 3 Specifically including:

[0069] S3-1. Calculate cloud scattering transmittance based on the cloud parameterization method in the FARMS model. Figure 3 As shown in Figure 2, the estimated values ​​of two key parameters output by the multi-task learning model with physical constraints are input into the FARMS model to calculate the cloud scattering transmittance based on the following formula:

[0070]

[0071] is the cloud scattering transmittance, is the cloud effective particle radius, is the cloud optical thickness, is the cosine of the solar zenith angle, It is adjustment The scaling factor, yes The cloud optical thickness corresponding to the maximum value is, is the cloud diffuse transmittance for each given solar zenith angle The maximum value of is related to the cloud scattering transmittance Width-related functions. and The calculation formula is as follows:

[0072] For Water Cloud:

[0073]

[0074]

[0075] For Ice Cloud:

[0076]

[0077]

[0078] S3-2. Based on the FARMS model algorithm, calculate the cloud direct transmittance, the formula is as follows:

[0079]

[0080] is the cloud direct transmittance, is the cloud optical thickness, is the cosine of the solar zenith angle.

[0081] S3-3, based on the cloud parameterization method in the FARMS model, calculate the cloud reflectivity, the formula is as follows:

[0082] For Water Cloud:

[0083]

[0084] For Ice Cloud:

[0085]

[0086] is the cloud reflectivity, is the cloud optical thickness.

[0087] S3-4. Based on the cloud parameterization method in the FARMS model, the scattering transmittance of scattering through the cloud layer in the atmosphere is calculated as follows:

[0088]

[0089] is the scattering transmittance in the atmosphere through scattering through clouds, is the cosine of the zenith angle of a single sun beam, is the direct transmittance under clear sky.

[0090] S3-5. Based on the cloud parameterization method in the FARMS model and the variables obtained from S3-1 to S3-4 above, calculate the downlink shortwave radiation from the cloud-to-surface space using the following formula:

[0091]

[0092] is the downward shortwave radiation from the cloud surface, is the cosine of the solar zenith angle, is the radiation flux at the top of the atmosphere, is the cloud direct transmittance, is the cloud scattering transmittance, is the scattering transmittance in the atmosphere due to scattering through clouds, is the cloud reflectivity, is the clear sky direct transmittance, is the clear sky diffuse transmittance, is the albedo of the large balloon under clear sky.

[0093] S4. Based on the estimated value of the downward shortwave radiation from the cloud-to-ground surface and the ground observation value of the downward shortwave radiation from the cloud-to-ground surface, the loss function of the method for estimating the downward shortwave radiation from the cloud-to-ground surface is calculated, such as Figure 3 As shown in the figure, the parameters of the multi-task learning model with physical constraints are adjusted by back propagation of loss values, and the optimal cloud-to-ground downlink shortwave radiation estimation model (model training) is obtained through training. Specifically, it includes:

[0094] S4-1. Construct the loss function of the cloud-to-ground downward shortwave radiation estimation model. The loss function of the cloud-to-ground downward shortwave radiation estimation method is constructed based on the error of the cloud-to-ground downward shortwave radiation estimation value and the physical loss function in step S2-2, and is defined as follows:

[0095]

[0096] is the model loss function, Represents ground observations With model estimates Here we use the mean square error as the error between . is the general regularization term, are the solution parameters of the model. Physical loss function constructed for S2-2. and is a hyperparameter used to adjust the weights. is defined as follows:

[0097]

[0098] S4-2. Calculate the loss value of each training of the cloud-to-ground downward shortwave radiation estimation method based on the loss function, and transfer it back to the multi-task learning model with physical constraints to adjust its parameters. By minimizing the loss function, the optimal cloud-to-ground downward shortwave radiation estimation model is obtained.

[0099] S5. Model application: Based on the optimal cloud-to-ground downward shortwave radiation estimation model obtained through training, the cloud-to-ground downward shortwave radiation value of the target area is estimated.

[0100] Figure 4 The accuracy of the model on 30% of the independent validation sites of the FLUXNET observation network and the accuracy on the sites of the BSRN observation network that did not participate in the model training are shown. As can be seen from the figure, the accuracy of the model on both sets of validation sites is high, with R values ​​of 0.83 and 0.87, and rRMSE values ​​of 31.26% and 31.33%, respectively, which can meet the needs of actual applications. Figure 5 The estimation results of the model in two test areas are shown, namely the estimation results of the instantaneous surface downward shortwave radiation in the central and eastern part of China at 03:00 on November 30, 2013, and the estimation results of the instantaneous surface downward shortwave radiation in the central part of China at 03:10 on November 12, 2013. It can be seen from the figure that the estimation results can well reflect the spatial distribution of the surface downward shortwave radiation in the case of clouds, proving the effectiveness of the estimation method of the surface downward shortwave radiation in the cloud space.

[0101] By inputting the data of any target area into the optimal cloud-to-ground downward shortwave radiation estimation model obtained through training, the corresponding cloud-to-ground downward shortwave radiation estimation value can be obtained. Since the method of the present invention combines the deep learning model with the physical model, it has high accuracy and good transferability and generalization. Therefore, it can obtain high-precision estimation results for data in any region and time, and has high practical value.

[0102] The above embodiments are only specific implementation methods of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention is described in detail with reference to the above embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify or improve the technical solutions recorded in the above embodiments within the technical scope disclosed by the present invention, or replace some of the technical features therein by equivalents; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A method for estimating downward shortwave radiation from cloud-space surface, characterized in that: The following steps are involved: S1. Collect data and construct sample data sets; S2. Construct a multi-task learning model with physical constraints to obtain the estimated values ​​of two key parameters in the cloud-to-ground downlink shortwave radiation estimation model; S3. Based on the cloud-to-air radiation estimation method in the physical model FARMS, combined with the key parameter estimation values ​​obtained by the multi-task learning model of the physical constraints, the estimated value of the cloud-to-air surface downward shortwave radiation is obtained; S4. Based on the estimated value of the downward shortwave radiation from the cloud-space surface and the ground observation value of the downward shortwave radiation from the cloud-space surface, a loss function is calculated, and the parameters of the multi-task learning model of the physical constraint are adjusted by back propagation of the loss value to train an optimal cloud-space surface downward shortwave radiation estimation model; and S5. based on the optimal cloud-to-ground downward shortwave radiation estimation model obtained through training, estimate the cloud-to-ground downward shortwave radiation value of the target area; Wherein, step S2 comprises: S2-1. The multi-task learning model is constructed based on a fully connected neural network. The multi-task learning model first extracts common features contained in the input variables through two shared layers, then constructs independent task-specific layers for the target variables, respectively learns unique features related to the two target parameters in the input variables, and finally outputs two key parameters, namely, the cloud optical thickness and the adjustment scale factor when the cloud scattering transmittance reaches the maximum value; S2-2. Add physical constraints to the multi-task learning model by constructing a physical loss function based on physical laws. The physical loss function is defined as follows: L phy is the physical loss function, D e is the effective cloud particle radius, u is the cosine of the solar zenith angle, d cp is the cloud optical thickness corresponding to the maximum value of cloud scattering transmittance, n u For each D e The number of different u values ​​under the value, ReLU is the linear rectified unit function, Δd cp [D e ,i] is a certain D e Under this value, two adjacent d cp The difference is defined as: Δd cp [D e ,i]=d cp [D e ,μ i ]―d cp [D e ,μ i+1 ]u i <u i+1 The formula of the ReLU function is as follows: ReLU(x)=max(0,x).

2. The method for estimating downward shortwave radiation from cloud-space surface according to claim 1 is characterized in that: Step S1 includes: collecting input variable data and collecting ground observation data; and based on the ground observation data, extracting pixel values ​​of corresponding time and location from each input variable image, matching them with cloud-space surface downlink radiation data provided by the ground observation data, and constructing the sample data set.

3. The method for estimating downward shortwave radiation from cloud-space surface according to claim 1 is characterized in that: Step S3 includes: S3-1. Calculate the cloud scattering transmittance based on the cloud parameterization method in the FARMS model: Input the two key parameter estimates output by the multi-task learning model of the physical constraints into the FARMS model, and calculate the cloud scattering transmittance based on the following formula: is the cloud scattering transmittance, D e is the cloud effective particle radius, d c is the cloud optical thickness, μ is the cosine of the solar zenith angle, f(D e ,d c ,μ) is the adjustment The scaling factor, d cP (D e ,μ) is The cloud optical thickness corresponding to the maximum value is, is the cloud diffuse transmittance for each given solar zenith angle The maximum value of D(μ) is related to the cloud scattering transmittance A function related to width, The calculation formulas for D(μ) are as follows: For Water Cloud: D(μ)=-0.644531μ+1.20117+0.129807μ ―1 ―1.21096×10 ―3 μ ―2 +1.52587×10 ―7 μ ―3 For Ice Cloud: D(μ)=―0.0549531μ+0.617632+0.17876μ ―1 ―0.002174m ―2 ; S3-2. Based on the cloud parameterization method in the FARMS model, the cloud direct transmittance is calculated using the following formula: is the cloud direct transmittance, d c is the cloud optical thickness, μ is the cosine of the solar zenith angle; S3-3, based on the cloud parameterization method in the FARMS model, calculate the cloud reflectivity, the formula is as follows: For Water Cloud: For Ice Cloud: is the cloud reflectivity, d c is the cloud optical thickness; S3-4. Based on the cloud parameterization method in the FARMS model, the scattering transmittance of scattering through the cloud layer in the atmosphere is calculated as follows: is the scattered transmittance in the atmosphere through the clouds, μ0 is the cosine of the zenith angle of a single solar beam, is the clear sky direct transmittance; S3-5. Based on the cloud parameterization method in the FARMS model and the variables obtained from steps S3-1 to S3-4, calculate the downlink shortwave radiation from the cloud-to-surface space using the following formula: R cld is the downward shortwave radiation from the cloud surface, μ is the cosine value of the solar zenith angle, R0 is the radiation flux at the top of the atmosphere, is the cloud direct transmittance, is the cloud scattering transmittance, is the scattering transmittance in the atmosphere due to scattering through clouds, is the cloud reflectivity, is the clear sky direct transmittance, is the clear sky diffuse transmittance, is the albedo of the large balloon under clear sky.

4. The method for estimating downward shortwave radiation from cloud-space surface according to claim 1 is characterized in that: Step S4 includes: S4-1. Constructing the loss function of the cloud-to-ground surface downward shortwave radiation estimation model: Based on the error of the cloud-to-ground downward shortwave radiation estimation value and the physical loss function in step S2-2, the loss function of the cloud-to-ground surface downward shortwave radiation estimation method is constructed, which is defined as follows: L=L d (x true ―x pred )+αR(w)+βL phy (x pred ) L is the model loss function, L d (x true ―x pred ) represents the ground observation value x true With the model estimate x pred The mean square error between them, R(w) is the general regularization term, w is the solution parameter of the model, L phy (x pred ) is the physical loss function constructed for S2-2, α and β are hyperparameters used to adjust the weights, L d (x true ―x pred ) is defined as follows: S4-2. Calculate the loss value of each training of the cloud-to-ground surface downward shortwave radiation estimation method based on the loss function, transfer it back to the multi-task learning model with physical constraints, adjust its parameters, and obtain the optimal cloud-to-ground surface downward shortwave radiation estimation model by minimizing the loss function.

Images