Surface solar multiband radiation inversion method fusing physical information and deep learning

Abstract

The invention discloses an earth surface solar multiband radiation inversion method fusing physical information and deep learning, and belongs to the technical field of solar radiation prediction. The problem of remarkably reducing the calculation cost is solved. The method comprises the steps of selecting input features and output features; data acquisition is carried out, then data preprocessing is carried out, and all the preprocessed data are randomly divided into a training set, a verification set and a test set; constructing a multi-task deep learning model structure, wherein the multi-task deep learning model structure comprises an expert sub-network, a gating sub-network, an independent sub-network, a self-attention mechanism and physical constraints; training the multi-task deep learning model structure by using the normalized training set to obtain a trained multi-task deep learning model structure; according to the method, the calculation cost can be remarkably reduced, the physical consistency among different wavebands can be ensured, and reliable technical support is provided for scientific research and engineering application in related fields.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of solar radiation prediction, and particularly relates to a method for ground surface solar multi-band radiation inversion fusing physical information and deep learning. BACKGROUND

[0002] The energy system of modern society is still dominated by fossil fuels such as oil and coal. Although these traditional energies have provided strong impetus for economic development, they have also brought increasingly severe environmental challenges: the continuous rise in carbon dioxide emissions has led to global climate anomalies, the overexploitation of non-renewable resources has triggered an energy crisis, and the emission of various pollutants has exacerbated the deterioration of the ecological environment. Solar energy, with its clean and environmentally friendly and inexhaustible characteristics, has become a key pillar of renewable energy transformation, and its energy source is the eternal natural gift of solar radiation.

[0003] As an important part of the electromagnetic spectrum, solar radiation covers a broad range of wavelengths from 150 to 4000 nm. The energy of radiation in different bands shows significant differences: the ultraviolet region has sterilization and disinfection effects, the visible light region (400-700 nm) is not only the basis for human visual perception, but also the main energy source for plant photosynthesis (i.e., photosynthetically active radiation PAR); the infrared region mainly reflects thermal effects. This spectral characteristic makes solar radiation exhibit unique application value in multiple fields: in the field of architecture, the lighting effect of visible light, the thermal radiation characteristics of infrared rays, and the disinfection function of ultraviolet rays need to be considered; in agricultural production, PAR data is a key parameter for building crop growth models; the efficiency optimization of photovoltaic power generation systems cannot be achieved without accurate spectral response analysis; and the development of new building materials such as Low-e glass also needs detailed spectral data support.

[0004] Ground surface solar spectrum measurement equipment is expensive and complex to maintain, and conventional weather observations usually only provide wide-band irradiance data (285-2800 nm). Ground surface solar spectrum instruments can provide more accurate and comprehensive narrow-band irradiance information, but the measurement equipment is expensive and complex to maintain, and only a few ground radiation observation stations are equipped with such instruments. More importantly, ground observation stations, as point measurements, can only provide irradiance information at their locations and do not have spatial continuity, which makes it difficult to analyze areas where ground observation stations cannot be set up, such as urban buildings and islands.

[0005] Although semi-empirical models have a wider spatial coverage, their accuracy is poor due to the use of empirical models, especially their robustness under extreme sky conditions, which makes them unsuitable for applications that require more accurate spectral data.

[0006] Traditional machine learning methods are mostly single-task methods, that is, the output only includes narrow-band irradiance information, which requires modeling the solar irradiance on each target band separately, which greatly increases the complexity of the model and the consumption of computing resources. Meanwhile, as a data-driven modeling method, the output of the model may not meet the physical constraints and lacks physical consistency. SUMMARY

[0007] The problem to be solved by the present application is to significantly reduce the computing cost and ensure the physical consistency between different bands, and a method for retrieving ground surface solar multi-band radiation by fusing physical information and deep learning is proposed.

[0008] To achieve the above-mentioned purpose, the present application realizes the following technical solutions:

[0009] A method for retrieving ground surface solar multi-band radiation by fusing physical information and deep learning, comprising the following steps:

[0010] S1. Select input features and output features;

[0011] S2. Collect data on the input features and output features selected in step S1, and then perform data preprocessing. All preprocessed data is randomly divided into a training set, a validation set and a test set;

[0012] S3. Normalize the data in the training set, the validation set and the test set obtained in step S2 to obtain the normalized training set, the normalized validation set and the normalized test set;

[0013] S4. Construct a multi-task deep learning model structure, including expert subnetworks, gated subnetworks, independent subnetworks, self-attention mechanisms and physical constraints;

[0014] S5. Train the multi-task deep learning model structure obtained in step S4 using the normalized training set obtained in step S3 to obtain a trained multi-task deep learning model structure;

[0015] S6. Validate and test the trained multi-task deep learning model structure obtained in step S5 using the validation set and the test set, and then denormalize the output results to obtain the retrieval results of ground surface solar multi-band radiation by fusing physical information and deep learning.

[0016] Further, the input features selected in step S1 include four solar geometric parameters, five solar irradiance-related variables, six weather variables, two atmospheric meteorological variables and two special meteorological parameters.

[0017] The four solar geometric parameters include the solar zenith angle Z, the solar azimuth angle , the atmospheric mass AM and the true solar time AST.

[0018] The five solar irradiance variables include total horizontal irradiance GHI, normal incident irradiance BNI, horizontal diffuse irradiance DHI, extraterrestrial irradiance and surface albedo;

[0019] The six common weather variables include ground temperature, relative humidity, dew point, pressure, wind speed and wind direction;

[0020] The two atmospheric meteorological variables include cloud cover and aerosol optical depth;

[0021] The two special meteorological parameters include the k-index and the Perer parameter;

[0022] The output features include the horizontal total irradiance in the visible light region VIS, the short-wave near-infrared region NIR, the long-wave ultraviolet region UVA, the short-wave ultraviolet region UVB, and the photosynthetically active radiation region PAR.

[0023] Further, the specific implementation method of step S2 includes the following steps:

[0024] S2.1. Collecting input feature data and output feature data, performing time zone alignment, and then performing timestamp alignment;

[0025] S2.2. Quality control is performed on the timestamp-aligned data, irradiance limit test and closure relationship test are performed on the solar irradiance variables, and manual screening is performed on the remaining meteorological elements to remove extreme values and abnormal values;

[0026] The integral values of the five narrow-spectrum irradiance in the full waveband are compared with the wide-band irradiance data to detect abnormal values, and if the integral value of the narrow-band irradiance is less than the wide-band irradiance data, the timestamp data is retained;

[0027] S2.3. All data are randomly divided into a training set, a validation set and a test set according to a ratio of 6:2:2, the training set is used to train model parameters, the validation set is used to debug model parameters to prevent overfitting, and the test set is used to finally evaluate the generalization ability of the model.

[0028] Further, in step S4, the multi-task deep learning model structure sets each expert subnetwork to be shared by all output variables, each output variable is configured with a separate gating subnetwork and an independent subnetwork, the expert subnetwork and the independent subnetwork both adopt a multi-layer perception MLP structure, the gating subnetwork is composed of a multi-layer perception combined with a Softmax function, the multi-layer perception is responsible for extracting high-level features of input data, and then the Softmax function maps it into a probability vector, indicating the participation degree of each expert subnetwork, and finally generates a weight distribution for each expert subnetwork.

[0029] Further, the multi-layer perception in step S4 includes an input layer, a hidden layer and an output layer;

[0030] The hidden layer of the expert sub-network in step S4 adopts 3, and the number of neural nodes of each hidden layer is 144, 96 and 48 respectively; the hidden layer of the gating sub-network and the independent sub-network adopts 2, and the number of neural nodes of each hidden layer is 144 and 96 respectively.

[0031] Further, the basic formula for setting the physical constraint in step S4 is:

[0032] ;

[0033] ;

[0034] The above formula is converted into a specific mathematical form, and the physical constraint relationship is embedded into the multi-task deep learning model structure using the corresponding function;

[0035] Firstly, it is converted into the form of an equality constraint, and is set as The difference between the two sides of the inequality is:

[0036]

[0037]

[0038] Then divide the two sides of the equation by GHI, respectively, to obtain:

[0039]

[0040] The changed physical constraint satisfies the mathematical form that the sum of variables in the 0-1 interval is 1, and then a Softmax function is added before the output layer of the multi-task deep learning model structure to realize the physical constraint formula;

[0041] For , a penalty term is added in the loss function:

[0042]

[0043] Wherein, is a hyperparameter for controlling the penalty strength, represents the mean value of the sample;

[0044] That is, the part exceeding the proportion range is added with a penalty constraint, wherein the part with a comparison value less than 0.45 is squared, and the part with a comparison value greater than 0.55 is squared.

[0045] Further, step S5 sets the loss function formula of model training as follows:

[0046]

[0047]

[0048] wherein, represents the loss function corresponding to the i-th output variable, represents the total loss function, N represents the total number of samples, S represents the total number of output variables, y, , respectively represent the observed value, the estimated value and the observed value mean of the output variable.

[0049] Further, the multi-task deep learning model structure in step S5 is built and trained using the pytorch library; after inputting the data, the forward propagation of the model is carried out, the estimated values of multiple tasks are obtained after the expert subnetwork, the gate subnetwork, the independent subnetwork and the self-attention mechanism and the physical constraint, each estimated value is compared with the true label to calculate the sub-loss, and the gradient of the loss to the model parameters is calculated after the sub-loss is weighted as the total loss, the error is propagated layer by layer backward through the chain rule, the model parameters are updated through the Adam optimizer with a learning rate of 0.0001, and after 10000 times of repeated forward propagation and parameter update iteration process, the final stopping condition is reached, and the training is ended.

[0050] The beneficial effects of the application are as follows:

[0051] The method for fusing physical information and deep learning to invert ground surface solar multi-band radiation provided by the application proposes a modeling method based on a multi-input multi-output neural network, and calculates multiple narrow spectral band irradiance information by integrating easily obtained meteorological parameters, including photosynthetically active radiation region, visible light region, short wave near infrared region, long wave ultraviolet region and short wave ultraviolet region. This method not only can significantly reduce the calculation cost, but also can ensure the physical consistency between different bands, and provides reliable technical support for scientific research and engineering application in related fields.

[0052] The method for fusing physical information and deep learning to invert ground surface solar multi-band radiation provided by the application uses a multi-output deep learning neural network model, which can obtain time and space continuous, relatively accurate high temporal and spatial resolution data, and can model multiple target variables, thereby simplifying the model, and adding physical constraints to guide multiple target variables to conform to certain physical facts, and improve the physical consistency of the model. BRIEF DESCRIPTION OF DRAWINGS

[0053] Figure 1 ​A flow chart of a method for fusing physical information and deep learning to retrieve ground surface solar multi-band radiation according to the present application;

[0054] Figure 2 A model structure diagram according to the present application;

[0055] Figure 3 A closure relationship test result graph of the present application in which the ratio of the calculated GHI to the observed GHI is arranged by month;

[0056] Figure 4 A closure relationship test result graph of the present application in which the ratio of the calculated GHI to the observed GHI is arranged by month;

[0057] Figure 5 A multi-spectral data integral value and wide-band GHI comparison result graph according to the present application. DETAILED DESCRIPTION

[0058] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application is further described in detail below in combination with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application, that is, the described specific embodiments are only a part of the embodiments of the present application, but not all the specific embodiments. The components of the specific embodiments of the present application described and shown in the accompanying drawings can be arranged and designed in various different configurations, and the present application can also have other embodiments.

[0059] Therefore, the detailed description of the specific embodiments of the present application provided in the accompanying drawings below is not intended to limit the scope of the claimed present application, but only represents selected specific embodiments of the present application. Based on the specific embodiments of the present application, all other specific embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0060] In order to further understand the inventive content, characteristics and effects of the present application, the following specific embodiments are exemplified, and the accompanying drawings are used for reference. Figure 1 - the accompanying drawings Figure 5 The detailed description is as follows:

[0061] Example 1:

[0062] A method for fusing physical information and deep learning to retrieve ground surface solar multi-band radiation, comprising the following steps:

[0063] S1. Select input features and output features;

[0064] Further, the input features selected in step S1 include four solar geometry parameters, five solar irradiance related variables, six weather variables, two atmospheric meteorological variables, and two special meteorological parameters;

[0065] The four solar geometry parameters include solar zenith angle Z, solar azimuth angle , atmospheric mass AM, and apparent solar time AST.

[0066] The five solar irradiance variables include global horizontal irradiance GHI, normal incidence irradiance BNI, horizontal diffuse irradiance DHI, extraterrestrial irradiance, and ground albedo.

[0067] The six common weather variables include ground temperature, relative humidity, dew point, pressure, wind speed, and wind direction.

[0068] The two atmospheric meteorological variables include cloud cover and aerosol optical thickness.

[0069] The two special meteorological parameters include k-index and Perze parameter.

[0070] Further, the k-index can more clearly understand the characteristics and variation dynamics of the ground solar irradiance, and the Perze parameter is a key index for describing the sky condition.

[0071] The k-index includes total transmittance , direct transmittance , diffuse transmittance , and diffuse fraction k, which are given by the formulas:

[0072] The Perze parameter includes sky brightness and sky clarity , which are obtained by the formulas and .

[0073] In the above two formulas, represents horizontal extraterrestrial irradiance, represents normal extraterrestrial irradiance, represents relative optical atmospheric mass, which can be calculated using the approximate formula , where γ is the solar elevation angle, and GHI, BNI, and DHI represent total horizontal irradiance, normal direct irradiance, and horizontal diffuse irradiance, respectively.

[0074] ​Further, in order to obtain accurate ground narrow-band data for use as true value labels in deep learning model training, ground measurement data from the benchmark measurement system of the Solar Radiation Research Laboratory under the U.S. National Renewable Energy Laboratory is selected. The site contains a variety of measurement instruments, and the operation and maintenance are in good condition, which can provide the required radiation data in multiple bands. The output features include the total horizontal irradiance in the visible light region VIS, the short-wave near-infrared region NIR, the long-wave ultraviolet region UVA, the short-wave ultraviolet region UVB, and the photosynthetically active radiation region PAR.

[0075] S2. Data collection is performed on the input features and output features selected in step S1, and then data preprocessing is performed. All preprocessed data is randomly divided into a training set, a validation set, and a test set;

[0076] Further, the specific implementation method of step S2 includes the following steps:

[0077] S2.1. Collect input feature data and output feature data, perform time zone alignment, and then perform timestamp alignment;

[0078] Further, since different meteorological data have different time resolutions (ground observation 1 min / satellite inversion 30 min), in order to facilitate subsequent data processing, all variable timestamps need to be aligned: first, ensure that the timestamps of all data are located in the same time zone, and if the time zones are different, the time zones need to be converted, and then the 1 min resolution data of the ground observation is aggregated to align with the satellite inversion data. For example, for the data with a timestamp of 18:30 in satellite inversion, 30 data from 18:16 to 19:15 in ground observation need to be aggregated to match the satellite inversion data.

[0079] S2.2. Quality control is performed on the data after timestamp alignment. Irradiance limit test and closure relationship test are performed on the solar irradiance variable, and manual screening is performed on the remaining meteorological elements to remove extreme values and abnormal values;

[0080] The integral values of the five narrow-band irradiance in the full-band are compared with the wide-band irradiance data to detect abnormal values. If the integral value of the narrow-band irradiance is less than the wide-band irradiance data, the timestamp data is retained;

[0081] Further, the limit test adopts extreme rare test to exclude extreme rare values in each irradiance data. The closure relationship test is based on the closure relationship equation of three solar irradiance components, i.e. to examine the compliance of the data with the closure relationship equation to screen data points that deviate too much from the equation. The specific screening condition is: let the observed irradiance component be , and let the total horizontal irradiance calculated by the closure relationship equation be then when , when ,

[0082] S2.3. All data is randomly divided into training set, validation set and test set according to the ratio of 6:2:2, the training set is used to train the model parameters, the validation set is used to debug the model parameters to prevent overfitting, and the test set is used to finally evaluate the generalization ability of the model.

[0083] S3. The data in the training set, the validation set and the test set obtained in step S2 are normalized to obtain the normalized training set, the normalized validation set and the normalized test set;

[0084] S4. A multi-task deep learning model structure is constructed, including expert subnetworks, gate subnetworks, independent subnetworks, self-attention mechanisms and physical constraints;

[0085] Further, in step S4, the multi-task deep learning model structure sets each expert subnetwork to be shared by all output variables, each output variable is configured with a gate subnetwork and an independent subnetwork, the expert subnetwork and the independent subnetwork both adopt a multi-layer perception (MLP) structure, the gate subnetwork is composed of a multi-layer perception combined with a Softmax function, the multi-layer perception is responsible for extracting high-level features of the input data, and then the Softmax function maps it to a probability vector, which represents the participation degree of each expert subnetwork, and finally generates a weight distribution for each expert subnetwork.

[0086] Further, the Softmax formula can be expressed as: for an input vector z=[ ], the i-th output of the Softmax function is: .

[0087] Further, the number of expert subnetworks is 8.

[0088] Further, in step S4, the multi-layer perception includes an input layer, a hidden layer and an output layer.

[0089] In step S4, the hidden layer of the expert subnetwork adopts 3, and the number of neural nodes of each hidden layer is 144, 96 and 48 respectively; the hidden layer of the gate subnetwork and the independent subnetwork adopts 2, and the number of neural nodes of each hidden layer is 144 and 96 respectively.

[0090] Further, the nonlinear activation function used in the hidden layer of the multi-layer perception is Mish: At the same time, the number of hidden layers and the number of neural nodes of each layer of each subnetwork are optimized as hyperparameters during model training,

[0091] Further, the self-attention mechanism realizes the information interaction and dependency modeling between different tasks, and takes the feature representation of all tasks as a task sequence, and stacks it into a matrix: Then the self-attention mechanism is used to capture the correlation between tasks: So as to adjust the representation of each task, and further improve the multi-task prediction performance, wherein is the representation vector of each task, , and is the trainable parameter matrix, is the scaling factor of the attention mechanism, and represents the dimension of

[0092] Further, the basic formula for setting the physical constraint in step S4 is:

[0093] ;

[0094] ;

[0095] The above formula is converted into a specific mathematical form, and the physical constraint relationship is embedded into the multi-task deep learning model structure using the corresponding function;

[0096] Firstly, it is converted into the form of an equality constraint, and set is the difference on both sides of the inequality, that is, the irradiance of the remaining wave band:

[0097]

[0098]

[0099] Then divide both sides of the equation by GHI to get:

[0100]

[0101] The changed physical constraint satisfies the mathematical form that the sum of variables with multiple values in the interval 0-1 is 1, so a Softmax function is added before the output layer of the multi-task deep learning model structure to realize the physical constraint formula;

[0102] For , a penalty term is added in the loss function:

[0103]

[0104] Wherein, is a hyperparameter for controlling the penalty strength, represents the mean value of the sample;

[0105] That is, to impose a penalty on the part that exceeds the ratio range, where The part with a contrast value less than 0.45 is subjected to square penalty. A square penalty is applied to the part with a contrast value greater than 0.55.

[0106] S5. Using the normalized training set obtained in step S3 to train the multi-task deep learning model structure obtained in step S4, a trained multi-task deep learning model structure is obtained;

[0107] Furthermore, step S5 sets the loss function formula for model training as follows:

[0108] ,

[0109]

[0110] in, represents the loss function corresponding to the i-th output variable, Represents the total loss function, N represents the total number of samples, S represents the total number of output variables, y, 、 Represent the observed value, estimated value and mean of the observed value of the output variable respectively. In order to balance the training effect of the neural network on different tasks, a balance coefficient is set for the loss function of each task. Each sub-loss function is first multiplied by its own balance coefficient as a weight, and then added to form the total loss function. The balance coefficient is selected as the mean of the corresponding output variable.

[0111] Furthermore, in step S5, the multi-task deep learning model structure is built and trained using the pytorch library; after the data is input, the model is forward propagated, and after passing through the expert sub-network, gated sub-network, independent sub-network, self-attention mechanism, and physical constraints, estimated values ​​of multiple tasks are obtained. Each estimated value is compared with the true label to calculate the sub-loss, and the sub-loss is weighted as the total loss to calculate the gradient of the loss with respect to the model parameters. The error is propagated backward layer by layer through the chain rule, and the model parameters are updated through the Adam optimizer with a learning rate of 0.0001. After 10,000 iterations of forward propagation and parameter update, the stopping condition is finally reached and the training ends.

[0112] S6. Use the validation and test sets to verify and test the trained multi-task deep learning model structure obtained in step S5. After denormalizing the output results, obtain the surface solar multi-band radiation inversion results that integrate physical information and deep learning.

[0113] Compared with ground observation methods, the proposed method can provide spatially continuous multi-spectral irradiance information, breaking the dependence on expensive ground-based spectral measurement equipment and achieving efficient acquisition of high-quality multi-spectral data. Compared with traditional semi-empirical models, the proposed method has significant advantages in terms of accuracy and spatial and temporal resolution. Compared with single-task machine learning methods, the proposed method significantly reduces model complexity and saves computing resources while effectively improving the physical consistency and reliability of the prediction results by introducing physical constraints. In addition, the proposed method simultaneously outputs five narrow-band radiation information, which has a significant comprehensive advantage compared with previous methods that only model a single waveband.

[0114] Figure 3 and Figure 4 In FIG. 6, the gray data points are the data that pass the screening, and the red data points are the data that need to be screened out. Figure 5 In FIG. 7, the black data points are the data that pass the screening, and the red and green data points are the data that need to be screened out. Table 1 is an error analysis of the estimated values of each target variable generated by the proposed method on the test set and the true value labels in the test set. The error indicators are the normalized root mean square error (nRMSE), the mean absolute error (nMAE), the mean bias (nMBE), and the determination coefficient (R2). As can be seen from Table 1, the results are good.

[0115] Table 1

[0116]

[0117] It should be noted that the relational terms, such as first and second, and the like, are used solely to distinguish one from another entity or action, without necessarily requiring or implying any actual relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without more limitations, an element preceded by "comprises... " does not exclude the existence of additional identical elements in the process, method, article, or apparatus that includes the element.

[0118] ​Although the present application has been described with reference to the specific embodiments thereof, it should be understood by those skilled in the art that various changes can be made and equivalents can be substituted for elements thereof without departing from the scope of the present application. In particular, various features and aspects of the present application can be used individually or in any combination depending on the specific application and implementation. Therefore, it is expressly intended that the specific embodiments of the present application both as set forth and including any equivalents thereof should not limit the present application or scope of the claims herein, but rather the overall scope of pertaining solely to the methods and the articles of manufacture specifically recited in the following claims.

Claims

1. A method for inverting multi-band solar radiation on the surface by integrating physical information and deep learning, characterized in that: The steps include: S1. Select input and output features; S2. Collect data on the input and output features selected in step S1, then perform data preprocessing, and randomly divide all preprocessed data into training, validation, and test sets; S3. The data in the training set, validation set, and test set obtained in step S2 are normalized to obtain normalized training set, validation set, and test set; S4. Build a multi-task deep learning model architecture, including an expert sub-network, a gating sub-network, an independent sub-network, a self-attention mechanism, and physical constraints. S5. Using the normalized training set obtained in step S3 to train the multi-task deep learning model structure obtained in step S4, a trained multi-task deep learning model structure is obtained; S6. Use the validation and test sets to verify and test the trained multi-task deep learning model structure obtained in step S5. After denormalizing the output results, obtain the surface solar multi-band radiation inversion results that integrate physical information and deep learning.

2. The method for inverting surface solar multi-band radiation by integrating physical information and deep learning according to claim 1, characterized in that: The input features selected in step S1 include four solar geometry parameters, five solar irradiance-related variables, six weather variables, two atmospheric meteorological variables, and two special meteorological parameters; The four solar geometric parameters include the solar zenith angle Z, the solar azimuth angle , atmospheric mass AM and true solar time AST; The five solar irradiance variables include total horizontal irradiance GHI, normal direct irradiance BNI, horizontal diffuse irradiance DHI, extraterrestrial irradiance and surface albedo; Six common weather variables include surface temperature, relative humidity, dew point, pressure, wind speed, and wind direction; Two atmospheric meteorological variables include cloud cover and aerosol optical depth; Two special meteorological parameters include the k index and the Perer parameter; The output characteristics include the total horizontal irradiance of the visible light region VIS, the shortwave near infrared region NIR, the longwave ultraviolet region UVA, the shortwave ultraviolet region UVB, and the photosynthetically active radiation region PAR.

3. The method for inverting surface solar multi-band radiation by integrating physical information and deep learning according to claim 1 or 2, characterized in that: The specific implementation method of step S2 includes the following steps: S2.

1. Collect input and output feature data, align their time zones, and then align their timestamps. S2.

2. Perform quality control on the timestamp-aligned data, perform irradiance limit and closure relationship tests on the solar irradiance variable, and manually screen the remaining meteorological elements to remove extreme values ​​and outliers. Compare the integral values ​​of the five narrow spectrum irradiances over the entire band with the wide band irradiance data to detect abnormal values. If the integral value of the narrow band irradiance is smaller than that of the wide band irradiance data, retain the timestamp data. S2.

3. All data are randomly divided into training set, validation set and test set in the ratio of 6:2:

2. The training set is used to train the model parameters, the validation set is used to debug the model parameters to prevent overfitting, and the test set is used to finally evaluate the generalization ability of the model.

4. The method for inverting surface solar multi-band radiation by integrating physical information and deep learning according to claim 3, characterized in that: In step S4, the multi-task deep learning model structure is set up so that each expert sub-network is shared by all output variables. Each output variable is separately configured with a gating sub-network and an independent sub-network. Both the expert sub-network and the independent sub-network adopt the multi-layer perceptron MLP structure. The gating sub-network is composed of a multi-layer perceptron combined with a Softmax function. The multi-layer perceptron is responsible for extracting high-level features of the input data, and then the Softmax function maps it into a probability vector, which represents the degree of participation of each expert sub-network, and finally generates a weight distribution for each expert sub-network.

5. The method for inverting surface solar multi-band radiation by integrating physical information and deep learning according to claim 3, characterized in that: In step S4, the multilayer perceptron includes an input layer, a hidden layer, and an output layer; In step S4, the expert sub-network has three hidden layers, and the number of neural nodes in each hidden layer is 144, 96, and 48 respectively; the gated sub-network and the independent sub-network have two hidden layers, and the number of neural nodes in each hidden layer is 144 and 96 respectively.

6. The method for inverting surface solar multi-band radiation by integrating physical information and deep learning according to claim 5, characterized in that: The basic formula for setting the physical constraints in step S4 is: ； Convert the above formula into a specific mathematical form and use the corresponding function to embed the physical constraint relationship into the multi-task deep learning model structure; First, transform it into the form of equality constraints, set The difference between the two sides of the inequality sign: Then divide both sides of the equation by GHI to get: The changed physical constraint satisfies the mathematical form that the sum of multiple variables with values ​​in the range of 0-1 is 1. A Softmax function is added before the output layer of the multi-task deep learning model structure to implement the physical constraint formula; for , by adding a penalty term to the loss function : in, is a hyperparameter used to control the penalty intensity, represents the mean of the sample; That is, to impose a penalty on the part that exceeds the ratio range, where The part with a contrast value less than 0.45 is subjected to square penalty. A square penalty is applied to the part with a contrast value greater than 0.

55.

7. The method for inverting multi-band solar radiation on the surface by integrating physical information and deep learning according to claim 6, characterized in that: Step S5 sets the loss function formula for model training as follows: , in, represents the loss function corresponding to the i-th output variable, Represents the total loss function, N represents the total number of samples, S represents the total number of output variables, y, 、 represent the observed value, estimated value and mean of the observed value of the output variable respectively.

8. The method for inverting multi-band solar radiation on the surface by integrating physical information and deep learning according to claim 7, characterized in that: In step S5, the multi-task deep learning model structure is built and trained using the pytorch library; after the data is input, the model is forward propagated, and after passing through the expert sub-network, gated sub-network, independent sub-network, self-attention mechanism, and physical constraints, estimated values ​​of multiple tasks are obtained. Each estimated value is compared with the true label to calculate the sub-loss, and the sub-loss is weighted as the total loss to calculate the gradient of the loss with respect to the model parameters. The error is propagated backward layer by layer through the chain rule, and the model parameters are updated using the Adam optimizer with a learning rate of 0.0001. After 10,000 iterations of forward propagation and parameter update, the stopping condition is finally reached and the training ends.

Images