A regional horizontal irradiance inversion method and system based on satellite data and deep learning

Abstract

The application discloses a kind of based on satellite data and deep learning's regional horizontal irradiance inversion method and system, belong to solar meteorology technical field, method includes: obtaining FY4A satellite data;GHI data of measuring station is obtained and GHI missing value is deleted, obtain GHI pretreatment data;Satellite data and GHI pretreatment data are corresponding, eliminate missing data in satellite data;According to the geographical time information where each station is located, the distance between the earth and the sun, the solar zenith angle and the solar elevation angle of each station at a certain time are calculated;Based on the distance between the earth and the sun and the solar zenith angle, extraterrestrial solar radiation is calculated, and based on extraterrestrial solar radiation, quality control is carried out on GHI pretreatment data, to obtain processed GHI training set;According to the processed GHI training set, the initial TabNet model is improved and trained to obtain an irradiance inversion model;The reprocessed satellite data feature set is input into the irradiance inversion model to obtain a high-precision inversion map.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of solar meteorology, and particularly relates to a regional horizontal irradiance inversion method and system based on satellite data and deep learning. BACKGROUND

[0002] Global horizontal irradiance (GHI) is the most important parameter in solar resource assessment. At present, the inversion of GHI is mainly completed by two methods, namely physical method and statistical method (including machine learning). In the physical method, the theory of radiation transmission is quite mature, but it has obvious limitations, because the physical inversion algorithm is essentially a local fitting function, and the atmospheric parameters required to drive the radiation transmission model are different for different regions, which means that even if the physical method is very mature, it is facing a bottleneck. Under the requirement of the solar resource map for higher precision and higher accuracy, compared with the physical method, the statistical inversion method (machine learning) is a very competitive alternative method. Machine learning can make full use of a large amount of high-quality observation data, and it can fit any complex nonlinear function to ensure its applicability after training. Although training such a model requires big data and computers, once trained, the network can quickly produce retrieval results. SUMMARY

[0003] The application aims to solve the problems of the prior art, and provides the following scheme:

[0004] A regional horizontal irradiance inversion method based on satellite data and deep learning, comprising the following steps:

[0005] Obtain FY4A full-disk L1-level 4KM resolution satellite data; obtain GHI data of a radiation measurement station and delete GHI missing values to obtain GHI pretreatment data;

[0006] Correspond the satellite data and the GHI pretreatment data, and eliminate missing data in the satellite data;

[0007] According to geographical time information of each radiation measurement station, calculate the distance between the earth and the sun, the solar zenith angle and the solar elevation angle of each station at a certain time;

[0008] Calculate extraterrestrial solar radiation based on the distance between the earth and the sun and the solar zenith angle, and perform quality control on the GHI pretreatment data based on the extraterrestrial solar radiation to obtain a processed GHI training set;

[0009] The initial TabNet model is improved according to the non-0 characteristics of the data of the processed GHI training set, the improved model is trained based on the distance between the earth and the sun, the solar zenith angle, the solar elevation angle and the processed GHI training set, and a radiation inversion model is obtained.

[0010] The reprocessed satellite data feature set is input into the radiation inversion model, and a high-precision inversion map is obtained.

[0011] Preferably, the method for eliminating missing data in the satellite data comprises:

[0012] Eliminate data in which all channels in the satellite data are 0.

[0013] Eliminate data in which any channel in the satellite data is missing.

[0014] Preferably, the method for obtaining the processed GHI training set comprises:

[0015] The extraterrestrial solar radiation is calculated based on the distance between the earth and the sun and the solar zenith angle:

[0016]

[0017] Wherein, R0 represents the extraterrestrial solar radiation, S0 represents the solar constant, r0 represents the average distance between the earth and the sun, r represents the distance between the earth and the sun on a certain day, represents the solar zenith angle.

[0018] The expected minimum value of GHI is set, the data in the preprocessed data which is greater than the extraterrestrial solar radiation and less than the expected minimum value of GHI is eliminated, and the processed GHI training set is obtained.

[0019] Preferably, the method for improving the initial TabNet model comprises: adding a ReLU activation function after the last fully connected layer in the initial TabNet model according to the non-0 characteristics of the GHI data.

[0020] The application also provides an area horizontal radiation inversion system based on satellite data and deep learning, which applies the inversion method of any one of the above, comprising: a data acquisition module, a satellite data processing module, a geographic data calculation module, a GHI data processing module, a model training module and an inversion module.

[0021] The data acquisition module is used for acquiring FY4A full-disk L1-level 4KM resolution satellite data, acquiring GHI data of a radiation measurement station and deleting GHI missing values, and obtaining GHI preprocessed data.

[0022] The satellite data processing module is configured to correspond the satellite data with the GHI pre-processing data, and eliminate missing data in the satellite data.

[0023] The geographic data calculation module is configured to calculate the distance between the Earth and the Sun, the solar zenith angle and the solar elevation angle of each station at a certain time according to geographic time information of each radiation measurement station.

[0024] The GHI data processing module is configured to calculate extraterrestrial solar radiation based on the distance between the Earth and the Sun and the solar zenith angle, and perform quality control on the GHI pre-processing data based on the extraterrestrial solar radiation to obtain a processed GHI training set.

[0025] The model training module is configured to improve an initial TabNet model according to the characteristic that data of the processed GHI training set is not 0, and train the improved model based on the distance between the Earth and the Sun, the solar zenith angle, the solar elevation angle and the processed GHI training set to obtain an irradiance inversion model.

[0026] The inversion module is configured to input the reprocessed satellite data feature set into the irradiance inversion model to obtain a high-precision inversion map.

[0027] Preferably, in the satellite data processing module, the process of eliminating missing data in the satellite data includes:

[0028] Eliminate data in which all channels of the satellite data are 0.

[0029] Eliminate data in which any channel of the satellite data is missing.

[0030] Preferably, the working process of the GHI data processing module includes:

[0031] Calculate extraterrestrial solar radiation based on the distance between the Earth and the Sun and the solar zenith angle:

[0032]

[0033] wherein R0 represents extraterrestrial solar radiation, S0 represents a solar constant, r0 represents an average distance between the Earth and the Sun, r represents the distance between the Earth and the Sun on a certain day, represents the solar zenith angle.

[0034] Set a minimum expected GHI value, eliminate data in the pre-processing data that is greater than the extraterrestrial solar radiation and less than the minimum expected GHI value to obtain the processed GHI training set.

[0035] Preferably, in the model training module, the process of improving the initial TabNet model comprises: adding a ReLU activation function after the last fully connected layer in the initial TabNet model according to the characteristics of non-0 GHI data.

[0036] Compared with the prior art, the present application has the following beneficial effects:

[0037] The present application optimizes the inversion reliability compared with the ordinary TabNet model, so that the inverted GHI does not appear abnormal negative value. The FY4A satellite data is used to realize the full coverage of the region, and the "boundary problem" of the previous GHI inversion is solved. At the same time, compared with other deep learning models, the improved TabNet model has interpretability, can output the global importance and local importance of features, and improves the inversion effect of GHI. The evaluation results of the GHI inversion method of the present application are compared with the random forest model, which improves the accuracy to a certain extent and achieves good results. BRIEF DESCRIPTION OF DRAWINGS

[0038] In order to more clearly illustrate the technical solutions of the present application, the following briefly introduces the drawings needed in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0039] Figure 1 The method flowchart of the embodiment of the present application is shown in the figure;

[0040] Figure 2 The figure shows that the 101 sites in the country are divided into training set and test set in the embodiment of the present application;

[0041] Figure 3 The scatter plot of the actual value and the inversion value of each site on the test set in the embodiment of the present application is shown in the figure, wherein (a) is the scatter plot of the actual value and the inversion value of the improved TabNet model, and (b) is the scatter plot of the actual value and the inversion value of the random forest model;

[0042] Figure 4 The spatial distribution diagram of the root mean square error RMSE, the mean bias error MBE and the correlation coefficient R of each site on the test set in the embodiment of the present application is shown in the figure, wherein (a) is the root mean square error RMSE of each site on the test set, (b) is the mean bias error MBE of each site on the test set, and (c) is the correlation coefficient R of each site on the test set;

[0043] Figure 5The inversion effect pictures of the embodiment of the present application at GHI from 00:00 to 11:00 on July 1, 2018 in world coordinated time (UTC) are shown in the following table, wherein a is the inversion effect picture at 00:00, b is the inversion effect picture at 01:00, c is the inversion effect picture at 02:00, d is the inversion effect picture at 03:00, e is the inversion effect picture at 04:00, f is the inversion effect picture at 05:00, g is the inversion effect picture at 06:00, h is the inversion effect picture at 07:00, i is the inversion effect picture at 08:00, j is the inversion effect picture at 09:00, k is the inversion effect picture at 10:00, and l is the inversion effect picture at 11:00. DETAILED DESCRIPTION

[0044] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the present application.

[0045] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0046] Embodiment one

[0047] In this embodiment, as shown in the following table, a regional horizontal irradiance inversion method based on satellite data and deep learning includes the following steps: Figure 1

[0048] S1. Obtain FY4A full-disk L1-level 4KM resolution satellite data; obtain GHI data of radiation measurement stations and delete GHI missing values to obtain GHI pretreatment data.

[0049] In this embodiment, FY4A full-disk L1-level 4KM resolution data from March 12, 2018 to December 13, 2018 is downloaded, including 14 spectral bands. The time period of the downloaded data is the whole time of each day, i.e. the whole time data from 00:00 to 23:00, and there are 24 satellite data in each day. The hourly GHI ground measurement data provided by 101 radiation measurement stations of the China Meteorological Administration in 2018 is pretreated, and the satellite data and the GHI data at each time are corresponded. The corresponding rule is to average the satellite data within 0.04° of the station latitude and longitude at the same time, so as to represent the satellite data of 14 spectral bands at this time.

[0050] S2. Correspond the satellite data and the GHI pretreatment data, and delete the missing data in the satellite data.

[0051] ​Methods for removing missing data from satellite data include: removing data where all channels are 0; and removing data where any channel in the satellite data is missing.

[0052] S3. Based on the geographic time information (including latitude and longitude, topography and time) of each radiation measurement station, calculate the Earth-Sun distance, solar zenith angle and solar altitude angle of each station at a certain moment, and use them as input features for the model.

[0053] S4. Calculate extraterrestrial solar radiation based on Earth-Sun distance and solar zenith angle, and perform quality control on preprocessed data based on extraterrestrial solar radiation to obtain the processed GHI training set.

[0054] The methods for obtaining the processed GHI training set include: calculating extraterrestrial solar radiation based on the Earth-Sun distance and the solar zenith angle.

[0055]

[0056] Where R0 represents extraterrestrial solar radiation, S0 represents the solar constant, r0 represents the average Earth-Sun distance, and r represents the Earth-Sun distance on a given day. This represents the solar zenith angle. A minimum expected GHI value is set, and data exceeding the extraterrestrial solar radiation but below the minimum expected GHI value are removed from the preprocessed data to obtain the processed GHI data. In this embodiment, the minimum expected GHI value is set to 10 W / m². 2 .

[0057] S5. Based on the non-zero characteristics of the processed GHI training set, the initial TabNet model is improved. The improved model is trained based on the Earth-Sun distance, solar zenith angle, solar altitude angle and processed GHI data to obtain the irradiance inversion model.

[0058] One method to improve the initial TabNet model is to add a ReLU activation function after the last fully connected layer in the initial TabNet model, taking advantage of the non-zero nature of GHI data.

[0059] The training method included dividing 101 stations nationwide into a training set and a test set at an 8:2 ratio, with 81 stations in the training set and 20 stations in the test set. Twenty features—including Earth-Sun distance, solar zenith angle, solar altitude angle, station latitude and longitude, topography, and 14 FY4A spectral channels—were used as input to the improved TabNet model, and preprocessed GHI data was used as output to train the improved model. Specifically, the improved TabNet model was trained using the Adam optimizer with a learning rate of 0.01, and its hyperparameters were set to the width n of the decision prediction layer. aThe attention embedding width n of each mask is 16 d The number n of decision steps in the structure is 16 steps The remaining settings are all default settings of the model, and the early stopping round is 20. After the training is completed, the model is compared with the random forest model to evaluate the model, and the statistical quantities used are the root mean square error RMSE, the correlation coefficient R, and the mean bias error MBE. Among them, RMSE and MBE are defined as follows:

[0060]

[0061] Where y i represents the measured GHI value, represents the predicted GHI value, cov represents the covariance, represents the standard deviation of the measured GHI value, represents the standard deviation of the predicted GHI value.

[0062] S6. Input the reprocessed satellite data feature set into the irradiance inversion model to obtain a high-precision inversion map.

[0063] In this embodiment, all 101 stations are input into the evaluated model. Specifically, an input feature data set is constructed, the data in each channel of the processed satellite data is flattened into one dimension, and the corresponding latitude, longitude and altitude are matched, the distance from the earth to the sun, the solar zenith angle and the solar azimuth angle are calculated, and a 4KM resolution input feature set is obtained. The input feature set is input into the irradiance inversion model to obtain the inverted GHI and visualize it, thereby obtaining a high-precision 4KM resolution inversion map of GHI.

[0064] Embodiment Two

[0065] In this embodiment, a regional horizontal irradiance inversion system based on satellite data and deep learning includes a data acquisition module, a satellite data processing module, a geographic data calculation module, a GHI data processing module, a model training module and an inversion module.

[0066] The data acquisition module is used to acquire FY4A full-disk L1-level 4KM resolution satellite data, and to acquire GHI data of radiation measurement stations and delete GHI missing values to obtain GHI preprocessed data.

[0067] The satellite data processing module is used to correspond the satellite data with the GHI preprocessed data, and to eliminate missing data in the satellite data. In the satellite data processing module, the process of eliminating missing data in the satellite data includes: eliminating data in which all channels of the satellite data are 0; and eliminating data in which any channel of the satellite data is missing.

[0068] The geographic data calculation module calculates the distance between the Earth and the Sun, the solar zenith angle and the solar elevation angle of each station at a certain time according to the geographic time information of each radiation measurement station.

[0069] The working process of the GHI data processing module includes: calculating the extraterrestrial solar radiation based on the distance between the Earth and the Sun and the solar zenith angle:

[0070]

[0071] wherein R0 represents the extraterrestrial solar radiation, S0 represents the solar constant, r0 represents the average distance between the Earth and the Sun, r represents the distance between the Earth and the Sun on a certain day, the solar zenith angle is represented by θ0; the expected minimum value of GHI is set, and the data greater than the extraterrestrial solar radiation and less than the expected minimum value of GHI in the preprocessed data are removed to obtain the processed GHI training set.

[0072] The GHI data processing module calculates the extraterrestrial solar radiation based on the distance between the Earth and the Sun and the solar zenith angle, and performs quality control on the GHI preprocessed data based on the extraterrestrial solar radiation to obtain the processed GHI training set.

[0073] The model training module improves the initial TabNet model according to the non-0 characteristics of the data of the processed GHI training set, trains the improved model based on the distance between the Earth and the Sun, the solar zenith angle, the solar elevation angle and the processed GHI training set, and obtains the irradiance inversion model. In the model training module, the process of improving the initial TabNet model includes: adding a ReLU activation function after the last fully connected layer in the initial TabNet model according to the non-0 characteristics of the GHI data.

[0074] The inversion module is used for inputting the reprocessed satellite data feature set into the irradiance inversion model to obtain a high-precision inversion map.

[0075] Embodiment three

[0076] In this embodiment, the feasibility and accuracy of the GHI inversion in the present application are verified by an actual case.

[0077] The GHI inversion method in the present application can invert all time instances with data information, and is all based on domestic data and completed on a domestic platform. Here, the results of one division are taken as an example, and the high-precision inversion effect of GHI4KM from 00:00 to 11:00 on July 1, 2018 (UTC) is displayed to illustrate the inversion effect of the present method on GHI.

[0078] The steps of obtaining the GHI distribution of July 1, 2018 using the GHI inversion method in the present application are as follows:

[0079] 1、asFigure 2 As shown in FIG. 6, 101 stations nationwide are divided into a training set and a test set in a ratio of 8:2, with a total of 81 stations in the training set and 20 stations in the test set, so as to test the spatial generalization ability of the improved TabNet model for GHI inversion (all GHI data of the above stations have been made into corresponding input features through S1 to S4 in the specific implementation steps).

[0080] 2. The training set and the test set are put into the model according to the hyperparameter settings mentioned in the specific implementation step S5, and verification is performed.

[0081] Figure 3 The scatter plot of the GHI values inverted by the method and the observed GHI values is given in FIG. 7, wherein, Figure 3 In FIG. 7(a), the scatter plot of the actual values and the inverted values of the improved TabNet model is given, Figure 3 In FIG. 7(b), the scatter plot of the actual values and the inverted values of the random forest model is given, the solid line is the standard line of 100% inversion, and the dotted line is the one-dimensional linear regression of the actual values and the inverted values. The one-dimensional linear regression function established by the predicted values and the observed values shows that the improved TabNet model can better invert the GHI values and has little difference with the actual values of the observed values. The random forest model and the improved TabNet model have similar inversion trends for the hourly GHI, but the inversion effect of the random forest model is obviously not as good as that of the improved TabNet model.

[0082] The score comparison of the improved TabNet model and the random forest model is given in Table 1. The improved TabNet model is better than the random forest model in terms of RMSE and R, but the MBE is larger than that of the random forest model. This may be caused by the mutual offset of the overestimated and underestimated hourly GHI values of the random forest model. In combination with the results of Table 1 and Figure 3 It can be concluded that the improved TabNet model has better effect on the inversion of the hourly GHI than the random forest model.

[0083] Table 1

[0084]

[0085] Figure 4 The spatial distribution of the root mean square error RMSE, the correlation coefficient R and the mean bias error MBE of each station on the test set is given in FIG. 8. Figure 4 In FIG. 8(a), it is shown that in the East China, Central China and North China regions with more training stations, the RMSE is smaller, while in the West China and Northeast China regions, the RMSE is larger. In addition, the accuracy of 70% of the test stations is better, and the RMSE is below 86 W / m2. Figure 4The middle (b) shows that the model tends to overestimate GHI at the test stations in North China, East China and Yunnan Province. In Qinghai, northern Xinjiang and southern Hainan Island, the model tends to underestimate GHI. In addition, the MBE of 50% of the stations is between -7 W / m2 and 9 W / m2, which means that the model has good accuracy. Figure 4 The middle (c) shows that the correlation coefficient R of all stations is above 0.92, which proves that the model can learn the time variation of GHI well. In summary, the model also has good spatial generalization ability in space. The model not only learns the correct time variation of GHI, but also has good accuracy on unknown data.

[0086] 3. The high-precision inversion of GHI4KM from 00:00 to 11:00 (UTC) on July 1, 2018 is obtained according to the method mentioned in embodiment one, and is shown in the middle. Figure 5 The middle.

[0087] As can be seen from Figure 5 In the morning at 8 o'clock, the east of China is brighter and the west of China is darker. With the passage of time, the value of GHI reaches the maximum at 12 o'clock and remains bright until 2 o'clock in the afternoon. After 2 o'clock in the afternoon, the value of GHI gradually decreases and all become dark at 7 o'clock in the evening, which means that the improved TabNet model can learn the time variation of GHI through the solar elevation angle and other characteristic quantities with time information. In addition, the improved TabNet model can also learn the high-precision cloud, water vapor and aerosol features from FY4A satellite data, so as to give different regional GHI values and have high precision, so that the power department can schedule photovoltaic resources according to the distribution of regional GHI, and contribute to the energy transformation of China.

[0088] The above-described embodiments are only descriptions of the preferred modes of the present application and do not limit the scope of the present application. Without departing from the design spirit of the present application, various modifications and improvements to the technical solutions of the present application made by those skilled in the art shall fall within the protection scope determined by the claims of the present application.

Claims

1. A method for retrieving regional horizontal irradiance based on satellite data and deep learning, characterized in that, The method comprises the following steps: Satellite data of FY4A full disc L1 level 4KM resolution is acquired; GHI data of a radiation measurement station is acquired, and missing values of GHI are deleted to obtain preprocessed GHI data; The satellite data is matched with the preprocessed GHI data, and missing data in the satellite data is deleted; According to geographical time information of each radiation measurement station, the distance between the earth and the sun, the solar zenith angle and the solar elevation angle of each station at a certain time are calculated; Extraterrestrial solar radiation is calculated based on the distance between the earth and the sun and the solar zenith angle, and the preprocessed GHI data is subjected to quality control based on the extraterrestrial solar radiation to obtain a processed GHI training set; According to the characteristic that the data of the processed GHI training set is not 0, an initial TabNet model is improved, and the improved model is trained based on the distance between the earth and the sun, the solar zenith angle, the solar elevation angle and the processed GHI training set to obtain an irradiance inversion model; The feature set of the reprocessed satellite data is input into the irradiance inversion model to obtain a high-precision inversion map; The method for improving the initial TabNet model comprises: adding a ReLU activation function after the last fully connected layer in the initial TabNet model according to the characteristic that the GHI data is not 0; The training method includes: dividing 101 stations nationwide into a training set and a test set at an 8:2 ratio, with 81 stations in the training set and 20 stations in the test set; using 20 features—including Earth-Sun distance, solar zenith angle, solar altitude angle, station latitude and longitude, topography, and 14 FY4A spectral channels—as input to the improved TabNet model, and preprocessed GHI data as output to train the improved model; specifically, the improved TabNet model is trained using the Adam optimizer with a learning rate of 0.01, and the hyperparameters are set to the width n of the decision prediction layer. a The attention embedding width for each mask is 16, and the width of each mask is n. d The number of decision steps n in the structure is 16. steps The initial value was set to 3, with all other settings remaining at the model's defaults. The number of early stopping rounds was 20. After training, the model was compared with a random forest model to evaluate it. The statistical measures used were the root mean square error (RMSE), correlation coefficient (R), and mean bias error (MBE). RMSE and MBE were defined as follows: ， ， ， wherein, represents the measured GHI value, represents the predicted GHI value, cov represents the covariance, represents the standard deviation of the measured GHI value, represents the standard deviation of the predicted GHI value.

2. The method according to claim 1, wherein, The method for deleting missing data in the satellite data comprises: Data in which all channels in the satellite data are 0 is deleted; Data in which any channel in the satellite data has missing data is deleted.

3. The method of claim 1, wherein the method is based on satellite data and deep learning. The method for obtaining the processed GHI training set comprises: Extraterrestrial solar radiation is calculated based on the distance between the earth and the sun and the solar zenith angle: ， where, R 0 represents the extraterrestrial solar radiation, S 0 represents the solar constant, r 0 represents the mean distance Earth-Sun, r represents the Earth-Sun distance for a certain day, φ represents the solar zenith angle; The minimum expected value of GHI is set, data in the preprocessed data that is greater than the extraterrestrial solar radiation and less than the minimum expected value of GHI is deleted, and the processed GHI training set is obtained.

4. A regional level irradiance retrieval system based on satellite data and deep learning, the retrieval system applying the retrieval method of any one of claims 1-3, characterized in that, It comprises: a data acquisition module, a satellite data processing module, a geographical data calculation module, a GHI data processing module, a model training module and an inversion module; The data acquisition module is used to acquire satellite data of FY4A full disc L1 level 4KM resolution, and acquire GHI data of a radiation measurement station and delete missing values of GHI to obtain preprocessed GHI data; The satellite data processing module is used to match the satellite data with the preprocessed GHI data, and delete missing data in the satellite data; The geographical data calculation module calculates the distance between the earth and the sun, the solar zenith angle and the solar elevation angle of each station at a certain time according to geographical time information of each radiation measurement station; The GHI data processing module calculates extraterrestrial solar radiation based on the distance between the earth and the sun and the solar zenith angle, and performs quality control on the preprocessed GHI data based on the extraterrestrial solar radiation to obtain a processed GHI training set; The model training module improves an initial TabNet model according to the characteristic that the data of the processed GHI training set is not 0, and trains the improved model based on the distance between the earth and the sun, the solar zenith angle, the solar elevation angle and the processed GHI training set to obtain an irradiance inversion model; The inversion module is configured to input the reprocessed satellite data feature set into the irradiance inversion model to obtain a high-precision inversion map; In the model training module, the process of improving the initial TabNet model includes: adding a ReLU activation function after the last fully connected layer in the initial TabNet model according to the non-0 characteristics of GHI data; The training method comprises the following steps: 101 stations nationwide are divided into a training set and a test set according to a ratio of 8:2, the training set comprises 81 stations, and the test set comprises 20 stations; 20 features including a distance between the earth and the sun, a solar zenith angle, a solar elevation angle, latitude and longitude of the station, terrain and 14 spectral channels of the FY4A are taken as inputs of an improved TabNet model, preprocessed GHI data is taken as output, and the improved model is trained; in the process of training the improved TabNet model, a specific training method of the improved TabNet model is to use an Adam optimizer to train, a learning rate is 0.01, and a width n of a decision prediction layer of the model is set to 16 a The attention embedding width n of each mask is 16 d The number n of decision steps in the structure is 16 steps The number n of decision steps in the structure is 16 The remaining settings are all default settings of the model, the early stopping round is 20, after the training is completed, a random forest model is compared, so that the model is evaluated, statistical quantities used are a root mean square error RMSE and a correlation coefficient R, and a mean bias error MBE; wherein the RMSE and the MBE are respectively defined as: ， ， ， wherein, represents the measured GHI value, represents the predicted GHI value, cov represents the covariance, represents the standard deviation of the measured GHI value, represents the standard deviation of the predicted GHI value.

5. The regional level irradiance retrieval system based on satellite data and deep learning according to claim 4, characterized in that, In the satellite data processing module, the process of eliminating missing data in the satellite data includes: Eliminate data in which all channels in the satellite data are 0; Eliminate data in which any channel in the satellite data is missing.

6. The regional level irradiance retrieval system based on satellite data and deep learning according to claim 4, characterized in that, The working process of the GHI data processing module includes: Calculate extraterrestrial solar radiation based on the distance between the earth and the sun and the solar zenith angle: ， where, R 0 represents the extraterrestrial solar radiation, S 0 represents the solar constant, r 0 represents the mean distance Earth-Sun, r represents the Earth-Sun distance for a certain day, Set the expected minimum value of GHI, eliminate data in the preprocessed data that is greater than the extraterrestrial solar radiation and less than the expected minimum value of GHI, and obtain the processed GHI training set. represents the solar zenith angle; ​

Images