Regional high-resolution horizontal irradiance forecasting method based on deep learning model chain
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- LONGNAN METEOROLOGICAL BUREAU
- Filing Date
- 2026-04-02
- Publication Date
- 2026-07-21
AI Technical Summary
Existing technologies struggle to achieve high-resolution, detailed solar irradiance forecasts, especially accurate predictions over small areas, and the spatial resolution of GHI data is insufficient to support the detailed requirements of solar power plants.
By combining FY4B satellite remote sensing data with a deep learning model chain, and by improving the TabNet and Earthformer models, a high-resolution horizontal irradiance forecasting method is constructed. This method includes data preprocessing, feature extraction, model training, and forecasting steps, ensuring that the output is a non-negative GHI inversion value, and using an improved ReLU activation function to improve forecast accuracy.
It has achieved a 4KM spatial resolution GHI resource distribution map, supports refined forecasting of solar photovoltaic power plants in any region, and achieves a high daily accuracy and excellent pass rate, meeting the industry standards in the field of energy meteorology.
Smart Images
Figure CN122432690A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of solar meteorology technology, and in particular to a method for forecasting regional high-resolution horizontal irradiance based on a deep learning model chain. Background Technology
[0002] Solar energy, as an important renewable energy source, has enormous potential to meet my country's energy needs and promote sustainable development. Photovoltaic systems can achieve zero-carbon electricity production, effectively alleviating energy shortages. Therefore, as the world's largest carbon emitter, my country urgently needs to expand the application of solar power generation to reduce carbon emissions and address global warming.
[0003] Global horizontal irradiance (GHI) is one of the most important meteorological parameters related to solar energy. Due to its highly volatile nature, accurate ultra-short-term GHI forecasting has long been a challenging problem for meteorologists. Furthermore, most current GHI data is station-based, and this low spatial resolution makes it difficult to support the need for small-scale, refined GHI forecasts. Therefore, this paper proposes an ultra-short-term high-resolution horizontal irradiance forecasting method based on an improved TabNet and Earthformer deep learning model chain, combining satellite data from Fengyun-4B with a deep learning model chain, to promote the development of high spatiotemporal resolution ultra-short-term solar energy forecasting. Summary of the Invention
[0004] The purpose of this invention is to provide a high-resolution regional horizontal irradiance forecasting method based on a deep learning model chain. The forecasted GHI resource distribution map has a spatial resolution of 4KM and supports refined forecasting of solar photovoltaic power generation resource stations in any region.
[0005] To achieve the above objectives, the present invention provides the following solution: A method for high-resolution horizontal irradiance forecasting of regions based on deep learning model chains includes the following steps: S1. Acquire and preprocess the measurement dataset, including remote sensing data from the FY4B satellite and ground-based GHI measurement data from ground radiation measurement stations; S2. Perform spatiotemporal matching between remote sensing data and GHI ground-based measured data, and perform quality control on the matched remote sensing data and GHI ground-based measured data, deleting outliers and missing values. S3. Calculate astronomical characteristic parameters, including Earth-Sun distance, solar zenith angle, and solar azimuth angle, based on the geographical and time information of the station, including latitude and longitude, topography, and time. S4. Construct an input feature set including ground-measured GHI data and corresponding remote sensing data, astronomical feature parameters and geographic time information, and use the input feature set to train a pre-set improved TabNet model. In this model, a ReLU activation function is added after the final fully connected layer of the TabNet model to obtain an improved TabNet model that outputs non-negative GHI inversion values. S5. Based on the trained improved TabNet model, input the dataset to be inverted to generate a GHI inversion image dataset with high spatiotemporal resolution covering the target area and time period; S6. Divide the GHI inversion image dataset into training set, validation set and test set, and use the training set, validation set and test set to train the pre-defined improved Earthformer model. S7. Input the dataset to be predicted into the trained improved Earthformer model to achieve accurate prediction of regional horizontal irradiance.
[0006] Preferably, in S2, the remote sensing data is spatiotemporally matched with the GHI ground-based measured data, specifically including: For each ground station, the measured GHI ground data at each time point are correlated with the remote sensing data. All remote sensing data whose latitude and longitude differences are within a preset threshold range are selected, and the average value of each spectral band data is calculated to represent the satellite observation value of the station at the corresponding time point.
[0007] Preferably, in S2, quality control is performed on the matched remote sensing data and the GHI ground-based measured data, specifically including: (1) Delete the data in the remote sensing data where all channels are 0 after matching, and delete the data in the remote sensing data where any channel has missing values after matching; (2) At the same geographical location, the matched GHI measured data is less than the extraterrestrial solar radiation reaching the upper atmospheric boundary at this location; under cloudy conditions, the matched GHI measured data that are less than the preset threshold will be deleted.
[0008] Preferably, the formula for calculating the extraterrestrial solar radiation reaching the upper atmospheric boundary at this location is:
[0009] in Represents the solar constant. Indicates the average distance between the Earth and the Sun. Indicates the Earth-Sun distance on a given day. It represents the solar zenith angle.
[0010] Preferably, in S4, a pre-defined improved TabNet model is trained using the input feature set, specifically including: The model was trained using the Adam optimizer with a learning rate of 0.01. The hyperparameters were set as follows: the width of the decision prediction layer was 16, the width of the attention embedding of each mask was 16, the number of decision steps in the structure was 3, and the rest of the settings were the default settings of the model. The number of early stopping rounds was 20.
[0011] Preferably, in S5, the improvement of the Earthformer model is as follows: a ReLU activation function is added to the output of the Earthformer model to obtain an improved Earthformer model, which takes historical GHI image sequences as input and outputs future non-negative GHI prediction image sequences.
[0012] Preferably, in S6, a pre-defined improved Earthformer model is trained using a training set, a validation set, and a test set, specifically including: Input the GHI inversion images of the first 8 time steps and output the GHI forecast images of the next 8 time steps; perform Min-Max normalization on the input data; use the AdamW optimizer with a learning rate of 0.001 and set the number of early stopping rounds to 10.
[0013] Preferably, in S7, one or more of the following indicators are used to evaluate the forecasting performance of the improved Earthformer model: root mean square error, mean absolute error, mean deviation error, correlation coefficient, daily accuracy, and daily pass rate.
[0014] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements a method for regional high-resolution horizontal irradiance forecasting based on a deep learning model chain as described above.
[0015] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects: This invention proposes an ultra-short-term high-resolution horizontal irradiance forecasting method based on an improved TabNet and an improved Earthformer deep learning model chain. A ReLU activation function is added to the end of both models, effectively solving the problem of deep learning models inverting or forecasting GHI as negative. Furthermore, the spatial resolution of the GHI resource distribution map predicted by this invention is 4 km, supporting refined forecasting of solar photovoltaic power generation resource stations in any region. Evaluation of this invention using the latest energy meteorology industry standard "Accuracy Evaluation Method for Short-Term and Ultra-Short-Term Solar Radiation Forecasts of Photovoltaic Power Stations" (QX / T 771—2025) issued by the China Meteorological Administration shows that the forecasting model achieved high daily accuracy and excellent pass rate across all time periods. Attached Figure Description
[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0017] Figure 1 A flowchart illustrating the high-resolution horizontal irradiance forecasting method for regions based on deep learning model chains provided by this invention. Figure 2 This is a schematic diagram of the neural network structure of the deep learning model chain proposed in this invention; Figure 3 This is a graph showing the accuracy variation trend of each time step in southern my country from June to December 2023 in an embodiment of the present invention. Among them, (a) is the graph of average RMSE changing over time, (b) is the graph of average MBE changing over time, (c) is the graph of average daily accuracy changing over time, and (d) is the graph of average daily pass rate changing over time. Detailed Implementation
[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0019] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0020] like Figure 1 As shown, the present invention provides a method for predicting high-resolution horizontal irradiance in a region based on a deep learning model chain, comprising the following steps: S1. Acquire and preprocess the measurement dataset, including remote sensing data from the FY4B satellite and ground-based GHI measurement data from ground radiation measurement stations; S2. Perform spatiotemporal matching between remote sensing data and GHI ground-based measured data, and perform quality control on the matched remote sensing data and GHI ground-based measured data, deleting outliers and missing values. S3. Calculate astronomical characteristic parameters, including Earth-Sun distance, solar zenith angle, and solar azimuth angle, based on the geographical and time information of the station, including latitude and longitude, topography, and time. S4. Construct an input feature set including ground-measured GHI data and corresponding remote sensing data, astronomical feature parameters and geographic time information, and use the input feature set to train a pre-set improved TabNet model. In this model, a ReLU activation function is added after the final fully connected layer of the TabNet model to obtain an improved TabNet model that outputs non-negative GHI inversion values. S5. Based on the trained improved TabNet model, input the dataset to be inverted to generate a GHI inversion image dataset with high spatiotemporal resolution covering the target area and time period; S6. Divide the GHI inversion image dataset into training set, validation set and test set, and use the training set, validation set and test set to train the pre-defined improved Earthformer model. S7. Input the dataset to be predicted into the trained improved Earthformer model to achieve accurate prediction of regional horizontal irradiance.
[0021] Specifically, this invention proposes a high-resolution regional horizontal irradiance forecasting method based on a deep learning model chain. This method utilizes full-disk L1-level data from the FY4B geostationary meteorological satellite, sourced from the advanced geostationary orbit radiometric imager on FY4B, comprising 15 spectral bands with a temporal resolution of 15 minutes and a spatial resolution of 4 km. It also uses hourly ground-based GHI data from 100 radiation measurement stations of the China Meteorological Administration for 2022-2023 to train an improved TabNet model. Input variables include longitude, latitude, and topography for each station; 15 channels of FY4A satellite data; Earth-Sun distance; solar zenith angle; and solar azimuth angle, totaling 21 different feature variables; the inverted variable is GHI. After training, post-processing yields a high-precision 4 km resolution inverted GHI map. This inverted map is then used as training data to train the improved Earthformer model, which outputs forecast results for the next 0-2 hours. The method includes the following steps: Step S1: Download the FY4B full-disk L1 level 4KM resolution data from June 1, 2022 to December 31, 2023, a total of 15 spectral bands. All data for each day need to be downloaded. Preprocess the hourly GHI ground-based measured data from June 2022 to May 2023 provided by 100 radiation measurement stations of the China Meteorological Administration, and delete the missing GHI values.
[0022] Step S2: Match the satellite data at each moment with the GHI data. The matching rule is to calculate the average of the satellite data that are within 0.04° of the station's latitude and longitude, so as to represent the satellite data of 15 spectral bands at this moment. At the same time, missing data in the FY4B satellite need to be removed.
[0023] Step S201: Remove data from the FY4B satellite L1 level data where all channels are 0.
[0024] Step S202: Remove data with missing values in any channel of the FY4B satellite L1 level data.
[0025] Step S3: Calculate the Earth-Sun distance, solar zenith angle, and solar altitude angle of each station at a given moment based on the station's geographic time information, including latitude and longitude, topography, and time, and use these as input features for the model.
[0026] Step S4: Further quality control of the GHI data is performed, following the following two principles: (1) At the same geographical location, the GHI needs to be less than the extraterrestrial solar radiation reaching the upper atmospheric boundary at that location. (2) Under cloudy conditions, the GHI needs to be greater than the expected minimum value.
[0027] Step S401: Extraterrestrial solar radiation reaching the upper atmospheric boundary at a certain location The calculation formula is: (1)
[0028] In the formula Represents the solar constant. Indicates the average distance between the Earth and the Sun. Indicates the Earth-Sun distance on a given day. This represents the solar zenith angle. Calculate... Then, the GHI at each time point is subjected to quality control to handle outliers.
[0029] Step S402: Remove values of GHI less than 10 W / m2.
[0030] Step S5: Improve the TabNet model by adding a ReLU activation function after the final fully connected layer of TabNet, taking advantage of the non-zero nature of GHI. Input all 100 stations into the improved TabNet model and perform modeling to obtain the inverse GHI map.
[0031] Step S501: During the training of the improved Tabnet model, the specific training method of the improved Tabnet model is to use the Adam optimizer for training, with a learning rate of 0.01. The hyperparameters of the model are set as follows: the width of the decision prediction layer n_a is 16, the width of the attention embedding of each mask n_d is 16, the number of decision steps in the structure n_steps is 3, and the remaining settings are the default settings of the model. The number of early stopping rounds is 20.
[0032] Step S502: After training is completed, it is necessary to construct an input feature dataset. The data of each channel in the FY4B satellite data from June 1, 2022 to December 31, 2023 is flattened into one dimension and matched with the corresponding latitude, longitude and altitude. The Earth-Sun distance, solar zenith angle and solar azimuth angle are calculated to obtain an input feature set with a resolution of 4KM.
[0033] Step S503: Input the input feature set into the trained model to obtain the inverted GHI and visualize it, thereby obtaining a high-precision 4KM resolution inversion map of GHI from June 1, 2022 to December 31, 2023.
[0034] Step S6: Improve the Earthformer model by adding a ReLU activation function at the end of the Earthformer, taking advantage of the non-zero nature of GHI. Then, divide the inverted GHI images into training, validation, and test sets. The training set covers data from June 2022 to May 2023, 4:00 AM to 8:00 PM; the test set covers data from June to December 2023, 4:00 AM to 8:00 PM. The validation set consists of 10% randomly selected data from the training set. The spatiotemporal resolution of the GHI images is 15 minutes and 4 km. When feeding images into the model for training, the first 8 images are input, and the next 8 images are predicted. Min-Max normalization is performed to scale the data to the 0-1 range. During the training of the improved Earthformer model, the Adamw optimizer is used with a learning rate of 0.001, 10 early stopping epochs, and the hyperparameters follow the model's default settings.
[0035] Step S7: Evaluate the Earthformer model's prediction performance on the test set. The statistics used in the evaluation are the root mean square error (RMSE), correlation coefficient (R), mean bias error (MBE), and mean absolute error (MAE). RMSE, MBE, and MAE are defined as follows:
[0036] In the formula To improve the TabNet model's GHI value inversion, To improve the Earthformer model's prediction of GHI values, the correlation coefficient R is the Pearson correlation coefficient, calculated using the following formula:
[0037] In the formula To improve the covariance of the TabNet model's inverted GHI values and the Earthformer model's predicted values, To improve the standard deviation of the GHI value inversion in the TabNet model, To improve the standard deviation of the Earthformer model's prediction of GHI values.
[0038] Secondly, daily accuracy and daily pass rate are also used. The daily accuracy of the current short-term or ultra-short-term solar radiation forecast is defined as follows:
[0039] In the formula This indicates the total number of irradiance samples. Indicates the number of days The forecast total irradiance at a given time, expressed in watts per square meter (W / m²). 2 ); Indicates the number of days The measured total irradiance at a given time, expressed in watts per square meter (W / m²). 2 ); This indicates that the measured average total irradiance within the available sunshine hours on that day does not exceed 250 watts per square meter (W / m²). 2 )hour, The value is 600; the measured average total irradiance within the range of available sunshine hours on that day is greater than 250 watts per square meter (W / m²). 2 When I is taken as the maximum value of the total irradiance measured on that day, the unit is watts per square meter (W / m²). 2 (The same applies below.)
[0040] The daily short-term or ultra-short-term solar radiation forecast. The pass rate at any time ( ) and daily pass rate ( ) is defined as:
[0041] In the formula, Indicates the number of days The accuracy rate of solar radiation forecasts at any given time. If the forecast is good, it can be used to make GHI forecasts for the next 2 hours at a 4KM resolution.
[0042] Example 1 The following practical case will be used to verify the feasibility and accuracy of the GHI prediction in this invention.
[0043] The deep learning model chain ultra-short-term GHI forecasting method proposed in this invention, once trained, can forecast GHI resource distribution maps at any time with a 4KM resolution, occurring every 15 minutes. Furthermore, it is entirely based on domestically produced data and performed on a domestically developed platform. Here, we use a single forecast result as an example, demonstrating the high-resolution 4KM GHI forecast performance from 12:00 to 13:00 Beijing time on June 1, 2023, to illustrate the effectiveness of this method for GHI forecasting.
[0044] The steps for obtaining the GHI resource distribution from 12:00 to 13:00 on June 1, 2023 using the GHI forecasting method of this invention are as follows: 1. For example Figure 1 As shown, firstly, the inversion data training set of the improved TabNet model is prepared through steps S1-S4. Then, the training set is fed into the improved TabNet model for training according to the hyperparameter settings mentioned in the specific implementation step S501. After training, a GHI resource distribution map with a resolution of 4KM can be obtained.
[0045] 2. Following the methods described in implementation steps S502 and S503, obtain the GHI resource distribution map from June 1, 2022 to December 31, 2023. Then, improve the Earthformer model and set the hyperparameters according to the method in implementation step S6. Divide the GHI resource distribution map from June 1, 2022 to December 31, 2023 into training, validation, and test sets. The training set covers the period from 4:00 AM to 8:00 PM from June 2022 to May 2023, and the test set covers the period from 4:00 AM to 8:00 PM from June to December 2023. The validation set consists of 10% of the data randomly selected from the training set. Figure 2 It demonstrates the architecture of the entire deep learning model chain.
[0046] The accuracy and pass rate levels are classified as follows: For the grid dispatching side, the monthly average accuracy of day-ahead short-term and ultra-short-term solar radiation forecasts for photovoltaic power plants is divided into three levels: high, medium, and low. The accuracy thresholds for different levels should conform to the provisions of Table 1. For the grid dispatching side, the monthly average pass rate of day-ahead short-term and ultra-short-term solar radiation forecasts for photovoltaic power plants is divided into three levels: excellent, good, and poor. The pass rate thresholds for different levels should conform to the provisions of Table 2.
[0047] 3. Evaluate the accuracy of the improved Earthformer model on the test set.
[0048] Table 3 presents the overall accuracy ranking of the deep learning model chain on the test set in southern my country from June to December 2023. It can be seen that the errors are low in each month, and the daily accuracy is high, the daily pass rate is excellent, and the correlation coefficient is also high, proving that the improved Earthformer model can effectively learn the temporal variations of GHI.
[0049] from Figure 3 As shown in (a)-(d), the prediction error gradually increases with the increase of the prediction time. However, the daily accuracy and daily pass rate are both at a high and excellent level in the 0-2 hour forecast, which proves that the deep learning model chain proposed in this invention has excellent prediction capabilities. Secondly, the mean bias error (MBE) is low at all time steps, which reflects that the improved Earthformer model does not have a significant underestimation or overestimation phenomenon due to the increase of time steps.
[0050] The above results demonstrate that the deep learning model chain ultra-short-term horizontal irradiance prediction method proposed in this invention has excellent applicability.
[0051] Table 1
[0052] Table 2
[0053] Table 3
[0054] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements a method for regional high-resolution horizontal irradiance forecasting based on a deep learning model chain as described above.
[0055] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.
[0056] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.
Claims
1. A method for predicting high-resolution horizontal irradiance in a region based on a deep learning model chain, characterized in that, Includes the following steps: S1. Acquire and preprocess the measurement dataset, including remote sensing data from the FY4B satellite and ground-based GHI measurement data from ground radiation measurement stations; S2. Perform spatiotemporal matching between remote sensing data and GHI ground-based measured data, and perform quality control on the matched remote sensing data and GHI ground-based measured data, deleting outliers and missing values. S3. Calculate astronomical characteristic parameters, including Earth-Sun distance, solar zenith angle, and solar azimuth angle, based on the geographical and time information of the station, including latitude and longitude, topography, and time. S4. Construct an input feature set including ground-measured GHI data and corresponding remote sensing data, astronomical feature parameters and geographic time information, and use the input feature set to train a pre-set improved TabNet model. In this model, a ReLU activation function is added after the final fully connected layer of the TabNet model to obtain an improved TabNet model that outputs non-negative GHI inversion values. S5. Based on the trained improved TabNet model, input the dataset to be inverted to generate a GHI inversion image dataset with high spatiotemporal resolution covering the target area and time period; S6. Divide the GHI inversion image dataset into training set, validation set and test set, and use the training set, validation set and test set to train the pre-defined improved Earthformer model. S7. Input the dataset to be predicted into the trained improved Earthformer model to achieve accurate prediction of regional horizontal irradiance.
2. The method for high-resolution horizontal irradiance forecasting of regions based on a deep learning model chain according to claim 1, characterized in that, In step S2, the remote sensing data is spatiotemporally matched with the GHI ground-based measured data, specifically including: For each ground station, the measured GHI ground data at each time point are correlated with the remote sensing data. All remote sensing data whose latitude and longitude differences are within a preset threshold range are selected, and the average value of each spectral band data is calculated to represent the satellite observation value of the station at the corresponding time point.
3. The method for high-resolution horizontal irradiance forecasting of regions based on a deep learning model chain according to claim 1, characterized in that, In step S2, quality control is performed on the matched remote sensing data and the GHI ground-based measured data, specifically including: (1) Delete the data in the remote sensing data where all channels are 0 after matching, and delete the data in the remote sensing data where any channel has missing values after matching; (2) At the same geographical location, the matched GHI measured data is less than the extraterrestrial solar radiation reaching the upper atmospheric boundary at this location; under cloudy conditions, the matched GHI measured data that are less than the preset threshold will be deleted.
4. The method for regional high-resolution horizontal irradiance forecasting based on a deep learning model chain according to claim 3, characterized in that, The formula for calculating the extraterrestrial solar radiation reaching the upper atmospheric boundary at this location is: in Represents the solar constant. Indicates the average distance between the Earth and the Sun. Indicates the Earth-Sun distance on a given day. It represents the solar zenith angle.
5. The method for high-resolution horizontal irradiance forecasting of regions based on a deep learning model chain according to claim 1, characterized in that, In step S4, training a pre-defined improved TabNet model using the input feature set specifically includes: The model was trained using the Adam optimizer with a learning rate of 0.
01. The hyperparameters were set as follows: the width of the decision prediction layer was 16, the width of the attention embedding of each mask was 16, the number of decision steps in the structure was 3, and the rest of the settings were the default settings of the model. The number of early stopping rounds was 20.
6. The method for high-resolution horizontal irradiance forecasting of regions based on a deep learning model chain according to claim 1, characterized in that, In S5, the improvement of the Earthformer model is as follows: a ReLU activation function is added to the output of the Earthformer model to obtain an improved Earthformer model, which takes historical GHI image sequences as input and outputs future non-negative GHI prediction image sequences.
7. The method for high-resolution horizontal irradiance forecasting of regions based on a deep learning model chain according to claim 1, characterized in that, In step S6, a pre-defined improved Earthformer model is trained using a training set, a validation set, and a test set, specifically including: Input the GHI inversion images of the first 8 time steps and output the GHI forecast images of the next 8 time steps; perform Min-Max normalization on the input data; use the AdamW optimizer with a learning rate of 0.001 and set the number of early stopping rounds to 10.
8. The method for high-resolution horizontal irradiance forecasting of regions based on a deep learning model chain according to claim 1, characterized in that, S7 further includes using one or more of the following indicators to evaluate the forecasting performance of the improved Earthformer model: root mean square error, mean absolute error, mean deviation error, correlation coefficient, daily accuracy, and daily pass rate.
9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements a method for predicting regional high-resolution horizontal irradiance based on a deep learning model chain as described in any one of claims 1-8.