Solar radiation forecasting method fusing decomposition algorithm and GraphCast model

Abstract

The invention discloses a solar radiation forecasting method fusing a decomposition algorithm and a GraphCast model, and relates to the technical field of solar radiation forecasting. Comprising the following steps: acquiring meteorological variable data in an ERA5 reanalysis open source database; preprocessing the data to obtain meteorological variable data conforming to the input format of the GraphCast model; a GraphCast model is constructed to carry out weather forecast; acquiring actually measured meteorological variable data and solar radiation data of a certain region site; through an XGBoost feature screening algorithm, meteorological variables with high importance relative to solar radiation are screened out; preprocessing the screened meteorological variables by using a VMD decomposition algorithm, and constructing a training set and a test set by using the preprocessed data; taking actually measured solar radiation as a label of the model, and performing training optimization on the RF model by using data of the training set; and on the basis of a GraphCast coupling VMD decomposition algorithm hybrid RF model, solar radiation is predicted according to meteorological data of a test set. According to the invention, solar radiation forecasting with higher precision can be achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of solar radiation prediction, and in particular to a solar radiation prediction method integrating a decomposition algorithm and a GraphCast model. Background Art

[0002] Solar radiation is a clean, renewable energy source that reaches the Earth's surface directly from the sun without further processing. It plays a vital role in maintaining the Earth's atmospheric system. Solar radiation can be predicted based on empirical and physical methods, such as the sunshine duration-based (SDB) model, the cloud cover-based (CB) model, and the temperature-based (TB) model. These methods are sensitive to different geographical locations and different climatic conditions, and often require targeted modeling to make high-precision forecasts. Physically based methods, such as the Radiation Transfer Library (LibRadtran) and the Weather Research and Forecasting (WRF) model. However, these physical models are logically complex and computationally expensive. Secondly, these models require a large amount of meteorological data, which is difficult to obtain in some areas, which directly leads to the inability to forecast solar radiation.

[0003] The complex nonlinearity of solar radiation has become a difficult problem for accurate forecasting. At the same time, there is a relatively complex linear relationship between solar radiation and other meteorological variables. Machine learning has become a mainstream weather forecasting method. It effectively solves the complex linear relationship between meteorological variables and solves the nonlinear problem in solar radiation prediction caused by the randomness of climate conditions. In the forecast of solar radiation, the forecast effect of machine learning has been recognized, and the expected effect can be achieved with less meteorological data, such as support vector machine (SVM), artificial neural network (ANN), convolutional neural network (CNN) and long short-term memory (LSTM). The combination of preprocessing methods such as wavelet transform and bionic algorithm with machine learning has also been proven to have better performance in solar radiation forecasting. However, traditional machine learning algorithms still face the problems of high computational cost and complex parameter adjustment in mesoscale global weather forecasting.

[0004] GraphCast is a large-scale global weather forecast model based on machine learning. It makes up for the computational cost and parameter complexity of traditional machine learning algorithms. GraphCast can predict the weather conditions for the next ten days within 1 minute. It processes large amounts of data far better than traditional machine learning algorithms. At the same time, it captures the complex nonlinear relationships of multiple meteorological variables very well, achieving accurate forecasts. However, GraphCast does not forecast solar radiation.

[0005] Therefore, it is an urgent problem for those skilled in the art to propose a solar radiation forecasting method that integrates the decomposition algorithm and the GraphCast model to solve the difficulties existing in the prior art. Summary of the invention

[0006] In view of this, the present invention provides a solar radiation forecasting method that integrates the decomposition algorithm and the GraphCast model. The meteorological variables predicted by GraphCast can be used to further improve the performance of the RF model forecast through the VMD decomposition method to achieve a more accurate solar radiation forecast.

[0007] In order to achieve the above object, the present invention adopts the following technical solution:

[0008] A solar radiation forecasting method integrating a decomposition algorithm and a GraphCast model, comprising:

[0009] S1. Obtain meteorological variable data in the ERA5 reanalysis open source database according to the GraphCast model input requirements;

[0010] S2, preprocessing the acquired meteorological variable data to obtain meteorological variable data that conforms to the GraphCast model input format;

[0011] S3, build GraphCast model for weather forecast;

[0012] S4, obtaining the measured meteorological variable data and solar radiation data of a station in a certain area;

[0013] S5. Using the XGBoost feature screening algorithm, based on the measured meteorological variable data and solar radiation data of the site in step S4, screen out meteorological variables with high importance relative to solar radiation;

[0014] S6, using the VMD decomposition algorithm to preprocess the meteorological variables screened in step S5, and construct a training set and a test set with the preprocessed data;

[0015] S7, using the measured solar radiation as the label of the model, and using the data of the training set to train and optimize the RF model;

[0016] S8. Based on the hybrid RF model coupled with GraphCast VMD decomposition algorithm, solar radiation is predicted according to the test set meteorological data.

[0017] In the above method, optionally, in S1, the meteorological variable data includes: surface data and meteorological variables at 13 levels above the ground;

[0018] Surface data include: easterly component of wind speed at 10 meters above the ground, northerly component of wind speed at 10 meters above the ground, temperature at 2 meters above the ground, land-sea mask, mean sea level pressure, incident shortwave solar radiation at the top of the atmosphere, and 6-hour total precipitation;

[0019] Meteorological variables at 13 levels above the ground include: gravitational potential energy per unit mass relative to mean sea level, specific humidity of water vapor per kilogram of moist air, air temperature, easterly component of wind speed, northerly component of wind speed, and vertical air velocity upward or downward;

[0020] Among them, there are 13 levels from the ground: 50m, 100m, 150m, 200m, 250m, 300m, 400m, 500m, 600m, 700m, 850m, 925m and 1000m from the ground.

[0021] In the above method, optionally, in S2, the downloaded ERA5 data is preprocessed as follows:

[0022] The python netCDF4 library was used to rename the meteorological variables according to the input requirements of the GraphCast model. The python NumPyd library was used to mirror the ERA5 data. Through multiple for loops, ERA5 was injected into the generated network common data format file to generate the input file of the GraphCast model.

[0023] In the above method, optionally, in S3, constructing the GraphCast model specifically includes: constructing a checkpoint parameter, a high-resolution GraphCast pre-trained model with a resolution of 0.25 degrees, including 37 pressure levels, and a model trained with ERA5 data from 1979 to 2017 to forecast the meteorological variables included in step S1.

[0024] In the above method, optionally, in S5, XGBoost is used to screen out meteorological variables predicted by the GraphCast model that are more important than solar radiation. The XGBoost feature screening is expressed as:

[0025]

[0026] Where j represents the feature, N j is the number of times feature j is used for splitting, ΔLoss i It is the reduction in the loss function when the feature is used for splitting for the i-th time.

[0027] In the above method, optionally, in step S6, VMD decomposition is performed on the result of feature screening in step S5, and the meteorological data predicted by the GraphCast model is denoised. The VMD decomposition algorithm optimizes the meteorological data predicted by the GraphCast model, which is expressed as:

[0028]

[0029] Where f(t) is the original signal, u k (t) is the kth mode, ω k is the corresponding center frequency, means moving the mode to its center frequency, represents the time derivative.

[0030] In the above method, optionally, in S7, each mode after VMD decomposition is put into the RF model, and the measured value of solar radiation in step S4 is used as a label for training, and the predicted value of each mode is added to obtain the final solar radiation forecast value, which is specifically expressed as:

[0031]

[0032] Where f(k) is the RF model for the input u k The predicted value, u k is the kth mode of VMD decomposition, N is the total number of decision trees in random forest, T i is the i-th tree for input u k The predicted value of .

[0033] It can be seen from the above technical solution that, compared with the prior art, the present invention provides a solar radiation forecasting method that integrates a decomposition algorithm and a GraphCast model, which has the following beneficial effects: the present invention utilizes the meteorological variables predicted by the GraphCast model and decomposes and preprocesses the forecast results through a VMD decomposition algorithm to construct an improved GraphCast RF model for forecasting solar radiation. By using the VMD method to decompose the GraphCast meteorological variable data, the nonlinearity of the data is reduced, the noise is removed, and the forecasting performance of the RF model is improved, thereby improving the gap in GraphCast forecasting solar radiation, and the accuracy and forecasting effect of the solar radiation forecast are improved through the VMD decomposition method. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0035] Figure 1 A flow chart of a solar radiation forecasting method that integrates a decomposition algorithm and a GraphCast model provided by the present invention;

[0036] Figure 2 A comparison chart of meteorological variables based on the GraphCast model and meteorological variables based on the site provided by the present invention;

[0037] Figure 3 A cross-validation graph of solar radiation forecast results of the RF model based on GraphCast after VMD decomposition provided by the present invention as input and GraphCast meteorological variables without VMD decomposition as RF input;

[0038] Figure 4 A cross-validation scatter plot of the predicted values ​​and measured values ​​of the specific embodiments provided by the present invention. DETAILED DESCRIPTION

[0039] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0040] In this application, relational terms such as first and second, etc. are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. The terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of more restrictions, the elements defined by the sentence "comprise one..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0041] Reference Figure 1 As shown, the present invention discloses a solar radiation forecasting method integrating a decomposition algorithm and a GraphCast model, comprising:

[0042] S1. Obtain meteorological variable data in the ERA5 reanalysis open source database according to the GraphCast model input requirements;

[0043] S2, preprocessing the acquired meteorological variable data to obtain meteorological variable data that conforms to the GraphCast model input format;

[0044] S3, build GraphCast model for weather forecast;

[0045] S4, obtaining the measured meteorological variable data and solar radiation data of a station in a certain area;

[0046] S5. Using the XGBoost feature screening algorithm, based on the measured meteorological variable data and solar radiation data of the site in step S4, screen out meteorological variables with high importance relative to solar radiation;

[0047] S6, using the VMD decomposition algorithm to preprocess the meteorological variables screened in step S5, and construct a training set and a test set with the preprocessed data;

[0048] S7, using the measured solar radiation as the label of the model, and using the data of the training set to train and optimize the RF model;

[0049] S8. Based on the hybrid RF model coupled with GraphCast VMD decomposition algorithm, solar radiation is predicted according to the test set meteorological data.

[0050] Furthermore, in S1, the meteorological variable data include: surface data and meteorological variables at 13 levels above the ground;

[0051] Surface data include: easterly component of wind speed at 10 meters above the ground, northerly component of wind speed at 10 meters above the ground, temperature at 2 meters above the ground, land-sea mask, mean sea level pressure, incident shortwave solar radiation at the top of the atmosphere, and 6-hour total precipitation;

[0052] Meteorological variables at 13 levels above the ground include: gravitational potential energy per unit mass relative to mean sea level, specific humidity of water vapor per kilogram of moist air, air temperature, easterly component of wind speed, northerly component of wind speed, and vertical air velocity upward or downward;

[0053] Among them, there are 13 levels from the ground: 50m, 100m, 150m, 200m, 250m, 300m, 400m, 500m, 600m, 700m, 850m, 925m and 1000m from the ground.

[0054] Furthermore, in S2, the downloaded ERA5 data is preprocessed as follows:

[0055] The python netCDF4 library was used to rename the meteorological variables according to the input requirements of the GraphCast model. The python NumPyd library was used to mirror the ERA5 data. Through multiple for loops, ERA5 was injected into the generated network common data format file to generate the input file of the GraphCast model.

[0056] Furthermore, in S3, constructing the GraphCast model specifically includes: constructing a checkpoint, a high-resolution GraphCast pre-trained model with a resolution of 0.25 degrees, including 37 pressure levels, a model trained with ERA5 data from 1979 to 2017, and forecasting the meteorological variables included in step S1, and the forecasted meteorological variables are the same as the input meteorological variables.

[0057] Furthermore, in S5, XGBoost is used to screen out meteorological variables predicted by the GraphCast model that are more important than solar radiation. The XGBoost feature screening is expressed as:

[0058]

[0059] Where j represents the feature, N j is the number of times feature j is used for splitting, ΔLoss i It is the reduction of the loss function when the feature is used for splitting for the i-th time.

[0060] Further, in S6, VMD decomposition is performed on the result of feature screening in step S5, and the meteorological data predicted by the GraphCast model is denoised. The VMD decomposition algorithm optimizes the meteorological data predicted by the GraphCast model, which is expressed as:

[0061]

[0062] Where f(t) is the original signal, u k (t) is the kth mode, ω k is the corresponding center frequency, means moving the mode to its center frequency, represents the time derivative.

[0063] Further, in S7, each mode after VMD decomposition is put into the RF model, and the measured value of solar radiation in step S4 is used as the label for training, and the predicted value of each mode is added to obtain the final solar radiation forecast value, which is specifically expressed as:

[0064]

[0065] Where f(k) is the RF model for the input u k The predicted value, u k is the kth mode of VMD decomposition, N is the total number of decision trees in random forest, T i is the i-th tree for input u k The predicted value of .

[0066] In a specific embodiment, the relevant data are explained by taking a certain region as the study area. The certain region (33°53′~38°54′N, 114°63′~122°02′W) spans inland deserts, high mountains and the Pacific coast, with a total area of ​​about 423,970 square kilometers. The climate types in this region are diverse, including Mediterranean climate, desert climate and mountain climate, and the solar radiation conditions are complex and changeable;

[0067] The California region was selected as the object of the study of solar radiation forecasting because of its diverse geographical locations and climates, which make the temporal and spatial variations of solar radiation complex. The accurate solar radiation forecasting for this region can better demonstrate the reliability of the present invention.

[0068] The forecast meteorological data for 2020-2021 were obtained from the ERA5 open source website. At the same time, the historical meteorological data of a certain place from 2020 to 2021 were collected. The ERA5 forecast meteorological data include: the easterly component of the wind speed 10 meters above the ground (10U), the northerly component of the wind speed 10 meters above the ground (10V), the temperature 2 meters above the ground (2T), the land-sea mask (LSM), the mean sea level pressure (MSL), the incident shortwave solar radiation at the top of the atmosphere (TOA), and the 6-hour total precipitation (TP). The historical meteorological data include: rainfall (p), maximum temperature (Tmax), minimum temperature (Tmin), daily average temperature (T), maximum relative humidity (RHmax), minimum relative humidity (RHmin), and daily average relative humidity (RH).

[0069] The data is preprocessed according to the meteorological data obtained in step S2, and the python netCDF4 library is used to rename the meteorological variables according to the GraphCast input requirements. In order to meet the GraphCast input latitude and longitude requirements, the python NumPyd library is used to mirror the ERA5 data. Finally, ERA5 is injected into the generated network common data format (NC) file through multiple for loops to generate the GraphCast input file.

[0070] In this embodiment, since some of the meteorological data obtained from 2020 to 2021 may not exist or be abnormal due to environmental interference or human operation, the abnormal data is eliminated, and 6,844 pieces of data are finally obtained after screening, and the training set and test set are divided by 5-fold cross-validation.

[0071] Since the data in the dataset have different dimensions and large values, the training set and test set data need to be preprocessed; specifically, the values ​​of the training set and test set data need to be standardized to improve the training performance of the model.

[0072] By standardizing the meteorological data measurements and using the maximum and minimum normalization method to standardize the meteorological data measurements to the range of [0, 1], the data of different dimensions are scaled to the same numerical unit, the forecasting performance of the model is improved and the overfitting phenomenon of the model is reduced.

[0073] A GraphCast model is constructed based on the processed input data to perform weather forecasting. The forecasted meteorological variables are the same as the input meteorological variables. In this example, the construction parameters are checkpoints, and a high-resolution GraphCast pre-trained model (0.25 degree resolution, 37 pressure levels, and a model trained based on ERA5 data from 1979 to 2017) is used to forecast global meteorological variables such as those in S2.

[0074] The XGBoost feature screening algorithm is used to screen out the meteorological variables predicted by GraphCast in step S4 that are more important than the measured solar radiation.

[0075] The VMD decomposition algorithm is used to preprocess the screened meteorological variables, and the training set and test set are constructed with the preprocessed data. The measured solar radiation is used as the label of the model, and the data of the training set is used to train and optimize the RF model.

[0076] Since GraphCast predicts many meteorological variables, it is necessary to consider reducing the impact of the model on high nonlinearity in the forecast of solar radiation. Therefore, the XGBoost feature screening algorithm is used to sort the feature importance values ​​of solar radiation measured at a certain site from high to low with a value of 0 to 1, and the meteorological variables with lower importance than the measured values ​​are screened out as the objects of the VMD decomposition algorithm. This can greatly improve the training performance and stability of the model. XGBoost feature screening can be expressed as:

[0077]

[0078] Where j represents the feature, N j is the number of times feature j is used for splitting, ΔLossi It is the reduction of the loss function when the feature is used for splitting for the i-th time.

[0079] The VMD decomposition algorithm is a further data preprocessing method for meteorological variables after XGBoost feature screening. In order to maximize the information extracted from the data and improve the forecast performance of solar radiation, the hybrid method is crucial. The VMD decomposition method decomposes the original signal into multiple sub-modes and a residual. The sub-modes are from the highest frequency to the lowest frequency, and the nonlinearity gradually decreases, which allows the model to more clearly capture the sub-mode information of each meteorological variable, thereby achieving a higher accuracy solar radiation forecast. VMD decomposition can be expressed as:

[0080]

[0081] Where f(t) is the original signal, u k (t) is the kth mode, ω k is the corresponding center frequency, means moving the mode to its center frequency, represents the time derivative.

[0082] Due to the preprocessing of meteorological variables predicted by GraphCast by XGBoost’s feature screening and VMD decomposition method, the training of the RF model can better capture the relationship between solar radiation and the preprocessed GraphCast meteorological variables, and better capture the characteristic relationship between them; in order to further improve the forecasting performance of the RF model, a larger number of decision trees of RF are set for training, and the number of decision trees is set to 200. After testing, when the other parameters are all default, satisfactory forecasting results can still be achieved. The forecast of solar radiation of RF forecast gas based on VMD decomposition of GraphCast can be expressed as:

[0083]

[0084]

[0085] Where f(k) is the RF model for the input u k The predicted value, u k is the kth mode of VMD decomposition, N is the total number of decision trees in random forest, T i is the i-th tree for input u k The predicted value of .

[0086] To verify the prediction effect of the method for predicting solar radiation by improving GraphCast based on the VMD decomposition algorithm and hybrid machine learning in the present invention, three prediction methods are compared: improving GraphCast based on the VMD decomposition algorithm and hybrid machine learning, GraphCast without the VMD decomposition algorithm and hybrid machine learning, and predicting solar radiation using RF based on the data of a certain site. The coefficient of determination (R 2 ), mean absolute error (MAE), and root mean square error (RMSE) are used to evaluate the performance of the model.

[0087] The calculation formulas for the above evaluation parameters are as follows:

[0088]

[0089] Among them, y i,o , y i,f and represent the observed values of solar radiation at the site, the predicted solar radiation by the model, and the average observed values of solar radiation at the site respectively. N represents the sample size. The range of R 2 is [0, 1]. The closer the MAE and RMSE are to 0, the higher the prediction performance of the model.

[0090] Referring to Figure 2 shown, in this embodiment, from the perspective of the prediction results of GraphCast for solar radiation and the prediction results based on the data of a certain site, the curve of the predicted value and the measured value of predicting solar radiation based on GraphCast meteorological variables is closer, and it has higher prediction accuracy.

[0091] To determine the fitting performance of the test dataset, cross-validation analysis of the calculated values and simulated values of the test dataset is carried out.

[0092] Referring to Figure 3 and Figure 4 shown, the coefficient of determination (R 2 ) and fitting accuracy of the method for predicting solar radiation by improving GraphCast based on the VMD decomposition algorithm and hybrid machine learning and the GraphCast hybrid machine learning without decomposition are both higher than 0.86. Compared with the model of GraphCast hybrid machine learning without VMD decomposition improvement, the prediction effect of the method for predicting solar radiation by improving GraphCast based on the VMD decomposition algorithm and hybrid machine learning is significantly better than the model without VMD decomposition, and it can more effectively construct the complex non-linear relationship between solar radiation and other meteorological driving factors.

[0093] In summary, the prediction accuracy of the method for improving GraphCast to predict solar radiation by combining the VMD decomposition algorithm with machine learning is higher than that of the RF model based on site meteorological data and the GraphCast hybrid machine learning model without VMD decomposition. The method for improving GraphCast to predict solar radiation by combining the VMD decomposition algorithm with machine learning can well capture the relationship between multiple meteorological variables and solar radiation. Without analyzing the underlying principles of meteorology, by leveraging the large model GraphCast and through pre-methods such as feature screening and VMD decomposition, the forecasting performance of the RF model is greatly improved, enabling accurate forecasting and thus promoting the development of large models in the fields of hydrology and meteorological science.

[0094] Each embodiment in this specification is described in a progressive manner. For the same or similar parts among the embodiments, reference can be made to each other. Each embodiment focuses on the differences from other embodiments. In particular, for a system or system embodiment, since it is basically similar to the method embodiment, the description is relatively simple. For the relevant parts, reference can be made to the partial description of the method embodiment. The systems and system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art can understand and implement it without creative efforts.

[0095] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be obvious to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but will be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A solar radiation forecasting method integrating decomposition algorithm and GraphCast model, characterized in that: include: S1. Obtain meteorological variable data in the ERA5 reanalysis open source database according to the GraphCast model input requirements; S2, preprocessing the acquired meteorological variable data to obtain meteorological variable data that conforms to the GraphCast model input format; S3, build GraphCast model for weather forecast; S4, obtaining the measured meteorological variable data and solar radiation data of a station in a certain area; S5. Using the XGBoost feature screening algorithm, based on the measured meteorological variable data and solar radiation data of the site in step S4, screen out meteorological variables with high importance relative to solar radiation; S6, using the VMD decomposition algorithm to preprocess the meteorological variables screened in step S5, and construct a training set and a test set with the preprocessed data; S7, using the measured solar radiation as the label of the model, and using the data of the training set to train and optimize the RF model; S8. Based on the hybrid RF model coupled with GraphCast VMD decomposition algorithm, solar radiation is predicted according to the test set meteorological data.

2. The solar radiation forecasting method of the fusion decomposition algorithm and the GraphCast model according to claim 1, characterized in that: In S1, meteorological variable data include: surface data and meteorological variables at 13 levels above the ground; Surface data include: easterly component of wind speed at 10 meters above the ground, northerly component of wind speed at 10 meters above the ground, temperature at 2 meters above the ground, land-sea mask, mean sea level pressure, incident shortwave solar radiation at the top of the atmosphere, and 6-hour total precipitation; Meteorological variables at 13 levels above the ground include: gravitational potential energy per unit mass relative to mean sea level, specific humidity of water vapor per kilogram of moist air, air temperature, easterly component of wind speed, northerly component of wind speed, and vertical air velocity upward or downward; Among them, there are 13 levels from the ground: 50m, 100m, 150m, 200m, 250m, 300m, 400m, 500m, 600m, 700m, 850m, 925m and 1000m from the ground.

3. The solar radiation forecasting method of integrating the decomposition algorithm and the GraphCast model according to claim 1, characterized in that: In S2, the downloaded ERA5 data is preprocessed as follows: The python netCDF4 library was used to rename the meteorological variables according to the input requirements of the GraphCast model. The python NumPyd library was used to mirror the ERA5 data. Through multiple for loops, ERA5 was injected into the generated network common data format file to generate the input file of the GraphCast model.

4. The solar radiation forecasting method of integrating the decomposition algorithm and the GraphCast model according to claim 1, characterized in that: In S3, building a GraphCast model specifically includes: building a checkpoint, a high-resolution GraphCast pre-trained model with a resolution of 0.25 degrees, including 37 pressure levels, and a model trained with ERA5 data from 1979 to 2017 to forecast the meteorological variables included in step S1.

5. The solar radiation forecasting method of integrating the decomposition algorithm and the GraphCast model according to claim 1, characterized in that: In S5, XGBoost is used to screen out meteorological variables predicted by the GraphCast model that are more important than solar radiation. The XGBoost feature screening is expressed as: Where j represents the feature, N j is the number of times feature j is used for splitting, ΔLoss i It is the reduction of the loss function when the feature is used for splitting for the i-th time.

6. The solar radiation forecasting method of integrating the decomposition algorithm and the GraphCast model according to claim 1, characterized in that: In S6, VMD decomposition is performed on the result of feature screening in step S5, and the meteorological data predicted by the GraphCast model is denoised. The VMD decomposition algorithm optimizes the meteorological data predicted by the GraphCast model, which is expressed as: Where f(t) is the original signal, u k (t) is the kth mode, ω k is the corresponding center frequency, means moving the mode to its center frequency, represents the time derivative.

7. The solar radiation forecasting method of integrating the decomposition algorithm and the GraphCast model according to claim 1, characterized in that: In S7, each mode after VMD decomposition is put into the RF model, and the measured value of solar radiation in step S4 is used as the label for training. The predicted value of each mode is added to obtain the final solar radiation forecast value, which is specifically expressed as: Where f(k) is the RF model for the input u k The predicted value, u k is the kth mode of VMD decomposition, N is the total number of decision trees in random forest, T i is the i-th tree for input u k The predicted value of .

Images