A deep learning-based all-weather solar radiation remote sensing forecasting method

Abstract

The present invention relates to a method for all-weather solar radiation remote sensing forecasting supported by deep learning, comprising the following steps: S1, acquiring time series multispectral remote sensing data; S2, performing data preprocessing on the acquired time series multispectral remote sensing data to obtain a preprocessed time series multispectral remote sensing dataset; S3, inputting the acquired preprocessed time series multispectral remote sensing dataset into a ConvLSTM network model to obtain instantaneous solar shortwave radiation forecast data or multi-timescale solar shortwave radiation forecast; the ConvLSTM network model is a deep learning network model oriented towards remote sensing images and incorporating a spatial attention mechanism. The forecasting method provided by the present invention fully considers the spatial and time series characteristics of solar radiation, and adds a spatial attention mechanism to the deep neural network to more effectively capture the sudden changes in solar radiation in nature.

Description

Technical Field

[0001] The present invention belongs to the technical field of remote sensing forecasting, and in particular relates to an all-weather solar radiation remote sensing forecasting method supported by deep learning. Background Art

[0002] In the field of solar photovoltaic forecasting (solar radiation forecasting), accurately predicting solar radiation trends and predicting solar radiation values in advance are extremely important for new energy industries such as photovoltaic power generation peak prediction and photovoltaic power station site selection. With the continuous development of deep neural network methods, some neural network models have begun to be applied to solar radiation forecasting and have demonstrated good performance. However, when predicting solar radiation, traditional deep neural network models rely on training and testing data that are mostly site data, that is, point data at a single location. In addition, the true values used by traditional numerical analysis methods are mostly measured data from a single site or various types of reanalysis data, which cannot accurately capture the spatial distribution characteristics and mutation characteristics of solar radiation in nature.

[0003] Given that solar radiation forecasting requires considering both spatial and temporal characteristics, previous deep learning algorithm models that only consider a single characteristic of solar radiation have unavoidable limitations. Furthermore, conventional deep learning networks often fail to achieve satisfactory prediction accuracy for time-series solar radiation, as they are unable to account for sudden changes in solar radiation caused by random variations in cloud position and structure. Summary of the Invention

[0004] (1) Technical issues to be resolved

[0005] In order to solve the above-mentioned problems of the prior art, the present invention provides an all-weather solar radiation remote sensing forecasting method supported by deep learning, which breaks through the problem that traditional neural networks cannot simultaneously take into account the absolute numerical accuracy and temporal and spatial characteristic distribution of solar radiation in the process of solar shortwave radiation prediction.

[0006] (2) Technical solution

[0007] In order to achieve the above objectives, the main technical solutions adopted by the present invention include:

[0008] A deep learning-supported all-weather solar radiation remote sensing forecasting method includes the following steps:

[0009] S1. Acquire time series multispectral remote sensing data;

[0010] S2. performing data preprocessing on the acquired time series multispectral remote sensing data to obtain a preprocessed time series multispectral remote sensing dataset;

[0011] S3. Input the obtained pre-processed time series multispectral remote sensing data set into the ConvLSTM network model to obtain instantaneous solar shortwave radiation prediction data or multi-time scale solar shortwave radiation prediction;

[0012] The ConvLSTM network model is a deep learning network model for remote sensing images and combines a spatial attention mechanism.

[0013] Preferably, the basic model formula of the ConvLSTM network model is as follows:

[0014] x=SA(x t ),H t-1 =SA(H t-1 )

[0015] i t =σ(W xi *X t +W hi *H t-1 +b i )

[0016] f t =σ(W xf *X t +W hf *H t-1 +b f )

[0017] g t =tanh(W xc *X t +W hc *H t-1 +b c )

[0018]

[0019] o t =σ(W xo *X t +W ho *H t-1 +b o )

[0020]

[0021] Among them, SA is the spatial attention module, which is used to participate in the distribution of weights between images. and It is a module aggregated by the attention mechanism, that is, the sum of the weighted values of the selectively acquired features at all positions at each time step;

[0022] W iis the weight matrix corresponding to each feature; H t is the feature of the current input time step;

[0023] i t is the input gate, which corresponds to the eigenvalue g after fusion t ;

[0024] By outputting the updated original memory feature o t , and then the weights of the attention mechanism are redistributed so that the features of each time step are updated;

[0025] * means convolution operation; o means corresponding multiplication, and the result of multiplication is Hadamard product;

[0026] b i is the bias of the corresponding item; σ is the activation function corresponding to the neural network, which is the ReLU function in the model; C t The middle layer is the memory of the previous moment, that is, the memory layer;

[0027] Each network layer is connected through convolution operations.

[0028] Preferably, the custom loss function of the ConvLSTM network model is as follows:

[0029] k×SSIM+RMSE

[0030] Among them, the K value is a weight adjustment factor, which is used to adjust the weight distribution between the structural similarity function SSIM and the root mean square error RMSE between images to achieve the optimal extreme value.

[0031] Preferably, the formula of the structural similarity function SSIM is as follows:

[0032] SSIM is written as S(x,y), which is numerically decomposed into a linear function combination of three components: l, c, and s:

[0033] S(x,y)=f(l(x,y),c(x,y),s(x,y))

[0034]

[0035] SSIM(x,y)=[l(x,y)] a [c(x,y)] β [s(x,y)] γ

[0036] Where x and y represent the predicted solar radiation image and the actual image, respectively; σ is the image variance; μ is the image mean; C1, C2, and C3 are stabilization functions used to balance the dynamic range of pixel values; α, β, and γ are adjustment coefficients, which usually default to 1; and the range of the structural similarity function is (-1, 1).

[0037] Preferably, the calculation formula of the root mean square error RMSE is:

[0038]

[0039] Among them, x i is the predicted data, y i is the real data, and N is the total amount of data.

[0040] Preferably, the method further comprises: before step S3, using part of the data set in step S2 to train the ConvLSTM network model to obtain a ConvLSTM network model with prediction accuracy that meets the requirements;

[0041] Before training the ConvLSTM network model, the method further includes: setting relevant model parameters of the ConvLSTM neural network;

[0042] The relevant model parameters include at least: the number of neural network iterative calculations, the batch input network data size, the activation function setting, the loss function setting and the number of neural network layers.

[0043] Preferably, when predicting instantaneous solar shortwave radiation prediction data,

[0044] The time series remote sensing data in step S1 includes three types of original data: TOA reflectivity data of 6 optical bands, solar zenith angle data of 1 band, and solar shortwave radiation data;

[0045] The data preprocessing in step S2 includes two parts:

[0046] The first part is data quality control, which includes checking the quality of the obtained data, removing outliers in the data, and separating the TOA reflectivity data and solar zenith angle data from the solar shortwave radiation data for future use.

[0047] The second part is to transform the time series remote sensing data into a dimension that matches the ConvLSTM neural network model, that is, to reconstruct the data dimension into a five-dimensional array tensor of data number, time step, row, column, and eigenvalue.

[0048] Preferably, the method further comprises: normalizing the reconstructed data preprocessed in step S2.

[0049] Preferably, when predicting multi-timescale solar shortwave radiation,

[0050] The time series remote sensing data in step S1 includes three types of original data: TOA reflectivity data of 6 optical bands, solar zenith angle data of 1 band, and solar shortwave radiation data;

[0051] The data preprocessing in step S2 includes three parts:

[0052] The first part is data quality control, which includes checking the quality of the obtained data, removing outliers in the data, and separating the TOA reflectivity data and solar zenith angle data from the solar shortwave radiation data for future use.

[0053] The second part is to perform multi-time scale averaging processing on the instantaneous satellite remote sensing data;

[0054] The third part is to transform the time series remote sensing data into a dimension that matches the ConvLSTM neural network model, that is, to reconstruct the data dimension into a five-dimensional array tensor of data number, time step, row, column, and eigenvalue.

[0055] Preferably, the method further comprises: normalizing the reconstructed data preprocessed in step S2.

[0056] (3) Beneficial effects

[0057] The beneficial effects of the present invention are as follows: the all-weather solar radiation remote sensing forecasting method supported by deep learning provided by the present invention has the following beneficial effects:

[0058] This application fully considers the spatial and time series characteristics of solar radiation, and adds a spatial attention mechanism to the deep neural network to more effectively capture the sudden changes in solar radiation in nature. For some high-frequency time series multispectral optical remote sensing satellite observation data, this application has the ability to predict "solar irradiation" at multiple time scales, and the overall prediction of the absolute value of instantaneous solar radiation in the region is RMSE ≤ 90w / m 2 , the overall prediction RMSE of the absolute value of solar radiation at multiple time scales is ≤30w / m 2 .

[0059] At the same time, this application can also provide early warning references for fluctuations in power generation of photovoltaic power stations under complex meteorological conditions in multiple regions. It only requires time series multispectral remote sensing satellite data to complete large-scale, quantitative, and near-real-time regional comprehensive solar radiation forecasts. The forecast data has good generalization and high accuracy, the forecast model is highly robust, and has better anti-interference capabilities. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1A schematic diagram of the instantaneous solar radiation prediction process in the all-weather solar radiation remote sensing prediction method supported by deep learning provided by the present invention;

[0061] Figure 2 This is a multi-time-scale solar radiation prediction flowchart in the all-weather solar radiation remote sensing forecasting method supported by deep learning provided by the present invention. DETAILED DESCRIPTION

[0062] In order to better explain the present invention and facilitate understanding, the present invention is described in detail below through specific implementation methods in conjunction with the accompanying drawings.

[0063] Example 1

[0064] like Figure 1 and Figure 2 This embodiment discloses a method for all-weather solar radiation remote sensing forecasting supported by deep learning, including the following steps:

[0065] S1. Acquire time series multispectral remote sensing data. It should be noted that the time series multispectral remote sensing data here includes continuous time series multispectral remote sensing data and instantaneous time series multispectral remote sensing data.

[0066] S2. performing data preprocessing on the acquired time series multispectral remote sensing data to obtain a preprocessed time series multispectral remote sensing dataset;

[0067] S3. Input the obtained pre-processed time series multispectral remote sensing data set into the ConvLSTM network model to obtain instantaneous solar shortwave radiation forecast data or multi-time scale solar shortwave radiation forecast; the ConvLSTM network model obtains normalized solar shortwave radiation data, and further multiply the obtained normalized solar shortwave radiation data by the maximum solar radiation value of 1367w / m 2 By performing the inverse normalization operation, we can obtain the multi-time scale mean solar shortwave radiation data starting from the current instantaneous moment or the real instantaneous solar shortwave radiation prediction data in nature.

[0068] The ConvLSTM network model is a deep learning network model for remote sensing images and combines a spatial attention mechanism. This network model only needs to input sequential multispectral optical remote sensing satellite data with an advance time (e.g., the previous 30 minutes) to achieve instantaneous solar radiation and multi-time scale solar radiation prediction for a certain area at the 10-30 minute level.

[0069] Compared with existing prediction models, this model that simultaneously considers multiple influencing factors can better perceive sudden changes in solar radiation in nature. Model training requires fewer input parameters, the amount of data required for model prediction is smaller, and the model has stronger robustness and generalization capabilities during the prediction process.

[0070] In detail, the basic model formula of the ConvLSTM network model described in this embodiment is as follows:

[0071] x=SA(x t ),H t-1 =SA(H t-1 )

[0072] i t =σ(W xi *X t +W hi *H t-1 +b i )

[0073] f t =σ(W xf *X t +W hf *H t-1 +b f )

[0074] g t =tanh(W xc *X t +W hc *H t-1 +b c )

[0075]

[0076] o t =σ(W xo *X t +W ho *H t-1 +b o )

[0077]

[0078] Among them, SA is the spatial attention module, which is used to participate in the distribution of weights between images. and is the module obtained by aggregation through the attention mechanism, that is, the sum of the weighted values of the selectively acquired features at all positions in each time step; W i is the weight matrix corresponding to each feature; H t is the feature of the current input time step; i t is the input gate, which corresponds to the eigenvalue g after fusiont ; By outputting the updated original memory feature o t , and then the weights of the attention mechanism are redistributed so that the features of each time step are updated; * means convolution operation; o means corresponding multiplication, and the multiplication result is the Hadamard product; b i is the bias of the corresponding item; σ is the activation function corresponding to the neural network, which is the ReLU function in the model; C t The middle layer stores the memory of the previous moment, that is, the memory layer; the layers are connected by convolution operations.

[0079] In this embodiment, in order to solve the problem of simultaneously predicting the absolute value of solar radiation and the spatiotemporal distribution of radiation, a new neural network loss function (i.e., a new custom Loss function) is first designed based on taking into account both spatial position prediction and absolute value prediction.

[0080] This application takes full account of:

[0081] (1) A method commonly used in image classification and recognition to determine the similarity error between two images by combining the structural similarity function between images (SSIM).

[0082] (2) In regression analysis, the root mean square error (RMSE) is often used to judge the degree of deviation between the predicted data and the actual data; therefore, by reasonably adjusting the weight ratio distribution of the above-mentioned structural similarity function and the root mean square error, it is possible to simultaneously take into account the spatial position distribution and absolute numerical error in the solar radiation prediction process.

[0083] Based on this idea, the new custom loss function of the ConvLSTM network model described in this embodiment is as follows:

[0084] k×SSIM+RMSE

[0085] Among them, the K value is a weight adjustment factor, which is used to adjust the weight distribution between the structural similarity function SSIM and the root mean square error RMSE between images to achieve the optimal extreme value.

[0086] Specifically, the formula of the structural similarity function SSIM is as follows: SSIM is written as S(x, y), which is numerically decomposed into a linear function combination of three components l, c, and s:

[0087] S(x,y)=f(l(x,y),c(x,,y),,s(x,y))

[0088]

[0089] SSIM(x,y)=[l(x,y)] a [c(x,y)] β [s(x,y)] γ

[0090] Where x and y represent the predicted solar radiation image and the actual image, respectively; σ is the image variance; μ is the image mean; C1, C2, and C3 are stabilization functions used to balance the dynamic range of pixel values; α, β, and γ are adjustment coefficients, which usually default to 1; the range of the structural similarity function calculated by the above formula is (-1, 1).

[0091] The calculation formula of the root mean square error RMSE is:

[0092]

[0093] Among them, x i is the predicted data, y i is the real data, and N is the total amount of data.

[0094] In order to avoid overfitting of the neural network and enhance the generalization and robustness of the network, this application adopts the means of simultaneously adding a neuron random dropout mechanism (Dropout) and data random batch normalization (BN), supplemented by data shuffling operations on the training data input into the neural network, which effectively increases the robustness, generalization and portability of the neural network.

[0095] For a more detailed description of the data input to the neural network model in this application, please refer to Table 1.

[0096] Table 1 Detailed description of neural network input data

[0097]

[0098]

[0099] The method described in this embodiment also includes: before step S3, using part of the data set in step S2 to train the ConvLSTM network model to obtain a ConvLSTM network model with prediction accuracy that meets the requirements; before training the ConvLSTM network model, it also includes: setting relevant model parameters of the ConvLSTM neural network; the relevant model parameters include at least: the number of neural network iterative calculations, the batch input network data size, the activation function setting, the loss function setting and the number of neural network layers.

[0100] In this embodiment, when predicting instantaneous solar shortwave radiation prediction data, the time series remote sensing data in step S1 includes three types of original data: TOA reflectivity data of 6 optical bands, solar zenith angle data of 1 band, and solar shortwave radiation data; the data preprocessing in step S2 includes two parts.

[0101] The first part is data quality control, which includes checking the quality of the obtained data, removing outliers in the data, and separating the TOA reflectivity data and solar zenith angle data from the solar shortwave radiation data for future use.

[0102] The second part is to transform the time series remote sensing data into a dimension that matches the ConvLSTM neural network model, that is, to reconstruct the data dimension into a five-dimensional array tensor of data number, time step, row, column, and eigenvalue.

[0103] The method in this embodiment further includes: normalizing the reconstructed data that has been preprocessed in step S2.

[0104] In this embodiment, when predicting multi-time-scale solar shortwave radiation, the time series remote sensing data in step S1 includes three types of original data: TOA reflectivity data of 6 optical bands, solar zenith angle data of 1 band, and solar shortwave radiation data; the data preprocessing in step S2 includes three parts.

[0105] The first part is data quality control, which includes checking the quality of the obtained data, removing outliers in the data, and separating the TOA reflectivity data and solar zenith angle data from the solar shortwave radiation data for future use.

[0106] The second part is to perform multi-time scale averaging processing on the instantaneous satellite remote sensing data.

[0107] The third part is to transform the time series remote sensing data into a dimension that matches the ConvLSTM neural network model, that is, to reconstruct the data dimension into a five-dimensional array tensor of data number, time step, row, column, and eigenvalue.

[0108] The method in this embodiment further includes: normalizing the reconstructed data that has been preprocessed in step S2.

[0109] Example 2

[0110] In this embodiment, there are two parallel routes, namely:

[0111] 1. Instantaneous solar shortwave radiation prediction supported by deep learning technology;

[0112] 2. Multi-timescale solar shortwave radiation prediction supported by deep learning technology;

[0113] like Figure 1 As shown: Technical route 1 in this embodiment, that is, instantaneous solar shortwave radiation prediction supported by deep learning technology, the technical process is as follows:

[0114] Taking the Sunflower-8 multispectral optical remote sensing satellite as an example, we first acquire continuous time series multispectral optical satellite remote sensing data. The acquired time series remote sensing data includes three types of raw data: TOA (Top of Atmosphere) reflectance data of 6 optical bands, SOZ (Solar of Zenith) data of 1 band, and solar shortwave radiation data.

[0115] The acquired remote sensing data is preprocessed. The data preprocessing includes two parts: 1) data quality control, checking the quality of the obtained data, removing outliers in the data, and separating the TOA reflectivity data and solar zenith angle data from the solar shortwave radiation data for backup; 2) transforming the data into a dimension acceptable to the ConvLSTM neural network, that is, reconstructing the data dimension into a five-dimensional array tensor of (number of data, time steps, rows, columns, eigenvalues) (n, time_steps, cows, cols, features).

[0116] Note that the time step selected during the data reconstruction process directly determines the length of time the neural network can predict. For example, if the time step is set to 4, the solar radiation data, solar zenith angle, and TOA reflectivity data for the previous 40 minutes will be used to predict the solar radiation data for the next 10 minutes.

[0117] Normalize the reconstructed data set that has been preprocessed in step (2). Since the range of TOA reflectivity data is already (0, 1), it is only necessary to transform the range of solar shortwave radiation shortwave data and solar zenith angle data to (0, 1).

[0118] The normalization method chosen in this application is to directly divide the solar shortwave radiation data by the maximum solar radiation value of 1367w / m 2 The solar zenith angle data is directly divided by the maximum solar zenith angle of 90 degrees. Finally, the processed data is randomly divided into a test set and a training set at a ratio of 20% (i.e., 80% of the data is the training set and 20% of the data is the test set). The data is shuffled while being divided.

[0119] Set the relevant model parameters of the ConvLSTM neural network, including the number of neural network iterations, the batch input network data size, the activation function setting, the loss function setting, and the number of neural network layers. For specific neural network parameter settings, see Table 2.

[0120] The normalized and divided data set in (3) is input into the ConvLSTM neural network with set parameters, and the model training is started. During the training process of the neural network, the Loss and Accuracy values are continuously recorded and evaluated. Finally, when the accuracy is within an acceptable range, the normalized solar shortwave radiation data in the interval [0, 1] based on the predicted data value can be obtained.

[0121] Furthermore, the normalized solar shortwave radiation data is multiplied by the maximum solar radiation value of 1367w / m 2 By performing the inverse normalization operation, we can obtain the real predicted solar radiation data in nature.

[0122] like Figure 2 As shown: Technical route 2 in this embodiment, that is, multi-time scale solar shortwave radiation prediction supported by deep learning technology, the technical process is as follows:

[0123] Taking the Sunflower-8 multispectral optical remote sensing satellite data as an example, we first acquire continuous time series multispectral optical satellite remote sensing data. The acquired time series remote sensing data includes three types of raw data: TOA (Top of Atmosphere) reflectance data in six optical bands, SOZ (Solar of Zenith) data in one band, and solar shortwave radiation data.

[0124] The acquired data is preprocessed. The data preprocessing mainly includes three parts: 1) data quality control, checking the quality of the obtained data, removing outliers in the data, and separating the TOA reflectivity data and solar zenith angle data from the solar shortwave radiation data for backup; 2) the instantaneous satellite remote sensing data obtained is averaged at multiple time scales. Here, taking the half-hour mean prediction as an example, the instantaneous satellite remote sensing data is averaged at a time scale of every 30 minutes. After processing, three 10-minute instantaneous data are converted into one mean data; 3) the data is transformed into a dimension acceptable to the ConvLSTM neural network, that is, the data dimension is reconstructed into a five-dimensional array tensor of (number of data, time steps, rows, columns, eigenvalues) (n, time_steps, cows, cols, features).

[0125] Note that the time step selected during the data reconstruction process here will directly determine the length of time that the neural network can predict. For example, the time step here is calculated using the hourly mean as the time step unit, and each time step is the statistical hourly mean.

[0126] Normalize the reconstructed data set that has been preprocessed in step (2). Since the range of TOA reflectivity data is already (0, 1), it is only necessary to transform the range of solar shortwave radiation shortwave data and solar zenith angle data to (0, 1).

[0127] The normalization method chosen in this application is to directly divide the solar shortwave radiation data by the maximum solar radiation value of 1367w / m 2 ; The solar zenith angle data is directly divided by the maximum solar zenith angle of 90°; finally, the processed data is randomly divided into a test set and a training set at a ratio of 20% (that is, 80% of the data is the training set and 20% of the data is the test set), and the data is shuffled while dividing the data set.

[0128] Set the relevant model parameters of the ConvLSTM neural network, including the number of neural network iterations, batch input network data size, activation function settings, loss function settings, and neural network layer number settings. For specific neural network parameter settings, see Table 2.

[0129] The normalized and divided data set in (3) is input into the ConvLSTM neural network with set parameters, and the model training is started. During the training process of the neural network, the Loss and Accuracy values are continuously recorded and evaluated. Finally, when the accuracy is within an acceptable range, the normalized solar shortwave radiation data in the interval [0, 1] based on the predicted data value can be obtained.

[0130] Furthermore, the normalized solar shortwave radiation data is multiplied by the maximum solar radiation value of 1367w / m 2 By performing the inverse normalization operation, we can obtain the multi-time scale mean solar shortwave radiation data starting from the current instant.

[0131] Table 2 Detailed description of ConvLSTM network parameter settings

[0132]

[0133]

[0134] The technical principles of the present invention have been described above in conjunction with specific embodiments. These descriptions are intended solely to illustrate the principles of the present invention and are not to be construed in any way as limiting the scope of protection of the present invention. Based on the explanations herein, those skilled in the art will be able to devise other specific embodiments of the present invention without inventive effort, and such embodiments will fall within the scope of protection of the present invention.

Claims

1. A deep learning-supported all-weather solar radiation remote sensing forecasting method, characterized in that: The following steps are involved: S1. Acquire time series multispectral remote sensing data; The time series remote sensing data in step S1 includes three types of original data: TOA reflectivity data of 6 optical bands, solar zenith angle data of 1 band, and solar shortwave radiation data; S2. performing data preprocessing on the acquired time series multispectral remote sensing data to obtain a preprocessed time series multispectral remote sensing dataset; When predicting instantaneous solar shortwave radiation forecast data, The data preprocessing in step S2 includes two parts: The first part is data quality control, which includes checking the quality of the obtained data, removing outliers in the data, and separating the TOA reflectivity data and solar zenith angle data from the solar shortwave radiation data for future use. The second part is to transform the time series remote sensing data into a dimension that matches the ConvLSTM neural network model, that is, to reconstruct the data dimension into a five-dimensional array tensor of data number, time step, row, column, and eigenvalue; When predicting solar shortwave radiation on multiple time scales, The data preprocessing in step S2 includes three parts: The first part is data quality control, which includes checking the quality of the obtained data, removing outliers in the data, and separating the TOA reflectivity data and solar zenith angle data from the solar shortwave radiation data for future use. The second part is to perform multi-time scale averaging processing on the instantaneous satellite remote sensing data; The third part is to transform the time series remote sensing data into a dimension that matches the ConvLSTM neural network model, that is, to reconstruct the data dimension into a five-dimensional array tensor of data number, time step, row, column, and eigenvalue; Normalizing the reconstructed data preprocessed in step S2; S3. Input the obtained pre-processed time series multispectral remote sensing data set into the ConvLSTM network model to obtain instantaneous solar shortwave radiation prediction data or multi-time-scale solar shortwave radiation prediction data, multiply the obtained instantaneous solar shortwave radiation prediction data or multi-time-scale solar shortwave radiation prediction data by the maximum value of solar radiation, and perform an inverse normalization operation to obtain the real instantaneous solar shortwave radiation prediction data or multi-time-scale mean solar shortwave radiation data in nature with the current instantaneous moment as the starting point; The ConvLSTM network model is a deep learning network model for remote sensing images and combined with a spatial attention mechanism; The basic model formula of the ConvLSTM network model is as follows: x=SA(x t ),H t-1 =SA(H t-1 ) i t =σ(W xi *X t +W hi *H t-1 +b i ) f t =σ(W xf *X t +W hf *H t-1 +b f ) g t =tanh(W xc *X t +W hc *H t-1 +b c ) o t =σ(W xo *X t +W ho *H t-1 +b o ) Among them, SA is the spatial attention module, which is used to participate in the distribution of weights between images. and It is a module aggregated by the attention mechanism, that is, the sum of the weighted values of the selectively acquired features at all positions at each time step; W i is the weight matrix corresponding to each feature; H t is the feature of the current input time step; i t is the input gate, which corresponds to the eigenvalue g after fusion t ; By outputting the updated original memory feature o t , and then the weights of the attention mechanism are redistributed so that the features of each time step are updated; * means convolution operation; o means corresponding multiplication, and the result of multiplication is Hadamard product; b i is the bias of the corresponding item; σ is the activation function corresponding to the neural network, which is the ReLU function in the model; C t The middle layer is the memory of the previous moment, that is, the memory layer; Each network layer is connected through convolution operations.

2. The forecasting method according to claim 1, characterized in that: The custom loss function of the ConvLSTM network model is as follows: k×SSIM+RMSE Among them, the K value is a weight adjustment factor, which is used to adjust the weight distribution between the structural similarity function SSIM and the root mean square error RMSE between images to achieve the optimal extreme value.

3. The forecasting method according to claim 2, characterized in that: The formula of the structural similarity function SSIM is as follows: SSIM is written as S(x,y), which is numerically decomposed into a linear function combination of three components: l, c, and s: S(x,y)=f(l(x,y),c(x,y),s(x,y)) SSIM(x,y)=[l(x,y)] α ·[c(x,y)] β ·[s(x,y)] γ Where x and y represent the predicted solar radiation image and the actual image, respectively; σ is the image variance; μ is the image mean; C1, C2, and C3 are stabilization functions used to balance the dynamic range of pixel values; α, β, and γ are adjustment coefficients, which usually default to 1; and the range of the structural similarity function is (-1, 1).

4. The forecasting method according to claim 2, characterized in that: The calculation formula of the root mean square error RMSE is: Among them, x i is the predicted data, y i is the real data, and N is the total amount of data.

5. The forecasting method according to claim 1, characterized in that: The method further includes: before step S3, using part of the data set in step S2 to train the ConvLSTM network model to obtain a ConvLSTM network model with a prediction accuracy that meets the requirements; Before training the ConvLSTM network model, the method further includes: setting relevant model parameters of the ConvLSTM neural network; The relevant model parameters include at least: the number of neural network iterative calculations, the batch input network data size, the activation function setting, the loss function setting and the number of neural network layers.

Images