A ground-based cloud image repair method for solar radiation prediction
By using compressed excitation module with momentum characteristics and phased design in solar radiation prediction, combined with partial convolution and supervision attention module, the problem of foundation cloud map repair under the influence of the light shield is solved, the prediction accuracy and repair quality are improved, and the prediction accuracy and repair quality are reflected. Meteorological and astronomical information.
Patent Information
- Application Number
- CN202211560453.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-12-07
- Publication Date
- 2024-12-10
- Estimated Expiration
- 2042-12-07
AI Technical Summary
The existing solar radiation prediction algorithm based on foundation cloud maps is affected by the light blocking panels and it is difficult to extract sufficient information, resulting in low short-term and ultra-short-term prediction accuracy, and common image repair methods are difficult to reflect meteorological and astronomical information, increasing the calculation amount.
Using a compressed excitation module with momentum characteristics, combined with a phased forward design, weaving time series and image feature data are fused, and images are repaired through partial convolutional sequences and supervisory attention modules to form a combination of visual integrity and physical information. Fix the result.
It improves the accuracy of solar radiation prediction and the quality of repair images, reduces the amount of calculation, enhances the time series information extraction ability, and can effectively reflect meteorological and astronomical characteristics.
Smart Images
Figure CN115760642B_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of solar radiation prediction and deep learning technology, and specifically relates to a ground-based cloud image repair method for solar radiation prediction. Background Art
[0002] Photovoltaic power generation is one of the most important components of utilizing solar energy. It has attracted much attention due to its large total amount of solar energy, pollution-free, and wide distribution. However, solar radiation is random and unstable, which brings difficulties to power dispatching. Therefore, it is of great significance to predict solar radiation. However, the current prediction algorithms based on ground-based cloud images with high accuracy and strong robustness are usually affected by the sunshade in the ground-based cloud image shooting equipment. The images lose a lot of information near the sun, making it difficult for short-term and ultra-short-term solar radiation prediction neural networks to extract enough information from them, which has a great impact on the prediction algorithm. Common image restoration methods and image restoration indicators mainly repair images based on human vision, and it is difficult to reflect physical information such as meteorology and astronomy that affect solar radiation in the restoration results of ground-based cloud images. Therefore, the restoration of images has little improvement on the accuracy of solar radiation prediction, but has greatly increased the amount of calculation for solar radiation prediction. Summary of the invention
[0003] The technical problem solved by the present invention is: to provide a compression excitation module with momentum characteristics, combined with a phased forward design, to complete the fusion of meteorological time series and image feature data, so that the repaired image has both the visual integrity of the traditional image repair neural network and the characteristic advantages of physical information such as meteorology and astronomy, and is a ground-based cloud image repair method for solar radiation prediction.
[0004] Technical solution: In order to solve the above technical problems, the technical solution adopted by the present invention is as follows:
[0005] A ground-based cloud image repair method for solar radiation prediction comprises the following steps:
[0006] S1: Preprocess the ground-based cloud image by positioning the sun position and covering the mask, and apply the equipment field of view mask;
[0007] S2: forming a sequence based on time by processing the pre-processed ground-based cloud image and the meteorological and astronomical measurement data at the same time;
[0008] S3: Use the partial convolution sequence to repair the image blocked by the shading plate: input the ground-based cloud image sequence into the partial convolution module to obtain the ground-based cloud image time series feature data;
[0009] S4: Then output it and pass it through the supervised attention module to obtain the repaired image and features.
[0010] S5: A compression excitation module with momentum characteristics is used to further extract historical information of features, and it is used as the deviation value to correct the solar radiation time series to obtain the predicted sequence; during training, the weights of the entire neural network are updated by combining the repaired foundation cloud map and the predicted sequence, so that the repaired foundation cloud map conforms to the intuitive feeling and contains physical information.
[0011] Furthermore, in step S1, the ground cloud image taken during equipment maintenance is rotated according to the solar azimuth to fix the sun position, and a mask is added to simulate a sunshade to form a training data set. s The calculation formula is as follows,
[0012] H s =sin -1 (sinφ×sinδ+cosφ×cosδ×cost)
[0013]
[0014] In the formula, H s is the solar altitude angle; φ is the geographical latitude; δ is the solar declination; t is the hour angle.
[0015] Furthermore, the device field of view mask is a binary image with the same size as a single channel of the ground-based cloud image. The formula for applying the device field of view mask is as follows:
[0016]
[0017] I masked =I in ⊙M ds
[0018]
[0019] Where M is the binary device field of view mask; h and w are the height and width of the image respectively; Represents a two-dimensional tensor space, whose dimensions are h and w respectively; To perform three tensor stacking operations on the new dimension, M ds M is a three-dimensional tensor formed by stacking M along the new axis; M ij is the value of the i-th row and j-th column in the device field of view mask; I masked It is the ground cloud image after applying the equipment field of view mask; I in is the original ground cloud map obtained by shooting; ⊙ is the Haddam product.
[0020] Further, in step S2, a time sequence of a fixed length is taken out according to a fixed step length s and a fixed sampling interval τ as the image sequence S to be processed. Ipending，i , the sampling formula is as follows:
[0021] S Ipending，i (j) = S Imasked (i×s+j×τ)
[0022] i, τ∈N, j∈{0, 1, 2,…,l-1}
[0023] s∈N +
[0024] In the formula, i is the sequence number; τ is the fixed sampling interval; S Ipending，i (j) is the jth element in the image sequence i to be processed; S Imasked (i×s+j×τ) is the (i×s+j×τ)th element in the ground-based cloud image time series data set after processing in step S1.
[0025] Furthermore, in step S3, the multi-head partial convolution module is implemented by repeatedly inputting a single image in the image sequence to be processed until the entire sequence is cycled, wherein the image cycle is parallelly calculated in a multi-head manner and then pooled for dimensionality reduction. For the Mth sequence, M∈{0, 1, 2, …, l-1}, the process is as follows:
[0026]
[0027] F Pconved =maxpool4(S Pconved )
[0028] In the formula, S Pconved It is the 4-dimensional tensor output after partial convolution calculation. It is a four-dimensional tensor space, and its dimensions are h, w, c, and l respectively; S Ipending，M (j) is the jth element in the Mth sequence to be processed; Pconv j is the operation corresponding to the jth partial convolutional network; The tensor stacking operation is performed on the new dimension with j as the variable; maxpool4 is the maximum pooling according to the fourth dimension; F Pconved is the output feature after dimensionality reduction, It is a three-dimensional tensor space, and the dimensions of each dimension are h, w, and c respectively.
[0029] Furthermore, in step S4, for the Mth sequence, the calculation formula of the supervised attention module is as follows:
[0030] I inpainted,M =conv a，none (F Pconved )+S Ipending,M (l-1)
[0031] F SAMed，M =conv b，none (FPconved )⊙conv c,sigmoid (I inpainted )+F Pconved
[0032] In the formula, I inpainted，M is the restored image of the last ground-based cloud image in the Mth sequence to be processed, It is a three-dimensional tensor space, and its dimensions are h, w, c respectively; conv a，none and conv b，none Convolution is performed on two independent images. c，sigmoid For conv a，none and conv b，none Independent convolution calculations are performed, and a sigmoid function is calculated on the calculation results; F SAMed，M Take the Mth sequence to be processed as input and the output features of the supervised attention module, It is a three-dimensional tensor space, and the dimensions of each dimension are f1, f2, and f3 respectively.
[0033] Further, in step S5, the compression excitation module with momentum characteristics has a cycle count value n, which is calculated as follows:
[0034]
[0035] F MoSEd =conv f，none [F Mo，0 ⊙sigmoid(F Mo，n )]
[0036] n∈N + , m∈N +
[0037] In the formula, F Mo，n is the momentum term of the sequence with count n; conv d，hswish 、conv e，hswish 、conv f，none Convolutional layers are independent of each other. d，hswish 、conv e，hswish The activation functions are hswish, conv f，none The activation function is a linear activation function; r is the momentum parameter, r∈[0,1].
[0038] Furthermore, in step S5, the corrected features are added to the meteorological and astronomical parameter time series corresponding to the time series of the ground-based cloud image according to the elements to obtain a corrected meteorological and astronomical parameter sequence, and the error between the corrected meteorological and astronomical parameter sequence and the meteorological and astronomical parameter sequence at the next moment, as well as the error between the repaired image and the real image are calculated. The two errors are combined as a loss function to train the network.
[0039] Furthermore, in step S5, the loss function is the sum of the losses of meteorological and astronomical data and image restoration, and the calculation formula is as follows:
[0040] L total =pL inpainting +(1-p)L pred
[0041] =6pL hole +pL valid +0.05pL perceptual +120p(L styleout +L stylecomp )+0.1pL tv +(1-p)L pred
[0042] Where p is the loss coefficient, which is 0.5; L pred is the MSE loss of meteorological and astronomical parameters; L inpainting To repair the error between the image and the real image; L hole is the shading plate mask loss; L valid is the non-shade mask loss; L perceptual is the perceptual loss of image restoration; (L styleout +L stylecomp ) is the style loss of image restoration; L tv is the smoothness loss of image inpainting.
[0043] Beneficial effects: Compared with the prior art, the present invention has the following advantages:
[0044] (1) The algorithm disclosed in the present invention obtains a ground-based cloud map by rotating the device according to the solar azimuth during normal operation, fixes the sunshade position of the device, and uses a trained neural network to form a ground-based cloud map that is not affected by the sunshade and is convenient for solar radiation prediction or subsequent meteorological analysis.
[0045] (2) The present invention designs a compression excitation module with momentum characteristics, which enhances the ability to extract time series information and converts the feature dimensions to form correction items for meteorological and astronomical data.
[0046] (3) The present invention adopts a multi-stage network, first generates a repaired image, and predicts meteorological and astronomical data based on information such as the features of the repaired image, so that the network used in training removes the prediction of the second stage and directly serves as an image repair network, so that the repaired image can reflect physical properties such as meteorology and astronomy.
[0047] (4) The present invention combines partial convolution with attention mechanism to reduce the amount of computation and improve training efficiency.
[0048] (5) The present invention uses image sequences as input of the image restoration neural network, providing sufficient historical information for the restoration network, which is beneficial to improving the restoration accuracy. BRIEF DESCRIPTION OF THE DRAWINGS
[0049] Figure 1 It is a schematic diagram of the overall structure of the network during the training process of the present invention;
[0050] Figure 2 It is a schematic diagram of the structure of a compression excitation module with momentum characteristics;
[0051] Figure 3 It is a cloud map of the foundation before repair, with the equipment mask superimposed and after repair. DETAILED DESCRIPTION
[0052] The present invention is further illustrated below in conjunction with specific examples. The examples are implemented based on the technical solutions of the present invention. It should be understood that these examples are only used to illustrate the present invention and are not used to limit the scope of the present invention.
[0053] The ground-based cloud image repair method for solar radiation prediction of the present invention first performs sun position positioning and mask preprocessing on the ground-based cloud image; the preprocessed ground-based cloud image and the meteorological and astronomical measurement data at the same time form a sequence based on time; and the solar radiation is predicted through a multi-head partial convolutional network, a supervised attention module, and a compression excitation module with momentum characteristics. It mainly includes steps S1 to S5, and the specific contents are as follows:
[0054] S1: Preprocess the ground-based cloud image for sun position positioning and overlay mask, and apply the device field of view mask.
[0055] The ground cloud image taken during equipment maintenance is rotated according to the solar azimuth to fix the sun position, and a mask is added to it to simulate the sunshade to form a training data set. s The calculation formula is as follows,
[0056] H s =sin -1 (sinφ×sinδ+cosφ×cosδ×cost)
[0057]
[0058] In the formula, H s is the solar altitude angle; φ is the geographical latitude; δ is the solar declination; t is the hour angle.
[0059] The device field of view mask is a binary image with the same size as a single channel of the ground-based cloud image. The formula for applying the device field of view mask is as follows:
[0060]
[0061] I masked =I in ⊙M ds
[0062]
[0063] Where M is the binary device field of view mask; h and w are the height and width of the image respectively; is a two-dimensional tensor space, whose first two dimensions are h and w; To perform three tensor stacking operations on the new dimension, M ds M is a three-dimensional tensor formed by stacking M along the new axis; M ij is the value of the i-th row and j-th column in the device field of view mask; I masked It is the ground cloud image after applying the equipment field of view mask; I in is the original ground cloud map obtained by shooting; ⊙ is the Haddam product.
[0064] S2: The pre-processed ground-based cloud images and the meteorological and astronomical measurement data at the same time are sequenced based on time.
[0065] According to the fixed step size s and fixed sampling interval τ, a time series of fixed length is taken as the image sequence S to be processed. Ipending，i , the sampling formula is as follows:
[0066] S Ipending，i (j) = S Imasked (i×s+j×τ)
[0067] i, τ∈N, j∈{0, 1, 2,…,l-1}
[0068] s∈N +
[0069] In the formula, i is the sequence number; τ is the fixed sampling interval; S Ipending，i (j) is the jth element in the image sequence i to be processed; S Imasked (i×s+j×τ) is the (i×s+j×τ)th element in the ground-based cloud image time series data set after processing in step S1.
[0070] S3: Inpainting images occluded by occluders using partial convolution sequences.
[0071] The ground-based cloud image sequence is input into the partial convolution module to obtain the ground-based cloud image time series feature data; the multi-head partial convolution module is implemented by repeatedly inputting a single image in the image sequence to be processed until the entire sequence is cycled, where the image cycle is parallelly calculated in a multi-head manner and then pooled for dimensionality reduction. For the Mth sequence, M∈{0, 1, 2, …, l-1}, the process is as follows:
[0072]
[0073] F Pconved =maxpool4(S Pconved )
[0074] In the formula, S Pconved It is the 4-dimensional tensor output after partial convolution calculation. It is a four-dimensional tensor space, and its dimensions are h, w, c, and l respectively; S Ipending，M (j) is the jth element in the Mth sequence to be processed; Pconv j is the operation corresponding to the jth partial convolutional network; The tensor stacking operation is performed on the new dimension with j as the variable; maxpool4 is the maximum pooling according to the fourth dimension; F Pconved is the output feature after dimensionality reduction, It is a three-dimensional tensor space, and the dimensions of each dimension are h, w, and c respectively.
[0075] S4: The output is then passed through the supervised attention module to obtain the repaired image and features.
[0076] For the Mth sequence, the calculation formula of the supervised attention module is as follows:
[0077] I inpainted,M =conv a，none (F Pconved )+S Ipending,M (l-1)
[0078] F SAMed,M =conv b,none (F Pconved )⊙conv c，sigmoid (I inpainted )+F Pconved
[0079] In the formula, I inpainted，M is the restored image of the last ground-based cloud image in the Mth sequence to be processed, It is a three-dimensional tensor space, and its dimensions are h, w, c respectively; conv a，noneand conv b，none Convolution is performed on two independent images. c，sigmoid For conv a，none and conv b,none Independent convolution calculations are performed, and a sigmoid function is calculated on the calculation results; F SAMed，M Take the Mth sequence to be processed as input and the output features of the supervised attention module, It is a three-dimensional tensor space, and the dimensions of each dimension are f1, f2, and f3 respectively.
[0080] S5: Using a compressed excitation module with momentum characteristics to further extract feature history
[0081] The information is used as the deviation value to correct the solar radiation time series to obtain the predicted sequence. During the training, the weight of the entire neural network is updated by combining the repaired foundation cloud map and the predicted sequence, so that the repaired foundation cloud map is consistent with the intuitive feeling and contains physical information.
[0082] The compression excitation module with momentum characteristics has a cycle count value n, which is calculated as follows:
[0083]
[0084] F MoSEd =conv f，none [F Mo，0 ⊙sigmoid(F Mo，n )]
[0085] n∈N + , m∈N +
[0086] In the formula, F Mo，n is the momentum term of the sequence with count n; conv d，hswish 、conv e，hswish 、conv f，none Convolutional layers are independent of each other. d，hswish 、conv e，hswish The activation functions are hswish, conv f，none The activation function is a linear activation function; r is the momentum parameter, r∈[0,1].
[0087] Each time the compression excitation module with momentum characteristics is calculated, the values of n and m are updated. The value of n is selected in sequence from the sequence {0, 1, 2, 3, 4, 5, 6, 7}, and the value of m is increased to obtain the modified characteristics.
[0088] Add the corrected features to the meteorological and astronomical parameter time series corresponding to the time series of the ground-based cloud image according to the elements to obtain the corrected meteorological and astronomical parameter sequence, calculate the error between the corrected meteorological and astronomical parameter sequence and the meteorological and astronomical parameter sequence at the next moment, and the error L between the repaired image and the real image inpainting , combining the two errors as the loss function to train the network, L total is the loss function of the neural network in training. The loss function is the sum of the losses of meteorological and astronomical data and image restoration. The calculation formula is as follows:
[0089] L total =pL inpainting +(1-p)L pred
[0090] =6pL hole +pL valid +0.05pL perceptual +120p(L styleout +L stylecomp )+0.1pL tv +(1-p)L pred
[0091] Where p is the loss coefficient, which is 0.5; L pred is the MSE loss of meteorological and astronomical parameters; L inpainting To repair the error between the image and the real image; L hole is the shading plate mask loss; L valid is the non-shade mask loss; L perceptual is the perceptual loss of image restoration; (L styleout +L stylecomp ) is the style loss of image restoration; L tv is the smoothness loss of image inpainting.
[0092] In the specific implementation of the present invention, network training is first performed and then image restoration is performed. The training process is a multi-stage neural network, and the neural network structure of the restoration process is a stage of the neural network in the training process.
[0093] The network training process specifically includes steps S11 to S19:
[0094] S11: obtaining a time series data set of ground-based cloud images without a sunshade taken at fixed time intervals and meteorological and astronomical data at corresponding times, and preprocessing the time series data set of ground-based cloud images and the meteorological and astronomical data at corresponding times;
[0095] In this embodiment, the ground-based cloud image is collected from the TSI-880 automatic all-sky imager of the American YANKEE company, and the collection time interval is 10 minutes.
[0096] Obtain a time series dataset of ground-based cloud images without sunshades taken at 10-minute intervals and the corresponding meteorological and astronomical data, and rotate the ground-based cloud images clockwise. This positions the sun roughly at the bottom center of the image.
[0097] S12: taking out a time sequence of fixed length as an image sequence to be processed according to a fixed step size and a fixed sampling interval;
[0098] In this embodiment, the device field of view mask is applied to the ground-based cloud image data set processed in step S11, only the circular area of the sky part in the image is retained, and a 160-minute time series is taken out with a step length of 10 minutes and an interval of 10 minutes as the image sequence to be processed.
[0099] S13: taking out an image in the sequence of images to be processed, inputting it together with the device shading plate mask into the partial convolution module for mask shrinkage, and setting the next image in the sequence of images to be processed as the image to be processed;
[0100] In this embodiment, the image sequence to be processed and the rectangular occluder mask are used as inputs of the partial convolution module to obtain features after mask shrinkage.
[0101] S14: repeating step S13 until all images in the image time sequence complete mask shrinkage, and merging the mask-shrinked image sequence to form feature data corresponding to the image sequence to be processed;
[0102] S15: Input the feature data in step S14 and the last ground cloud image in the image sequence to be processed in step S12 into the supervised attention module to generate a repaired image and attention features;
[0103] S16: input the attention feature in step S15 into the compression excitation module with momentum characteristics and update the state therein, and obtain the correction features of meteorological and astronomical parameters according to the current feature and the existing state in the compression excitation module with momentum characteristics;
[0104] The momentum parameter r is set to 0.75. The values of n and m are updated each time the compression excitation module with momentum characteristics is calculated. The value of n is selected in sequence from the sequence {0, 1, 2, 3, 4, 5, 6, 7}, and the value of m is increased to obtain the modified characteristics.
[0105] S17: applying the correction feature to the meteorological and astronomical parameter time series corresponding to the time of the ground-based cloud image time series to obtain a corrected meteorological and astronomical parameter series;
[0106] S18: Calculate the error between the corrected meteorological and astronomical parameter sequence and the meteorological and astronomical parameter sequence at the next moment, and the error between the restored image in step S15 and the real image;
[0107] S19: Combine the two errors in step S18 as the loss function to train the network.
[0108] The image restoration process specifically includes the following steps S21 to S25:
[0109] S21: Acquire and record the ground-based cloud image at the current moment, calculate the current solar azimuth according to the shooting position and location of the ground-based cloud image, rotate the ground-based cloud image according to the azimuth, and apply the device field of view mask;
[0110] S22: Repeat step S21 at the time interval in the network training step S12 until the length of the recorded ground-based cloud image time series reaches the length of the time series in the network training step S12;
[0111] S23: inputting the ground-based cloud image sequence into the partial convolution module to obtain the ground-based cloud image time series feature data;
[0112] S24: Input the foundation cloud image time series feature data and the foundation cloud image at the current moment into the supervised attention module to generate a repaired foundation cloud image;
[0113] If the ground-based cloud image repair needs to be performed in real time, step S21 is repeated according to the time interval in the network training step S12, and the earliest ground-based cloud image data is discarded to form a new ground-based cloud image time series and repeat steps S23 to S25.
[0114] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.
Claims
1. A ground-based cloud image repair method for solar radiation prediction, characterized in that: The following steps are involved: S1: Preprocess the ground-based cloud image by positioning the sun position and covering the mask, and apply the equipment field of view mask; rotate the ground-based cloud image taken during equipment maintenance according to the sun azimuth to fix the sun position, add a mask to it to simulate the sunshade, and form a training data set; S2: forming a sequence based on time by processing the pre-processed ground-based cloud image and the meteorological and astronomical measurement data at the same time; S3: Use the partial convolution sequence to repair the image blocked by the shading plate: input the ground-based cloud image sequence into the partial convolution module to obtain the ground-based cloud image time series feature data; S4: The output is then passed through the supervised attention module to obtain the restored image and features; S5: A compression excitation module with momentum characteristics is used to further extract historical information of features, and it is used as a deviation value to correct the solar radiation time series to obtain a predicted sequence; during training, the weight of the entire neural network is updated by combining the repaired foundation cloud map and the predicted sequence, so that the repaired foundation cloud map conforms to the intuitive feeling and contains physical information; The compression excitation module with momentum characteristics has a cycle count value n, which is calculated as follows: F MoSEd =conv f,none [F Mo，0 ⊙sigmoid(F Mo，n )] n∈N + ,m∈N + In the formula, F Mo,n is the momentum term of the sequence with count n; conv d,hswish 、conv e,hswish 、conv f,none Convolutional layers are independent of each other. d,hswish 、conv e,hswish The activation functions are hswish, conv f,none The activation function is a linear activation function; r is the momentum parameter, r∈[0,1]; F SAMed, M is the Mth sequence to be processed as input, and the output features of the supervised attention module, It is a three-dimensional tensor space, and the dimensions of each dimension are f1, f2, and f3 respectively.
2. The ground-based cloud image repair method for solar radiation prediction according to claim 1, characterized in that: Solar azimuth A s The calculation formula is as follows, H s =that -1 (sinφ×sinδ+cosφ×cosδ×cost) In the formula, H s is the solar altitude angle; φ is the geographical latitude; δ is the solar declination; t is the hour angle.
3. The ground-based cloud image repair method for solar radiation prediction according to claim 1, characterized in that: In step S1, the mask simulation shading plate applies the device field of view mask, which is a binary image with the same size as a single channel of the ground-based cloud map. The formula for applying the device field of view mask is as follows: I masked =I in ☉M ds M ij ∈{0,1} Where M is the binary device field of view mask; h and w are the height and width of the image respectively; is a two-dimensional tensor space, whose first two dimensions are h and w; To perform three tensor stacking operations on the new dimension, M ds M is a three-dimensional tensor formed by stacking M along the new axis; M ij is the value of the i-th row and j-th column in the device field of view mask; I masked It is the ground cloud image after applying the equipment field of view mask; I in is the original ground cloud map obtained by shooting; ⊙ is the Haddam product.
4. The ground-based cloud image repair method for solar radiation prediction according to claim 1, characterized in that: In step S2, a time sequence of fixed length l is taken out as the image sequence S to be processed according to a fixed step length s and a fixed sampling interval τ. Ipending,i , the sampling formula is as follows: S Ipending，i (j)=S Imasked (i×s+j×τ) i,τ∈N,j∈{0,1,2,…,l-1} s∈N + In the formula, i is the sequence number; τ is the fixed sampling interval; S Ipending,i (j) is the jth element in the image sequence i to be processed; S Imasked (i×s+j×τ) is the (i×s+j×τ)th element in the ground-based cloud image time series data set after processing in step S1.
5. The ground-based cloud image repair method for solar radiation prediction according to claim 1, characterized in that: In step S3, the multi-head partial convolution module is implemented by repeatedly inputting a single image in the image sequence to be processed until the entire sequence is cycled, where the image cycle is parallelly calculated in a multi-head manner and then pooled for dimensionality reduction. For the Mth sequence, M∈{0,1,2,…,l-1}, the process is as follows: F Pconved =maxpool4(S Pconved ) In the formula, S Pconved It is the 4-dimensional tensor output after partial convolution calculation. It is a four-dimensional tensor space, and its dimensions are h, w, c, and l respectively; S Ipending,M (j) is the jth element in the Mth sequence to be processed; Pconv j is the operation corresponding to the jth partial convolutional network; is to perform l tensor stacking operations on the new dimension with j as the variable; maxpool4 is to perform maximum pooling according to the fourth dimension; F Pconved is the output feature after dimensionality reduction, It is a three-dimensional tensor space, and its dimensions are h, w, and c respectively; M ds Represents a 3D tensor formed by stacking M along the new axis.
6. The ground-based cloud image repair method for solar radiation prediction according to claim 5, characterized in that: In step S4, for the Mth sequence, the calculation formula of the supervised attention module is as follows: I inpainted,M =conv a，none (F Pconved )+S Ipending,M (l-1) F SAMed，M =conv b,none (F Pconved )⊙conv c，sigmoid (I inpainted )+F Pconved In the formula, I inpainted,M is the restored image of the last ground-based cloud image in the Mth sequence to be processed, It is a three-dimensional tensor space, and its dimensions are h, w, c respectively; conv a,none and conv b,none Convolution is performed on two independent images. c,sigmoid For conv a，none and conv b,none Independent convolution calculations are performed, and a sigmoid function is calculated on the calculation results; F SAMed,M Take the Mth sequence to be processed as input and the output features of the supervised attention module, is a three-dimensional tensor space, and its dimensions are f1, f2, and f3 respectively; S Ipending，M Represented as the element in the Mth sequence to be processed.
7. The ground-based cloud image repair method for solar radiation prediction according to claim 1, characterized in that: In step S5, the corrected feature is added to the meteorological and astronomical parameter time series corresponding to the time series of the ground-based cloud image according to the element, to obtain a corrected meteorological and astronomical parameter sequence, and the error between the corrected meteorological and astronomical parameter sequence and the meteorological and astronomical parameter sequence at the next moment, as well as the error between the repaired image and the real image are calculated, and the two errors are combined as the loss function to train the network.
8. The ground-based cloud image repair method for solar radiation prediction according to claim 7, characterized in that: In step S5, the loss function is the sum of the losses of meteorological and astronomical data and image restoration, and the calculation formula is as follows: L total =pL inpainting +(1-p)L pred =6pL hole +pL valid +0.05pL percepptual +120p(L styleout +L stylecomp )+0.1pL tv +(1-p)L pred Where p is the loss coefficient, which is 0.5; L pred is the MSE loss of meteorological and astronomical parameters; L inpainting To repair the error between the image and the real image; L hole is the shading plate mask loss; L valid is the non-shade mask loss; L perceptual is the perceptual loss of image restoration; (L styleout +L stylecomp ) is the style loss of image restoration; L tv is the smoothness loss of image inpainting.