Photovoltaic power station ultra-short-term power prediction method based on double-foundation cloud atlas

Abstract

The invention relates to the technical field of photovoltaic power generation, in particular to a photovoltaic power station ultra-short-term power prediction method based on double foundation cloud atlas, which comprises the following steps: collecting real-time foundation cloud atlas of two imagers of a photovoltaic power station, and carrying out image preprocessing; dividing the preprocessed foundation cloud picture into different cloud cluster areas by using an image segmentation algorithm; according to a binocular vision principle, calculating cloud bottom height and horizontal distance of each cloud cluster relative to an observation point; and simulating the ground cloud shadow coverage condition of the photovoltaic power station, obtaining an attenuation coefficient, correcting the irradiation intensity by using the attenuation coefficient, and inputting into the power prediction model to obtain an ultra-short-term power prediction value. Through the prediction method, the shielding condition of the cloud layer projection on each area of the photovoltaic power station can be accurately judged, and the ultra-short-term power prediction precision and accuracy of the large-scale photovoltaic power station under the cloudy weather condition are improved.

Description

Technical Field

[0001] The invention relates to the technical field of photovoltaic power generation, and in particular to an ultra-short-term power prediction method for a photovoltaic power station based on a double-foundation cloud map. Background Art

[0002] Due to complex and changeable weather conditions, cloud cover and other factors, the intensity of solar radiation fluctuates, making photovoltaic power generation random and intermittent. The current mainstream optical power prediction method is mainly applicable to photovoltaic power stations with meteorological monitoring stations, and power prediction is based on real-time meteorological data and historical power generation data. This method can predict more accurately on sunny days; however, under complex weather conditions such as cloudy days, some areas of the photovoltaic power station are blocked by clouds, and the meteorological monitoring station cannot obtain this information, resulting in low power prediction accuracy in a short period of time, and unable to provide reference information for the filling of power station dispatch. If the predicted power is significantly different from the power grid dispatch power, it will lead to a large amount of abandoned light, thereby reducing power generation revenue and even affecting the safe and stable operation of the power grid.

[0003] In the prior art, a Chinese invention patent document with publication number CN103353952A and publication date August 10, 2016 is proposed to solve the above-mentioned technical problems. The technical solution disclosed in the patent document is as follows: a photovoltaic power prediction method based on ground-based cloud maps, comprising the following steps: Step 1: cloud movement prediction; Step 2: ground irradiance prediction; Step 3: establishing a comprehensive photoelectric conversion model; Step 4: photovoltaic power prediction.

[0004] In the prior art, a Chinese invention patent document with publication number CN112507793A and publication date August 8, 2023 has also been proposed to solve the above-mentioned technical problems. The technical solution disclosed in the patent document is as follows: an ultra-short-term photovoltaic power prediction method, which uses a longitude and latitude correction method to correct fisheye cloud images; uses an HSV-SURF algorithm to quickly extract feature points from a ground-based cloud image sequence; quickly matches and corrects feature point pairs, extracts and predicts cloud movement trajectories; uses an improved threshold segmentation method to refine cloud extraction, and proposes an irradiation coefficient that intuitively characterizes the irradiance conditions at the next moment; establishes an ultra-short-term photovoltaic power prediction model based on an improved IAM-CNN-LSTM hybrid neural network; initializes the weights of the hybrid neural network and sets the maximum number of iterations; establishes a convolutional neural network and inputs the feature matrix into the model; uses a control variable method to optimize the network structure; if the maximum number of iterations is reached, the iteration terminates and outputs the network parameters; uses the trained composite network to perform ultra-short-term power prediction to obtain the predicted power at the time point to be predicted.

[0005] The above technical solutions all adopt a single ground-based cloud map and only use cloud cover characteristics for cloud occlusion assessment. They are unable to accurately judge the occlusion of cloud projection on various areas of the photovoltaic power station. The precision and accuracy of ultra-short-term power prediction of large photovoltaic power stations under cloudy weather conditions are low. Summary of the invention

[0006] In order to solve the above technical problems, the present invention proposes an ultra-short-term power prediction method for photovoltaic power stations based on dual-base cloud images. The method uses image processing technology and binocular vision principles to accurately judge the shielding of cloud projection on various areas of the photovoltaic power station, thereby improving the precision and accuracy of ultra-short-term power prediction of large photovoltaic power stations under cloudy weather conditions, which is beneficial to the scientific dispatch of the power grid and maintains the stable operation of the new power system.

[0007] The present invention is achieved by adopting the following technical solutions:

[0008] A method for ultra-short-term power prediction of a photovoltaic power station based on a dual-foundation cloud map comprises the following steps:

[0009] Step S1. Collect real-time ground cloud images of two imagers of the photovoltaic power station and perform image preprocessing;

[0010] Step S2. Using an image segmentation algorithm, the preprocessed ground-based cloud image is divided into different cloud cluster areas;

[0011] Step S3. Calculate the cloud base height and horizontal distance of each cloud cluster relative to the observation point based on the binocular vision principle;

[0012] Step S4. Simulate the cloud coverage of the photovoltaic power station, obtain the attenuation coefficient, use the attenuation coefficient to correct the radiation intensity, and input the power prediction model to obtain the ultra-short-term power prediction value.

[0013] The image preprocessing in step S1 specifically includes the following steps:

[0014] Step S 11 .Remove the pixels in the non-sky area of ​​the ground-based cloud image and mark them as masks to be repaired;

[0015] Step S 12 .Perform image restoration on the mask to be restored to complete the missing sky information;

[0016] Step S 13 .The distortion correction coefficient of the all-sky imager is determined through parameter calibration, and the distortion of the ground-based cloud image is corrected.

[0017] The step S2 specifically refers to: the image segmentation algorithm introduces a regression model into the U-Net network, is trained based on a labeled ground-based cloud image dataset, and the preprocessed ground-based cloud image is used as the model input to obtain a pixel classification image of the cloud cluster, wherein the image contains the pixel coordinates of each single cloud cluster and the classification result representing the thickness of the cloud cluster; the total number of pixels belonging to the same single cloud cluster is used as the cloud amount feature of the single cloud cluster.

[0018] The labeled ground-based cloud image dataset uses an open source dataset as its basis, calculates the mean grayscale value of the individual cloud cluster area in the ground-based cloud image, clusters it according to a pre-set number of classifications, and uses the clustering results as cloud thickness classification labels.

[0019] The step S3 specifically comprises the following steps:

[0020] Step S 31 . Obtain the angle α between the line connecting the two imagers and the east direction and the horizontal distance L;

[0021] Step S 32 .Extract the edge contours of individual clouds from the two ground-based cloud images respectively, perform similarity comparison analysis in turn, and match the objects with the lowest similarity index that is less than the preset threshold as the same cloud cluster;

[0022] Step S 33 .Get the center point of the cloud i that has a successful similarity match Polar coordinate parameters in two ground-based cloud images According to the calibration parameters, the pole diameter Convert to elevation

[0023] Step S 34 . Calculate the horizontal distance between the center of cloud cluster i and imager k by the following method: and cloud base height H i ;

[0024]

[0025]

[0026] Where α is the angle between the line between the two imagers and the east direction; L is the horizontal distance between the two imagers; k is the imager and the ground-based cloud map number taken, which can be 0 or 1; and are the azimuth and elevation of the center point of cloud cluster i in the ground-based cloud image taken by imager k.

[0027] The step S 32The similarity analysis of the edge contour of the single cloud cluster specifically refers to: using shape matching, returning the similarity index, if the lowest similarity index is greater than the preset threshold, transforming the matching object, and then performing similarity analysis, repeating this step until the preset number of times is reached, if the lowest similarity index is still greater than the preset threshold, then ignoring the cloud cluster object.

[0028] The horizontal distance L between the two imagers is between 0.8 and 1.5 km.

[0029] The simulation of the cloud shadow coverage on the ground of the photovoltaic power station in step S4 specifically includes:

[0030] Calculate the solar azimuth θ s and the altitude angle z s :

[0031]

[0032] ω=(t-12)×15°

[0033]

[0034]

[0035] Calculate the ground projection of the center point of cloud cluster i in sequence Polar coordinate parameters of relative imager k

[0036]

[0037]

[0038]

[0039] Determine in turn whether the ground projections of all clouds fall on each area of ​​the photovoltaic power station;

[0040] In the above formula, n is the day of the year; t is the collection time; is the latitude of the observation point; are the horizontal distance and azimuth of the ground projection of the center point of cloud cluster i relative to imager k; H i are the horizontal distance of the center of cloud cluster i relative to imager k and the cloud base height; z s ,θ s are the solar altitude and azimuth respectively.

[0041] In step S4, using the attenuation coefficient to correct the irradiation intensity specifically refers to:

[0042]

[0043] Where R is the real-time irradiation intensity; Rj ' is the corrected irradiation intensity for the photovoltaic power station area or subarray j; μ j is the attenuation coefficient; μ1 and μ2 are proportional coefficients; N j is the number of cloud clusters falling in region j; and are the cloud amount and cloud thickness parameters of cloud cluster i respectively; τ is the radiation intensity influencing factor, which is positively correlated with R.

[0044] The power prediction model adopts the LSTM-Attention model combined with the attention mechanism, and takes the real-time temperature, wind speed, irradiation intensity of the meteorological monitoring station of the photovoltaic power station, the irradiation intensity corrected by the dual-base cloud map, the real-time power generation power, the historical meteorological and power generation operation data as the power prediction model input to obtain the ultra-short-term power prediction value.

[0045] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0046] 1. The present invention extracts the same cloud features from the dual-base cloud images and uses the binocular vision principle to more accurately calculate the cloud base height and horizontal distance, simulate and restore the real-time shielding situation of the photovoltaic power station in different regions, and reduce the interference of the rapid formation, movement or dissipation of clouds on the ultra-short-term power forecast under cloudy weather conditions, improve the prediction accuracy, facilitate the accurate reporting of power station scheduling, and increase the power generation income of the power station.

[0047] 2. The present invention is based on the decisive indicators that affect the ultra-short-term power prediction, including: cloud amount, cloud thickness and cloud base height calculated by dual-base cloud images, which are quantified into attenuation coefficients using a weighted algorithm, and the real-time irradiation intensity collected by the meteorological instrument, and the real-time irradiation intensity of the photovoltaic power station is corrected by region, thereby avoiding the limitation that the data of a single-point meteorological monitoring station in a large photovoltaic power station cannot distinguish the difference in irradiation intensity caused by local cloud cover in different areas of the power station, and improving the accuracy of ultra-short-term power generation prediction under complex weather conditions such as cloudy weather. In addition, in the present invention, by designing the attenuation coefficient and directly correcting the irradiation intensity, not only the accuracy of the prediction can be improved, but also the output power on a sunny day without cloud cover can be predicted at the same time, which is also impossible to achieve with the existing technology.

[0048] 3. The present invention uses image segmentation algorithm and similarity comparison analysis to extract corresponding feature points to calculate the cloud base height and horizontal distance of each cloud cluster relative to the observation point, which can achieve accurate assessment of the cloud base height and cloud size.

[0049] 4. When calculating the cloud base height and the horizontal distance, the present invention adopts polar coordinate calculation to avoid using conventional pixel coordinate difference for calculation, so that the calculation of the cloud base height and the horizontal distance is more accurate.

[0050] 5. The present invention uses the LSTM-Attention model combined with the attention mechanism to perform the final power generation prediction. The real-time temperature, wind speed, irradiation intensity corrected by the dual-base cloud map, real-time power generation, historical meteorological and power generation operation data of the photovoltaic power station meteorological monitoring station are used as multi-dimensional feature vectors to input the model. After being processed by the LSTM hidden layer, several intermediate states are obtained, and then processed by the Attention layer. Through the attention coefficient, important information is strengthened and unimportant information is suppressed. Finally, the results are output through the fully connected layer, which improves the performance and interpretability of the prediction model. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments, wherein:

[0052] Figure 1 It is a schematic diagram of the process of the present invention;

[0053] Figure 2 Schematic diagram of the positional relationship of a single cloud center point in two all-sky imagers in the present invention;

[0054] Figure 3 This is a schematic diagram of the network structure of the LSTM-Attention model combined with the attention mechanism in the present invention. DETAILED DESCRIPTION

[0055] Example 1

[0056] As a basic implementation mode of the present invention, the present invention includes a method for ultra-short-term power prediction of a photovoltaic power station based on a dual-foundation cloud map, comprising the following steps:

[0057] Step S1. Collect real-time ground cloud images of two imagers of the photovoltaic power station and perform image preprocessing.

[0058] Step S2: Using image segmentation algorithm, the pre-processed ground-based cloud image is divided into different cloud cluster areas.

[0059] Step S3. Calculate the cloud base height and horizontal distance of each cloud cluster relative to the observation point based on the binocular vision principle.

[0060] Step S4. Simulate the cloud coverage of the photovoltaic power station, obtain the attenuation coefficient, use the attenuation coefficient to correct the radiation intensity, and input the power prediction model to obtain the ultra-short-term power prediction value.

[0061] Example 2

[0062] As a preferred embodiment of the present invention, the present invention includes a method for ultra-short-term power prediction of a photovoltaic power station based on a dual-foundation cloud map, comprising the following steps:

[0063] Step S1. Collect real-time ground cloud images of two imagers of the photovoltaic power station and perform image preprocessing. The imager may be Cloud Sky Imager (CSI) of Zhongkeda Company.

[0064] Among them, image preprocessing includes the following steps.

[0065] Step S 11 .According to the obstruction of buildings, brackets, shading belts, etc. around the installation locations of the two imagers, the pixels in the non-sky area in the ground-based cloud map are removed and marked as masks to be repaired.

[0066] Step S 12 .Perform image repair on the mask to be repaired to complete the missing sky information.

[0067] Step S 13 .The distortion correction coefficient of the imager is determined through parameter calibration, and the distortion of the ground-based cloud image is corrected.

[0068] Step S2: Using image segmentation algorithm, the pre-processed ground-based cloud image is divided into different cloud cluster areas.

[0069] The image segmentation algorithm introduces a regression model into the U-Net network, trains it based on a labeled ground-based cloud image dataset, and uses the preprocessed ground-based cloud image as the model input to obtain a pixel classification image of the cloud cluster. The image contains the pixel coordinates of each individual cloud cluster and the classification results that characterize the thickness of the cloud cluster; the total number of pixels belonging to the same individual cloud cluster is used as the cloud amount feature of the individual cloud cluster.

[0070] Step S3. Calculate the cloud base height and horizontal distance of each cloud group relative to the observation point based on the binocular vision principle. Step S4. Simulate the cloud shadow coverage of the photovoltaic power station, obtain the attenuation coefficient, use the attenuation coefficient to correct the irradiation intensity, and input the power prediction model to obtain the ultra-short-term power prediction value. The power prediction model can be the ultra-short-term photovoltaic power prediction model of the improved IAM-CNN-LSTM hybrid neural network mentioned in the invention patent with publication number CN112507793A.

[0071] Example 3

[0072] As another preferred embodiment of the present invention, the present invention includes a method for ultra-short-term power prediction of a photovoltaic power station based on a dual-foundation cloud map, comprising the following steps:

[0073] Step S1. Collect real-time ground cloud images of two imagers of the photovoltaic power station and perform image preprocessing.

[0074] Step S2: Using image segmentation algorithm, the pre-processed ground-based cloud image is divided into different cloud cluster areas.

[0075] Step S3. Calculate the cloud base height and horizontal distance of each cloud cluster relative to the observation point based on the binocular vision principle.

[0076] More specifically, the following steps are included:

[0077] Step S 31 . Obtain the angle α between the line connecting the two imagers and the due east direction and the horizontal distance L.

[0078] Step S 32 .The edge contours of individual clouds are extracted from the two ground-based cloud images respectively, and similarity comparison analysis is performed in turn. The objects with the lowest similarity index and less than the preset threshold are matched as the same cloud.

[0079] Step S 33 .Get the center point of the cloud i that has a successful similarity match Polar coordinate parameters in two ground-based cloud images According to the calibration parameters, the pole diameter Convert to elevation

[0080] Step S 34 . Calculate the horizontal distance between the center of cloud cluster i and imager k by the following method: and cloud base height H i ;

[0081]

[0082]

[0083] Where α is the angle between the line between the two imagers and the east direction; L is the horizontal distance between the two imagers; k is the imager and the ground-based cloud map number taken, which can be 0 or 1; and are the azimuth and elevation of the center point of cloud cluster i in the ground-based cloud image taken by imager k.

[0084] Among them, shape matching can be used to perform similarity analysis on the edge contour of a single cloud cluster, and a similarity index is returned. The lower the result, the higher the degree of similarity matching. If the lowest similarity index is greater than the preset threshold, the matching object can be stretched, scaled, and other scale transformations, and then similarity analysis is performed. This step is repeated until the preset number of times is reached. If the lowest similarity index is still greater than the preset threshold, the cloud cluster object is ignored.

[0085] Due to the length and width limitations of the photovoltaic power station area, the horizontal distance L between the two imagers is usually between 0.8 and 1.5 km. At this distance, the difference in lens distortion produced by the two imagers is small. Therefore, it is not necessary to perform scale transformation such as stretching and scaling on the cloud contour image. If scale transformation is performed, the transformation coefficient can be set to 0.8 to 1.2.

[0086] Step S4. Simulate the cloud coverage of the photovoltaic power station, obtain the attenuation coefficient, use the attenuation coefficient to correct the radiation intensity, and input the power prediction model to obtain the ultra-short-term power prediction value.

[0087] Example 4

[0088] As the best embodiment of the present invention, the present invention includes a method for ultra-short-term power prediction of photovoltaic power plants based on dual-foundation cloud images. Figure 1 , including the following steps:

[0089] Step S1: Collect real-time ground-based cloud images from two all-sky imagers of the photovoltaic power station and perform image preprocessing.

[0090] Among them, image preprocessing includes the following steps:

[0091] Step S 11 .According to the obstruction of buildings, brackets, shading belts, etc. around the installation locations of the two all-sky imagers, the pixels in the non-sky area in the ground-based cloud map are removed and marked as masks to be repaired.

[0092] Step S 12 .Use the Criminisi algorithm to perform image restoration on the mask to be repaired and complete the missing sky information.

[0093] Step S 13 .The distortion correction coefficient of the imager is determined through parameter calibration, and the distortion of the ground-based cloud image is corrected.

[0094] Step S2. Use the image segmentation algorithm to divide the preprocessed ground-based cloud image into different cloud cluster areas. Specifically, the image segmentation algorithm introduces a regression model into the U-Net network, trains based on a labeled ground-based cloud image dataset, and uses the preprocessed ground-based cloud image as the model input to obtain a pixel classification image of the cloud cluster, which contains the pixel coordinates of each single cloud cluster and the classification result representing the thickness of the cloud cluster; the total number of pixels belonging to the same single cloud cluster is used as the cloud amount feature of the single cloud cluster.

[0095] The labeled ground-based cloud image dataset uses the open source dataset Singapore Whole sky IMagingSEGmentation Database as its basis. The mean grayscale value of the individual cloud cluster area in the cloud image is calculated, and K-means clustering is performed according to the pre-set number of classifications. The clustering results are used as cloud thickness classification labels.

[0096] Step S3. Calculate the cloud base height and horizontal distance of each cloud cluster relative to the observation point based on binocular vision principle. Figure 2 , specifically including the following steps:

[0097] Step S 31 . Obtain the angle α between the line connecting the two all-sky imagers and the due east direction and the horizontal distance L.

[0098] Step S 32 .The edge contours of individual clouds are extracted from the two ground-based cloud images respectively, and similarity comparison analysis is performed in turn. The objects with the lowest similarity index and less than the preset threshold are matched as the same cloud.

[0099] Shape matching can be used to perform similarity analysis on the edge contours of a single cloud cluster and return a similarity index. The lower the result, the higher the degree of similarity matching. If the lowest similarity index is greater than the preset threshold, the matching object can be stretched, scaled, and scaled, and then similarity analysis can be performed. This step is repeated until the preset number of times is reached. If the lowest similarity index is still greater than the preset threshold, the cloud cluster object is ignored.

[0100] Due to the length and width limitations of the photovoltaic power station area, the horizontal distance L between the two all-sky imagers is usually between 0.8 and 1.5 km. At this distance, the difference in lens distortion produced by the two imagers is small. Therefore, it is not necessary to perform scale transformations such as stretching and scaling on the cloud contour image. If scale transformation is performed, the transformation coefficient can be set to 0.8 to 1.2.

[0101] Step S 33 .Get the center point of the cloud i that has a successful similarity match Polar coordinate parameters in two ground-based cloud images According to the calibration parameters, the pole diameter Convert to elevation

[0102] Step S 34 . Calculate the horizontal distance between the center of cloud cluster i and imager k by the following method: and cloud base height H i ;

[0103]

[0104]

[0105] Where α is the angle between the line between the two imagers and the east direction; L is the horizontal distance between the two imagers; k is the imager and the ground-based cloud map number taken, which can be 0 or 1; and are the azimuth and elevation of the center point of cloud cluster i in the ground-based cloud image taken by imager k;

[0106] Step S4. Simulate the cloud coverage of the photovoltaic power station, obtain the attenuation coefficient, use the attenuation coefficient to correct the radiation intensity, and input the power prediction model to obtain the ultra-short-term power prediction value.

[0107] Specifically, the simulation of cloud shadow coverage on the ground of a photovoltaic power station includes:

[0108] The solar azimuth angle θ is calculated according to the following method s and the altitude angle z s :

[0109]

[0110] ω=(t-12)×15°

[0111]

[0112]

[0113] And calculate the ground projection of the center point of cloud cluster i in turn Polar coordinate parameters of relative imager k

[0114]

[0115]

[0116]

[0117] It is determined in turn whether the ground projections of all clouds fall on the areas of the photovoltaic power station.

[0118] In the above formula, n is the day of the year; t is the collection time; is the latitude of the observation point; are the horizontal distance and azimuth of the ground projection of the center point of cloud cluster i relative to imager k; H i are the horizontal distance of the center of cloud cluster i relative to imager k and the cloud base height; z s ,θ s are the solar altitude and azimuth respectively.

[0119] The attenuation coefficient is the reciprocal of the product of the weighted sum of the accumulated cloud volume and cloud thickness of all clouds in the photovoltaic power station area or subarray and the radiation intensity influence factor. The smaller the coefficient, the greater the impact of clouds blocking photovoltaic panels on power generation. The calculation method of the corrected irradiation intensity is as follows;

[0120]

[0121] Where R is the real-time irradiation intensity; R j ' is the corrected irradiation intensity for the photovoltaic power station area or subarray j; μ j is the attenuation coefficient; μ1 and μ2 are proportional coefficients; N j is the number of cloud clusters falling in region j; and are the cloud amount and cloud thickness parameters of cloud cluster i respectively; τ is the radiation intensity influencing factor, which is positively correlated with R.

[0122] For further information, please refer to the attached manual. Figure 3 ,The power prediction model adopts the LSTM-Attention model combined with the attention mechanism, and ,takes the real-time temperature, wind speed, irradiation intensity, ,irradiation intensity corrected by the ground-based cloud map, real-time ,generated power, historical meteorological and power generation operation data of the PV ,station as the input of the power prediction model to obtain ,the ultra-short-term power prediction value.

[0123] In summary, after reading the present invention document, ordinary technicians in this field can make various other corresponding transformation schemes based on the technical scheme and technical concept of the present invention without creative mental labor, which all fall within the scope of protection of the present invention.

Claims

1. A method for ultra-short-term power prediction of photovoltaic power stations based on dual-foundation cloud images, characterized in that: The following steps are involved: Step S1. Collect real-time ground cloud images of two imagers of the photovoltaic power station and perform image preprocessing; Step S2. Using an image segmentation algorithm, the preprocessed ground-based cloud image is divided into different cloud cluster areas; Step S3. Calculate the cloud base height and horizontal distance of each cloud cluster relative to the observation point based on the binocular vision principle; Step S4. Simulate the cloud coverage of the photovoltaic power station, obtain the attenuation coefficient, use the attenuation coefficient to correct the radiation intensity, and input the power prediction model to obtain the ultra-short-term power prediction value.

2. The method for ultra-short-term power prediction of a photovoltaic power station based on a dual-foundation cloud map according to claim 1 is characterized in that: The image preprocessing in step S1 specifically includes the following steps: Step S 11 .Remove the pixels in the non-sky area of ​​the ground-based cloud image and mark them as masks to be repaired; Step S 12 .Perform image restoration on the mask to be restored to complete the missing sky information; Step S 13 .The distortion correction coefficient of the all-sky imager is determined through parameter calibration, and the distortion of the ground-based cloud image is corrected.

3. The method for ultra-short-term power prediction of a photovoltaic power station based on a dual-foundation cloud map according to claim 1 is characterized in that: The step S2 specifically refers to: the image segmentation algorithm introduces a regression model into the U-Net network, is trained based on a labeled ground-based cloud image dataset, and the preprocessed ground-based cloud image is used as the model input to obtain a pixel classification image of the cloud cluster, wherein the image contains the pixel coordinates of each single cloud cluster and the classification result representing the thickness of the cloud cluster; the total number of pixels belonging to the same single cloud cluster is used as the cloud amount feature of the single cloud cluster.

4. The method for ultra-short-term power prediction of a photovoltaic power station based on a dual-foundation cloud map according to claim 3 is characterized in that: The labeled ground-based cloud image dataset uses an open source dataset as its basis, calculates the mean grayscale value of the individual cloud cluster area in the ground-based cloud image, clusters it according to a pre-set number of classifications, and uses the clustering results as cloud thickness classification labels.

5. The method for ultra-short-term power prediction of a photovoltaic power station based on a dual-foundation cloud map according to claim 1 is characterized in that: The step S3 specifically comprises the following steps: Step S 31 . Obtain the angle α between the line connecting the two imagers and the east direction and the horizontal distance L; Step S 32 .Extract the edge contours of individual clouds from the two ground-based cloud images respectively, perform similarity comparison analysis in turn, and match the objects with the lowest similarity index that is less than the preset threshold as the same cloud cluster; Step S 33 .Get the center point of cloud i that has successful similarity matching Polar coordinate parameters in two ground-based cloud images According to the calibration parameters, the pole diameter Convert to elevation Step S 34 . Calculate the horizontal distance between the center of cloud cluster i and imager k by the following method: and cloud base height H i ; Where α is the angle between the line between the two imagers and the east direction; L is the horizontal distance between the two imagers; k is the imager and the ground-based cloud map number taken, which can be 0 or 1; and are the azimuth and elevation of the center point of cloud cluster i in the ground-based cloud image taken by imager k.

6. The method for ultra-short-term power prediction of a photovoltaic power station based on a dual-foundation cloud map according to claim 5 is characterized in that: The step S 32 The similarity analysis of the edge contour of the single cloud cluster specifically refers to: using shape matching, returning the similarity index, if the lowest similarity index is greater than the preset threshold, transforming the matching object, and then performing similarity analysis, repeating this step until the preset number of times is reached, if the lowest similarity index is still greater than the preset threshold, then ignoring the cloud cluster object.

7. The method for ultra-short-term power prediction of a photovoltaic power station based on a dual-foundation cloud map according to claim 6 is characterized in that: The horizontal distance L between the two imagers is between 0.8 and 1.5 km.

8. A method for ultra-short-term power prediction of a photovoltaic power station based on dual-foundation cloud images according to any one of claims 3 to 7, characterized in that: The simulation of the cloud shadow coverage on the ground of the photovoltaic power station in step S4 specifically includes: Calculate the solar azimuth θ s and the altitude angle z s : ω=(t-12)×15° Calculate the ground projection of the center point of cloud cluster i in sequence Polar coordinate parameters of relative imager k Determine in turn whether the ground projections of all clouds fall on each area of ​​the photovoltaic power station; In the above formula, n is the day of the year; t is the collection time; is the latitude of the observation point; are the horizontal distance and azimuth of the ground projection of the center point of cloud cluster i relative to imager k; H i are the horizontal distance of the center of cloud cluster i relative to imager k and the cloud base height; z s ,θ s are the solar altitude and azimuth respectively.

9. The method for ultra-short-term power prediction of a photovoltaic power station based on dual-foundation cloud images according to claim 8, characterized in that: In step S4, using the attenuation coefficient to correct the irradiation intensity specifically refers to: Where R is the real-time irradiation intensity; R j ' is the corrected irradiation intensity for the photovoltaic power station area or subarray j; μ j is the attenuation coefficient; μ1 and μ2 are proportional coefficients; N j is the number of cloud clusters falling in region j; and are the cloud amount and cloud thickness parameters of cloud cluster i respectively; τ is the radiation intensity influencing factor, which is positively correlated with R.

10. The method for ultra-short-term power prediction of a photovoltaic power station based on dual-foundation cloud images according to claim 1, characterized in that: The power prediction model adopts the LSTM-Attention model combined with the attention mechanism, and takes the real-time temperature, wind speed, irradiation intensity of the meteorological monitoring station of the photovoltaic power station, the irradiation intensity corrected by the dual-base cloud map, the real-time power generation power, the historical meteorological and power generation operation data as the power prediction model input to obtain the ultra-short-term power prediction value.

Images