Ultra-short-term photovoltaic power prediction method based on ground-based cloud atlas and multi-source data

Abstract

The invention provides an ultra-short-term photovoltaic power prediction method based on a foundation cloud picture and multi-source data, and relates to the technical field of solar radiation prediction, and the method comprises the steps: collecting a sky image data set which comprises a foundation cloud picture image, solar radiation data and meteorological information data; constructing a spherical sky image based on the sun position in the ground-based cloud picture image, extracting the spherical sky image, projecting the spherical sky image to a two-dimensional plane to obtain a planar sky image, and obtaining first theoretical total sun radiation through sun position information and a clear sky irradiance function; the meteorological information data and the total solar radiation are preprocessed, and standard meteorological information data and standard total solar radiation are obtained respectively; and constructing a sky image fusion prediction model, inputting the standard meteorological information data, the standard total solar radiation and the first theoretical total solar radiation into the sky image fusion prediction model, and obtaining predicted total solar radiation and predicted photovoltaic power. According to the method, limitation possibly brought by a single data source can be reduced, and comprehensiveness and accuracy of solar radiation prediction and photovoltaic power prediction are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of solar radiation prediction, and in particular to an ultra-short-term photovoltaic power prediction method based on ground-based cloud images and multi-source data. Background Art

[0002] As the global demand for renewable energy continues to grow, the development and utilization of solar energy as a clean and sustainable energy source has received widespread attention. Solar power generation, especially photovoltaic power generation, has become an important part of the new energy field due to its environmental friendliness and high energy conversion efficiency. However, the efficiency and stability of solar power generation are directly affected by solar radiation conditions, and the temporal and spatial variations of solar radiation are highly uncertain and complex. However, since solar radiation is affected by multiple factors such as weather conditions, seasonal changes, and geographical location, the output power of photovoltaic power generation has obvious intermittent and instability, which poses a challenge to the stable operation of the power grid and the dispatching management of the power system.

[0003] The Chinese patent with publication number CN113537561A discloses a method and system for predicting ultra-short-term solar radiation based on ground-based cloud maps, which includes the following steps: obtaining a time series data set of ground-based cloud map images and a time series data set of solar radiation before the prediction moment; inputting the time series data set of ground-based cloud map images into a pre-trained convolutional neural network model to extract a time series data set of cloud map feature vectors; inputting the time series data set of solar radiation and the time series data set of cloud map feature vectors into a pre-established recurrent neural network model to obtain solar radiation data at the prediction moment; and using the solar radiation data at the prediction moment as an ultra-short-term solar radiation prediction result. However, the prediction model based on historical solar radiation and numerical weather forecast data provided by the above application is restricted by data accuracy and algorithm principles, and it is difficult to adapt to the prediction requirements of ultra-short-term time scales, resulting in insufficient accuracy in solar radiation prediction. Therefore, it is very necessary to provide a method and system for tracing the fault source of wind power equipment based on a multi-level model of fault decoupling to improve the accuracy of fault source location. Summary of the invention

[0004] In view of this, the present invention proposes an ultra-short-term photovoltaic power prediction method based on ground-based cloud images and multi-source data. By combining sky images, solar radiation data and meteorological information data, multiple data sources are fully utilized to capture more factors affecting solar radiation, reduce the limitations that may be brought about by a single data source, and improve the comprehensiveness and accuracy of solar radiation prediction and photovoltaic power prediction.

[0005] The present invention provides an ultra-short-term photovoltaic power prediction method based on ground-based cloud images and multi-source data, the method comprising:

[0006] Collecting a sky image data set, wherein the sky image data includes ground-based cloud image images, solar radiation data, and meteorological information data;

[0007] Based on the sun position in the ground-based cloud image, a spherical sky image is constructed, the spherical sky image is extracted and projected onto a two-dimensional plane to obtain a planar sky image, and a first theoretical total solar radiation is obtained through the sun position information and the clear sky irradiance function;

[0008] Preprocessing the meteorological information data and the total solar radiation to obtain standard meteorological information data and standard total solar radiation respectively;

[0009] A sky image fusion prediction model is constructed, and the standard meteorological information data, the standard total solar radiation and the first theoretical total solar radiation are input into the sky image fusion prediction model to obtain predicted total solar radiation and predicted photovoltaic power.

[0010] On the basis of the above technical solution, preferably, the step of extracting the spherical sky image and projecting it onto a two-dimensional plane to obtain a planar sky image specifically includes:

[0011] Calculating the simulated sun position in the ground-based cloud image by a barycentric method, and projecting the spherical sky image onto a unit sphere according to the simulated sun position, the actual sun position and a reprojection function to obtain a unit sphere sky image;

[0012] Based on the cloud image color moment function, the texture matrix function and the unit sphere sky image, respectively calculating the color features of the unit sphere sky image and the cloud texture features;

[0013] The unit sphere sky image is expanded by using the latitude and longitude method to convert the unit sphere sky image into a plane sky image.

[0014] Based on the above technical solutions, preferably, the color features include a first-order matrix of the average color distribution in the unit sphere sky image, a second-order matrix of the discrete degree of color distribution in the unit sphere sky image, and a third-order matrix of the skewness of color distribution in the unit sphere sky image, and the cloud texture features include cloud texture energy, cloud texture entropy, cloud texture contrast, cloud texture correlation and cloud texture inverse difference.

[0015] More preferably, the expression of the reprojection function is:

[0016] θ d =k0θ+k1θ 3 +k2θ 5 +k3θ 7 +k4θ 9

[0017] Where θ represents the undistorted camera angle, θ d represents the camera angle after distortion, k0 represents the first projection coefficient, k1 represents the second projection coefficient, k2 represents the third projection coefficient, k3 represents the fourth projection coefficient, and k4 represents the fifth projection coefficient.

[0018] More preferably, the calculating the simulated sun position in the ground-based cloud image by the centroid method specifically includes:

[0019] Screening the ground-based cloud image to obtain a standard cloud image without cloud cover, and converting the standard cloud image into a grayscale cloud image;

[0020] Obtaining a cloud image pixel value threshold according to the maximum inter-class variance method, the brightness of the grayscale cloud image, and the contrast of the grayscale cloud image;

[0021] If the pixel value corresponding to the pixel point in the grayscale cloud image is greater than the cloud image pixel value threshold, the pixel point is determined as a sun pixel point, and the grayscale value and pixel coordinates of the sun pixel point are obtained;

[0022] Based on the barycentric method and the pixel coordinates of the sun pixel point, the sun barycentric coordinates are calculated to obtain the simulated sun position.

[0023] More preferably, projecting the spherical sky image onto a unit sphere to obtain the unit sphere sky image specifically includes:

[0024] Generating uniformly distributed pixel points on the unit sphere according to the Fibonacci grid method;

[0025] Each pixel point of the unit sphere is matched with the spherical sky image, and a pixel value corresponding to the pixel point is obtained through a bilinear interpolation function to obtain the unit sphere sky image.

[0026] More preferably, the preprocessing of the meteorological information data and the total solar radiation specifically includes:

[0027] Performing an anomaly identification operation on the meteorological information data and the total solar radiation based on the DBSCAN clustering algorithm to obtain data anomaly points in the meteorological information data and the total solar radiation;

[0028] A cubic spline fitting is performed on the data points around the data outlier point according to a preset outlier point interval, so that the spline interpolation replaces the outlier value corresponding to the data outlier point.

[0029] More preferably, the expression of the clear sky irradiance function is:

[0030]

[0031] τ=0.56(e -0.56AM +e -0.09AM )

[0032] AM=[(r+c e ) 2 cos 2 (θ)+(2r+1+c e )(1-c e )] 1 / 2 -(r+c e )cosθ

[0033]

[0034] Among them, E0 represents the solar radiation reaching the upper atmosphere before the solar radiation is attenuated by the atmosphere, and E sc represents the solar radiation constant, N represents the serial number of the current day in the year, k represents the adjustment parameter, AM represents the ratio of the path length of solar radiation passing through the atmosphere to the shortest path length, r represents the radius of the earth, and c e represents the cosine value of the solar zenith angle, τ represents the ratio of solar transmitted radiation to incident radiation, E clear Represents the first theoretical total solar radiation.

[0035] More preferably, the constructing of the sky image fusion prediction model, inputting the standard meteorological information data, the standard total solar radiation and the first theoretical total solar radiation into the sky image fusion prediction model to obtain the predicted total solar radiation and the predicted photovoltaic power, specifically includes:

[0036] The standard meteorological information data, the standard total solar radiation and the first theoretical total solar radiation are divided according to a preset ratio to obtain a cloud map training set, a cloud map verification set and a cloud map test set respectively;

[0037] Constructing a sky image fusion prediction model, inputting the cloud image training set into the sky image fusion prediction model for training to obtain an optimized sky image fusion prediction model, and inputting the cloud image test set into the optimized sky image fusion prediction model to obtain a second theoretical total solar radiation;

[0038] The difference between the second theoretical total solar radiation and the actual solar radiation in the solar radiation data is used as a regression target, and the second theoretical total solar radiation is compensated to obtain predicted total solar radiation and predicted photovoltaic power.

[0039] More preferably, the meteorological information data includes temperature, relative humidity, wind speed and direction, zenith angle and azimuth angle, cloud cover, air pressure, albedo and broadband turbidity.

[0040] The ultra-short-term photovoltaic power prediction method based on ground-based cloud images and multi-source data provided by the present invention has the following beneficial effects compared with the prior art:

[0041] (1) By combining sky images, solar radiation data, and meteorological information data, the multiple data sources are fully utilized to capture more factors affecting solar radiation, reduce the limitations that may be brought about by a single data source, and improve the comprehensiveness and accuracy of solar radiation prediction and photovoltaic power prediction. The ground-based cloud map is converted into a spherical sky image and then projected into a flat sky image. This process can better preserve the spatial information of cloud distribution and more accurately represent the impact of clouds on solar radiation at different solar altitudes. At the same time, feature screening and outlier processing are performed on multi-source data, which improves the accuracy of total solar radiation prediction and photovoltaic power prediction.

[0042] (2) The centroid method is used to calculate and simulate the sun's position, which can more accurately locate the sun in the image and adapt to different lighting conditions and cloud distributions, thereby improving the accuracy of subsequent data processing. Projecting the spherical sky image onto the unit sphere effectively standardizes the image space and eliminates the geometric distortion caused by shooting angles and equipment differences. The unit sphere projection retains the full range of spatial information of the sky. Combining the cloud image color moment function and texture matrix function, the color features and cloud texture features are extracted simultaneously, which can more comprehensively describe the properties of the cloud layer. The extraction of color features helps to distinguish different types of clouds and the clarity of the sky. The calculation of cloud texture features can capture the structure and distribution pattern of the cloud layer, which helps to predict short-term cloud changes and solar radiation fluctuations.

[0043] (3) The Fibonacci grid method is used to generate uniformly distributed pixel points on the surface of the unit sphere, ensuring that the sampling of the entire sky is consistent and comprehensive, avoiding the problem of over-sampling near the poles of traditional latitude and longitude grids. By evenly distributing the sampling points on the unit sphere, the spatial information of the original spherical sky image is better preserved. Compared with directly projecting the spherical image onto a plane, this method can significantly reduce geometric distortion, especially in the edge area of ​​the sphere, thereby maintaining the overall integrity and accuracy of the image. The Fibonacci grid method allows the number of sampling points to be flexibly adjusted, so that the image resolution can be adjusted according to needs, and can adapt to different computing resources and accuracy requirements. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0045] Figure 1 A schematic flow chart of an ultra-short-term photovoltaic power prediction method based on ground-based cloud images and multi-source data provided by the present invention. DETAILED DESCRIPTION

[0046] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0047] The present invention discloses an ultra-short-term photovoltaic power prediction method based on ground-based cloud images and multi-source data. Figure 1 The method includes steps S1 to S4.

[0048] Step S1, collecting a sky image data set, wherein the sky image data includes ground-based cloud image images, solar radiation data, and meteorological information data.

[0049] In this step, ground-based cloud imagery is taken using ground-based cloud imagery instruments. These ground-based cloud imagery images record cloud and solar information. A solar radiometer is used to measure the solar radiation energy reaching the ground. The solar radiation energy includes direct radiation (direct light from the sun) and scattered radiation (scattered light from the sky).

[0050] Furthermore, the all-sky imager collects ground-based cloud images, and first screens the images to select cloud images during the day. When screening, delete images outside the sunrise and sunset times according to the local sunrise and sunset times. In actual implementation, since the total solar radiation near sunrise and sunset is small, the collection time period can be reduced, such as 7:00 to 18:00, which can also meet the algorithm requirements. The collection frequency of cloud images is at least 10 minutes / time. Use a solar radiometer to collect total solar radiation, and the measurement unit is W / m2. Use meteorological sensors to observe meteorological information, including temperature, humidity, wind speed, wind direction, air pressure, etc.

[0051] In one example, an OTT Parsivel2 all-sky imager can be installed on an open rooftop platform at an altitude of 50 meters to collect cloud images with a resolution of 1600x1200 pixels every 10 minutes from 7:00 to 18:00 every day in JPEG format. At the same time, a Kipp & Zonen CMP21 high-precision thermopile radiometer is used to record total solar radiation data every minute with an accuracy of ±2% and a measurement range of 0-4000W / m 2. Meteorological data collection uses Vaisala WXT536 multi-parameter meteorological sensor, which records data every 5 minutes, including air temperature (accuracy ± 0.3 ° C), relative humidity (accuracy ± 3%), wind speed (accuracy ± 3% or ± 0.3m / s), wind direction (accuracy ± 3°) and air pressure (accuracy ± 0.5hPa). All equipment is installed in the same location to ensure continuous collection 365 days a year. After data collection, an automated script is used for preliminary screening, images before 7:00 and after 18:00 are deleted, and outliers (such as sudden changes in radiation values ​​or meteorological data outside the normal range) are marked. Dell PowerEdge R740 local server is used to store data, equipped with a 20TB RAID5 storage array, and each data point is attached with a timestamp accurate to seconds. Equipment inspection and cleaning are carried out weekly, and radiometers and meteorological sensors are calibrated monthly. In addition, the precise GPS coordinates and altitude of the collection site are recorded, and special weather events are recorded by the on-site operator.

[0052] In this embodiment, the meteorological information data includes temperature, relative humidity, wind speed and direction, zenith angle and azimuth, cloud cover, air pressure, albedo, and broadband turbidity. The historical total solar radiation data is calculated by collecting direct radiation and scattered horizontal radiation, and the calculation formula is:

[0053] GHI=DHI+(cosθ r ×DNI)

[0054] Among them, θ r represents the solar zenith angle, GHI represents the global solar historical radiation, DHI represents the diffuse horizontal radiation, and DNI represents the direct radiation.

[0055] Step S2, based on the sun position in the ground-based cloud image, construct a spherical sky image, extract the spherical sky image and project it onto a two-dimensional plane to obtain a planar sky image, and obtain the first theoretical total solar radiation through the sun position information and the clear sky irradiance function.

[0056] This step also includes steps S21 to S23.

[0057] Step S21, calculating the simulated sun position in the ground-based cloud image by the centroid method, and projecting the spherical sky image onto the unit sphere according to the simulated sun position, the actual sun position and the reprojection function to obtain the unit sphere sky image.

[0058] The expression of the reprojection function is:

[0059] θ d =k0θ+k1θ 3 +k2θ 5 +k3θ 7 +k4θ 9

[0060] Where θ represents the undistorted camera angle, θ d represents the camera angle after distortion, k0 represents the first projection coefficient, k1 represents the second projection coefficient, k2 represents the third projection coefficient, k3 represents the fourth projection coefficient, and k4 represents the fifth projection coefficient. By obtaining the sun position in each ground-based cloud map and its corresponding actual sun position, the first to fifth projection coefficients can be obtained by using the least squares method. In this way, the two-dimensional image can be reprojected onto a three-dimensional hemisphere to simulate the sky.

[0061] In this embodiment, the ground-based cloud map image is screened to obtain a standard cloud map image without cloud cover, and the standard cloud map image is converted into a grayscale cloud map image; the cloud map pixel value threshold is obtained according to the maximum inter-class variance method, the brightness of the grayscale cloud map image, and the contrast of the grayscale cloud map image; if the pixel value corresponding to the pixel point in the grayscale cloud map image is greater than the cloud map pixel value threshold, the pixel point is determined as a sun pixel point, and the grayscale value and pixel coordinates of the sun pixel point are obtained; based on the centroid method and the pixel coordinates of the sun pixel point, the sun centroid coordinates are calculated to obtain a simulated sun position, and the actual sun position is calculated according to the shooting date.

[0062] Filter the images with complete sun and select the pixel where the sun is located. Count the number of occurrences of each pixel value to get the image histogram. Traverse all possible thresholds from the minimum pixel value to the maximum pixel value. For each threshold, divide the image into foreground and background, and calculate the variance between the two parts. Select the threshold that maximizes the inter-class variance as the optimal threshold to segment the sun.

[0063] The Otsu method is applied to analyze the grayscale cloud image. An optimal threshold is automatically determined by maximizing the variance between the foreground (cloud) and background (sky) categories. The Otsu method is particularly suitable for bimodal histograms and can effectively separate clouds and sky. The average value of all pixel values ​​in the grayscale cloud image is calculated to obtain the overall brightness. This value reflects the overall lighting conditions of the image and helps to adjust the threshold to adapt to different lighting environments. The contrast is estimated by calculating the standard deviation of the pixel values ​​of the grayscale cloud image. The contrast reflects the degree of difference between clouds and sky. High contrast usually means that the boundary between clouds and sky is clearer. The above three factors are weighted averaged to obtain the final cloud image pixel value threshold. Specifically: a higher weight (such as 60%) is given to the threshold obtained by the Otsu method because it can usually provide a good initial threshold. Image brightness and contrast are each given a lower weight (such as 20% each) for fine-tuning the threshold. The final threshold = 0.6*Otsu threshold + 0.2*image brightness + 0.2*image contrast.

[0064] Furthermore, uniformly distributed pixel points are generated on the unit sphere according to the Fibonacci grid method; each pixel point of the unit sphere is matched with the spherical sky image, and the pixel value corresponding to the pixel point is obtained through a bilinear interpolation function to obtain the unit sphere sky image.

[0065] To solve the problem that pixel projection results in more concentrated pixels on both sides of the hemisphere and sparse pixels on the top of the sphere, the inverse projection method based on spherical uniform sampling is used. The Fibonacci grid method is used to generate uniformly distributed, dense and chaotic points on the sphere. The coordinates of the nth point are as follows:

[0066] z n =(2n-1) / L-1

[0067]

[0068] Among them, the constant x n Indicates the x-axis coordinate value of the nth point, y n Indicates the y-axis coordinate value of the nth point, z n Indicates the z-axis coordinate value of the nth point.

[0069] For each point on the sphere, the corresponding point on the two-dimensional plane can be obtained, and then the corresponding pixel value can be obtained through bilinear interpolation.

[0070] Furthermore, generating uniformly distributed pixel points on the unit sphere according to the Fibonacci grid method specifically includes: determining the total number L of points that need to be generated on the unit sphere surface, and using the golden angle as a rotation increment to ensure that the distribution of points does not show obvious regular clustering.

[0071] The Fibonacci grid method generates uniformly distributed pixel points on the surface of the unit sphere, ensuring that the sampling of the entire sky is consistent and comprehensive, avoiding the problem of over-sampling of traditional latitude and longitude grids near the poles. By evenly distributing sampling points on the unit sphere, the spatial information of the original spherical sky image is better preserved. Compared with directly projecting the spherical image onto a plane, this method can significantly reduce geometric distortion, especially in the edge area of ​​the sphere, thereby maintaining the overall integrity and accuracy of the image. The Fibonacci grid method allows flexible adjustment of the number of sampling points, so that the resolution of the image can be adjusted as needed, and can adapt to different computing resources and accuracy requirements.

[0072] Step S22, based on the cloud image color moment function, the texture matrix function and the unit sphere sky image, respectively calculate the color features of the unit sphere sky image and the cloud texture features.

[0073] In this step, the color features include a first-order matrix of the average color distribution in the unit sphere sky image, a second-order matrix of the discrete degree of color distribution in the unit sphere sky image, and a third-order matrix of the skewness of color distribution in the unit sphere sky image. The cloud texture features include cloud texture energy, cloud texture entropy, cloud texture contrast, cloud texture correlation, and cloud texture inverse difference.

[0074] The expression of the cloud image color moment function is:

[0075]

[0076] Among them, p i,j represents the i-th color component of the j-th pixel in the unit sphere sky image, m represents the number of pixels in the unit sphere sky image, μ i The first-order matrix representing the average value of the color distribution in the unit sphere sky image, σ i The second-order matrix representing the degree of discreteness of the color distribution in the unit sphere sky image, si represents the third-order matrix representing the degree of skewness of the color distribution in the unit sphere sky image.

[0077] Calculate the grayscale co-occurrence matrix of the spherical image and extract the texture features of the sky and clouds. The points uniformly sampled on the sphere also have equidistant points, so the grayscale co-occurrence matrix can be obtained as with a two-dimensional image. According to the grayscale co-occurrence matrix, the cloud texture information can be obtained, including cloud texture energy, cloud texture entropy, cloud texture contrast, cloud texture correlation, and cloud texture inverse difference. The calculation formula is as follows:

[0078]

[0079] Entropy=-∑ i ∑ j C ij logC ij

[0080] Contrast=∑ i ∑ j (ij) 2 C ij

[0081]

[0082] Among them, Engergy represents the cloud texture energy, Entropy represents the cloud texture entropy value, Contrast represents the cloud texture contrast, Correlation represents the cloud texture correlation, IDM represents the cloud texture inverse difference, C ij Whether to represent the normalized gray-level co-occurrence matrix.

[0083] Step S23, expanding the unit sphere sky image by using the latitude and longitude method, and converting the unit sphere sky image into a plane sky image.

[0084] In this step, the width and height of the plane sky image are determined. Usually, the width represents longitude (0 to 360 degrees) and the height represents latitude (-90 to 90 degrees). For each pixel (x, y) of the plane image, the corresponding longitude and latitude are calculated, which can be expressed as: longitude = (x / image width) * 360°-180°, latitude = 90 degrees - (y / image height) * 180°. The longitude and latitude are converted to 3D coordinates (x, y, z) on the unit sphere, x = cos (latitude) * cos (longitude); y = cos (latitude) * sin (longitude); z = sin (latitude).

[0085] In this embodiment, the expression of the clear sky irradiance function is:

[0086]

[0087] AM=[(r+c e ) 2 cos 2 (θ)+(2r+1+c e )(1-c e )] 1 / 2 -(r+c e )cosθ

[0088]

[0089] Among them, E0 represents the solar radiation reaching the upper atmosphere before the solar radiation is attenuated by the atmosphere, and E sc represents the solar radiation constant, N represents the serial number of the current day in the year, k represents the adjustment parameter, AM represents the ratio of the path length of solar radiation passing through the atmosphere to the shortest path length, r represents the radius of the earth, and c e represents the cosine value of the solar zenith angle, τ represents the ratio of solar transmitted radiation to incident radiation, E clear represents the first theoretical total solar radiation, E sc =1367W / m 2 .

[0090] The centroid method is used to calculate and simulate the sun's position, which can more accurately locate the sun in the image and adapt to different lighting conditions and cloud distributions, thereby improving the accuracy of subsequent data processing. Projecting the spherical sky image onto a unit sphere effectively standardizes the image space and eliminates geometric distortion caused by shooting angles and equipment differences. The unit sphere projection retains the sky's all-round spatial information. Combining the cloud image color moment function and texture matrix function, the color features and cloud texture features are extracted simultaneously, which can more comprehensively describe the properties of the clouds. The extraction of color features helps to distinguish different types of clouds and the clarity of the sky. The calculation of cloud texture features can capture the structure and distribution patterns of clouds, which helps to predict short-term cloud changes and solar radiation fluctuations.

[0091] Step S3, preprocessing the meteorological information data and the total solar radiation to obtain standard meteorological information data and standard total solar radiation respectively.

[0092] This step also includes steps S31 to S32.

[0093] Step S31, performing anomaly identification operations on meteorological information data and total solar radiation based on the DBSCAN clustering algorithm to obtain data anomaly points in the meteorological information data and total solar radiation.

[0094] In this step, based on the collected meteorological information data and total solar radiation, ensure that the above data contains timestamps for time series analysis, perform preliminary cleaning of the data to handle missing values, use the mean or median to fill, and perform preliminary screening of outliers to remove obviously unreasonable data points. Create new time-related features, such as the hour of the day, month, season, etc. These features help capture the periodicity and seasonality of meteorological data. Consider adding lagged features, such as data from the previous hour or day, to capture short-term trends.

[0095] StandardScaler is used to standardize all numerical features to ensure that features of different scales are treated fairly in the DBSCAN algorithm. After standardization, all features will have zero mean and unit variance. The K-distance graph method is used to select the appropriate eps value, and the K-distance graph is plotted to observe the inflection point of the curve. For minPts, it is generally selected to be about twice the number of features, but the size of the dataset and the expected cluster size must also be considered.

[0096] Run the DBSCAN algorithm with the selected eps and minPts parameters. The algorithm will classify the data points into core points, boundary points, and noise points. Noise points (points marked as -1) will be considered as potential outliers. Perform a detailed analysis of the outliers identified by DBSCAN, check the distribution of these points in various feature dimensions, pay special attention to the points with abnormal total solar radiation, and analyze their relationship with other meteorological parameters. Analyze the distribution of outliers in the time dimension to check whether there is a phenomenon that outliers are more concentrated in certain specific time periods (such as specific months or specific times of the day). Based on the results of DBSCAN, set a reasonable anomaly threshold for each feature. You can use the statistical characteristics of normal data points (non-outliers), such as the mean plus or minus 3 times the standard deviation, as the initial threshold.

[0097] Furthermore, after the outliers in the data set are identified by DBSCAN or other anomaly detection methods, the indexes or timestamps of these outliers are saved in a list, and a complete data set containing timestamps and corresponding values ​​(such as temperature, solar radiation, etc.) is prepared. A suitable time window is determined for each outlier. For example, for data recorded once an hour, the data 12 hours before and after the outlier can be selected, for a total of 24 hours.

[0098] In one example, the distance between data points is calculated first, and the neighborhood radius of the DBSCAN clustering algorithm is set between 0.5 and 2 times the average distance between data points. The minimum number of points is tried according to the density of the data, between 3 and 10.

[0099] Step S32, performing cubic spline fitting on the data points around the data outlier point according to the preset outlier point interval, so that the spline interpolation replaces the outlier value corresponding to the data outlier point.

[0100] In this step, for each identified outlier, the following steps are performed: determine the exact timestamp of the outlier; calculate the start and end time of the data to be collected based on the defined window size; extract all data points within this time range from the data set; remove the points marked as outliers from the extracted data and keep only normal data points; ensure that the number of normal data points collected is sufficient for reliable spline fitting. Use the Spearman correlation analysis method to perform correlation analysis on the data, and select features whose Spearman correlation coefficient with the target variable is greater than the threshold as important features, for example, 0.75 can be selected.

[0101] Step S4, constructing a sky image fusion prediction model, inputting standard meteorological information data, standard total solar radiation and the first theoretical total solar radiation into the sky image fusion prediction model, and obtaining predicted total solar radiation and predicted photovoltaic power.

[0102] In one example, the collected historical ground-based cloud images and the corresponding original data sets such as total solar radiation and meteorological information, as well as the calculated future theoretical total solar radiation data, are preprocessed and standardized for the training and prediction of the sky image fusion prediction model. The sky image fusion prediction model adopts a combined structure of CNN-Transform. CNN extracts image features, and Transform is used to capture time series information. The network's loss function is MSE, and since the sequence has autocorrelation, taking the difference between the current moment and the theoretical radiance as the regression target can eliminate the autocorrelation.

[0103] A sky image fusion prediction model is constructed, and the processed ground-based cloud map, meteorological information and theoretical total solar radiation data are input into the model. The CNN part processes the image data, and the Transformer part processes the numerical data, and the two are fused together through the attention mechanism. The model predicts the total solar radiation in the future based on the rules learned from the training data. The CNN neural network model is trained based on the historical ground-based cloud map image dataset, and the Transformer neural network model is trained based on the meteorological information and cloud map feature dataset and the cloud map image feature data obtained by CNN, and finally the total solar radiation data at the prediction time is obtained.

[0104] The loss function of the sky image fusion prediction model includes the following calculation formula:

[0105]

[0106] Among them, y represents the true value of the neural network input, which is the difference between the theoretical total solar radiation calculated on that day and the actual total solar radiation. It represents the predicted value output by the neural network. The predicted total solar radiation can be obtained by subtracting it from the theoretical total solar radiation on the predicted day.

[0107] This step also includes steps S41 to S42.

[0108] Step S41, dividing the standard meteorological information data, the standard total solar radiation and the first theoretical total solar radiation according to a preset ratio to obtain a cloud map training set, a cloud map verification set and a cloud map test set respectively.

[0109] In this step, the ratio of the cloud map training set, cloud map validation set, and cloud map test set is 8:1:1.

[0110] Step S42, constructing a sky image fusion prediction model, inputting the cloud image training set into the sky image fusion prediction model for training to obtain an optimized sky image fusion prediction model, and inputting the cloud image test set into the optimized sky image fusion prediction model to obtain a second theoretical total solar radiation.

[0111] In this step, the width and height of the plane sky image are scaled to W*H to meet the input requirements of the CNN neural network. For example, it can be scaled to 224*224. The loss function of the sky image fusion prediction model is set as follows. The parameters of the model are adjusted according to the experimental results. The convolution kernel size of the CNN network in the sky image fusion prediction model is usually 3x3 or 5x5; the number of attention heads of the Transformer part is generally 4-8, which can be adjusted according to the model complexity and computing resources. The number of encoder layers is generally 2-6 layers, and the number of decoder layers is generally the same as or slightly less than the number of encoder layers. The batch size can be set to 32 according to the video memory, and the number of training rounds is between 50 and 100.

[0112] Step S43, taking the difference between the second theoretical total solar radiation and the actual solar radiation in the solar radiation data as a regression target, compensating the second theoretical total solar radiation to obtain predicted total solar radiation and predicted photovoltaic power.

[0113] By combining sky images, solar radiation data and meteorological information data, and making full use of multiple data sources, we can capture more factors that affect solar radiation, reduce the limitations that may be brought about by a single data source, and improve the comprehensiveness and accuracy of solar radiation prediction and photovoltaic power prediction. The ground-based cloud map is converted into a spherical sky image and then projected into a flat sky image. This process can better retain the spatial information of cloud distribution and more accurately represent the impact of clouds on solar radiation at different solar altitude angles. At the same time, feature screening and outlier processing are performed on multi-source data, which improves the accuracy of total solar radiation prediction.

[0114] It should be understood that in the present application, "at least one item" means one or more, and "plurality" means two or more. "And / or" is used to describe the association relationship of associated objects, indicating that three relationships may exist. For example, "A and / or B" can mean: only A exists, only B exists, and A and B exist at the same time, where A and B can be singular or plural. The character " / " generally indicates that the previous and next associated objects are in an "or" relationship. "At least one of the following" or similar expressions refers to any combination of these items, including any combination of single or plural items. For example, at least one of A, B or C can mean: A, B, C, "A and B", "A and C", "B and C", or "A and B and C", where A, B, C can be single or multiple.

[0115] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. An ultra-short-term photovoltaic power prediction method based on ground-based cloud images and multi-source data, characterized in that: The method comprises: Collecting a sky image data set, wherein the sky image data includes ground-based cloud image images, solar radiation data, and meteorological information data; Based on the sun position in the ground-based cloud image, a spherical sky image is constructed, the spherical sky image is extracted and projected onto a two-dimensional plane to obtain a planar sky image, and a first theoretical total solar radiation is obtained through the sun position information and the clear sky irradiance function; Preprocessing the meteorological information data and the total solar radiation to obtain standard meteorological information data and standard total solar radiation respectively; A sky image fusion prediction model is constructed, and the standard meteorological information data, the standard total solar radiation and the first theoretical total solar radiation are input into the sky image fusion prediction model to obtain predicted total solar radiation and predicted photovoltaic power.

2. The method according to claim 1, characterized in that The step of extracting the spherical sky image and projecting it onto a two-dimensional plane to obtain a planar sky image specifically includes: Calculating the simulated sun position in the ground-based cloud image by a barycentric method, and projecting the spherical sky image onto a unit sphere according to the simulated sun position, the actual sun position and a reprojection function to obtain a unit sphere sky image; Based on the cloud image color moment function, the texture matrix function and the unit sphere sky image, respectively calculating the color features of the unit sphere sky image and the cloud texture features; The unit sphere sky image is expanded by using the latitude and longitude method to convert the unit sphere sky image into a plane sky image.

3. The method according to claim 2, characterized in that The color features include a first-order matrix of the color distribution average value in the unit sphere sky image, a second-order matrix of the color distribution discreteness in the unit sphere sky image, and a third-order matrix of the color distribution skewness in the unit sphere sky image. The cloud texture features include cloud texture energy, cloud texture entropy, cloud texture contrast, cloud texture correlation, and cloud texture inverse difference.

4. The method according to claim 2, characterized in that The expression of the reprojection function is: i d =k0θ+k1θ 3 +k2θ 5 +k3θ 7 +k4θ 9 Where θ represents the undistorted camera angle, θ d represents the camera angle after distortion, k0 represents the first projection coefficient, k1 represents the second projection coefficient, k2 represents the third projection coefficient, k3 represents the fourth projection coefficient, and k4 represents the fifth projection coefficient.

5. The method according to claim 2, characterized in that The step of calculating the simulated sun position in the ground-based cloud image by the centroid method specifically includes: Screening the ground-based cloud image to obtain a standard cloud image without cloud cover, and converting the standard cloud image into a grayscale cloud image; Obtaining a cloud image pixel value threshold according to the maximum inter-class variance method, the brightness of the grayscale cloud image, and the contrast of the grayscale cloud image; If the pixel value corresponding to the pixel point in the grayscale cloud image is greater than the cloud image pixel value threshold, the pixel point is determined as a sun pixel point, and the grayscale value and pixel coordinates of the sun pixel point are obtained; Based on the barycentric method and the pixel coordinates of the sun pixel point, the sun barycentric coordinates are calculated to obtain the simulated sun position.

6. The method according to claim 2, characterized in that The projecting the spherical sky image onto a unit sphere to obtain the unit sphere sky image specifically includes: Generating uniformly distributed pixel points on the unit sphere according to the Fibonacci grid method; Each pixel point of the unit sphere is matched with the spherical sky image, and a pixel value corresponding to the pixel point is obtained through a bilinear interpolation function to obtain the unit sphere sky image.

7. The method according to claim 1, characterized in that The preprocessing of the meteorological information data and the total solar radiation specifically includes: Performing an anomaly identification operation on the meteorological information data and the total solar radiation based on the DBSCAN clustering algorithm to obtain data anomaly points in the meteorological information data and the total solar radiation; A cubic spline fitting is performed on the data points around the data outlier point according to a preset outlier point interval, so that the spline interpolation replaces the outlier value corresponding to the data outlier point.

8. The method according to claim 1, characterized in that The expression of the clear sky irradiance function is: τ=0.56(e -0.56AM +e -0.09AM )k AM=[(r+c e ) 2 cos 2 (θ)+(2r+1+c e )(1-c e )] 1 / 2 -(r+c e )cosθ Among them, E0 represents the solar radiation reaching the upper atmosphere before the solar radiation is attenuated by the atmosphere, and E sc represents the solar radiation constant, N represents the serial number of the current day in the year, k represents the adjustment parameter, AM represents the ratio of the path length of solar radiation passing through the atmosphere to the shortest path length, r represents the radius of the earth, and c e represents the cosine value of the solar zenith angle, τ represents the ratio of solar transmitted radiation to incident radiation, E clear Represents the first theoretical total solar radiation.

9. The method according to claim 1, characterized in that The step of constructing a sky image fusion prediction model, inputting the standard meteorological information data, the standard total solar radiation, and the first theoretical total solar radiation into the sky image fusion prediction model, and obtaining predicted total solar radiation and predicted photovoltaic power specifically includes: The standard meteorological information data, the standard total solar radiation and the first theoretical total solar radiation are divided according to a preset ratio to obtain a cloud map training set, a cloud map verification set and a cloud map test set respectively; Constructing a sky image fusion prediction model, inputting the cloud image training set into the sky image fusion prediction model for training to obtain an optimized sky image fusion prediction model, and inputting the cloud image test set into the optimized sky image fusion prediction model to obtain a second theoretical total solar radiation; The difference between the second theoretical total solar radiation and the actual solar radiation in the solar radiation data is used as a regression target, and the second theoretical total solar radiation is compensated to obtain predicted total solar radiation and predicted photovoltaic power.

10. The method according to claim 1, characterized in that The meteorological information data includes temperature, relative humidity, wind speed and direction, zenith angle and azimuth, cloud cover, air pressure, albedo and broadband turbidity.

Images