Photovoltaic ultra-short-term power prediction method based on cloud image multi-feature extraction

Abstract

The application discloses a photovoltaic ultra-short-term power prediction method based on cloud image multi-feature extraction, and belongs to the field of electric power. The prediction method comprises the following steps: acquiring a change image of a cloud cluster, and obtaining a ground-based cloud image sequence; all cloud images in the ground-based cloud image sequence are preprocessed to obtain a preprocessed image sequence; cloud image features are acquired based on the preprocessed image sequence; historical photovoltaic output power data and the cloud image features are subjected to deep learning to acquire a photovoltaic ultra-short-term power prediction model, so that photovoltaic ultra-short-term power prediction can be performed. The problem of insufficient cloud image feature extraction features in the prior art, which further leads to the problem of excessively low prediction result accuracy, is solved. Through accurate analysis of the features of the cloud image, the CNN and Autoformer methods are combined to construct a photovoltaic ultra-short-term power prediction model, and then photovoltaic ultra-short-term power prediction is performed.

Description

Technical Field

[0001] This application belongs to the field of power, specifically, it relates to a photovoltaic ultra-short-term power prediction method based on cloud map multi-feature extraction. Background Technology

[0002] With the rapid growth of the global economy, energy crises and environmental problems have become increasingly prominent. China's proposed dual-carbon goals are driving energy transformation towards cleaner and lower-carbon development. Against this backdrop, vigorously developing clean and low-carbon renewable energy is a powerful way to help achieve "carbon peaking" and "carbon neutrality." By the end of 2024, China's installed renewable energy capacity reached 1.73 billion kW, of which solar power capacity reached 770 million kW, making it the world's largest solar energy utilization base. To provide accurate photovoltaic output data for real-time dispatch and control of the power system, ultra-short-term forecasts are used to predict photovoltaic output for the next 15 minutes to one hour. The shorter forecast timescale not only results in smaller prediction errors but also enables the system to respond quickly to sudden changes in photovoltaic power generation, enhancing system stability.

[0003] Traditional photovoltaic (PV) forecasting methods, such as physical and statistical methods, often fail to meet grid operation requirements when sudden climate changes cause significant fluctuations in PV output because they cannot adapt their forecasting strategies in a timely manner. However, with the development of computer science and technology, artificial intelligence algorithms, with their excellent self-learning and self-organizing capabilities, are widely used in the field of PV power forecasting.

[0004] The relative position of clouds to the sun is the most direct meteorological factor causing random fluctuations in photovoltaic power generation. Therefore, considering cloud features is a key method to improve the accuracy of ultra-short-term forecasts. In recent years, some technologies have begun to process acquired cloud images as input data for prediction models. Cloud mask detection and image fusion reconstruction were performed using cloud images acquired by the Australian geostationary orbit Himawari-8 / 9 satellites, improving the accuracy and efficiency of cloud detection from satellite remote sensing images and reducing the number of model parameters required for prediction. Different spectral band combinations were explored to delineate cloud and snow regions in satellite images. However, satellite cloud images have low spatiotemporal resolution and cannot precisely capture the edge features of small cloud clusters.

[0005] The advantage of ground-based cloud images lies in the fact that the images captured by the lens reflect the actual light received by the photovoltaic panels, and their acquisition is more convenient and flexible. Utilizing cloud images requires initial image restoration to reduce the impact of image recognition errors on prediction results. A camera calibration method is used to correct distortion in the cloud images, and threshold segmentation of the B-channel components of the RGB color channels is performed to improve the cloud cover detection algorithm. A two-stage cloud velocity calculation model is proposed, using cloud movement speed as an input feature. Dense optical flow algorithms are used to extract cloud image motion direction information. A novel deep model, CloudSwinNet, is proposed by combining the local feature extraction capabilities of CNNs with the global learning capabilities of Transformers. The method of fusing ground-based cloud images with multimodal meteorological factors significantly improves prediction accuracy under variable weather conditions.

[0006] Existing cloud image feature extraction methods can reduce prediction errors and ensure the safe and stable operation of the power grid when clouds affect photovoltaic power output. However, existing methods are not sufficiently effective in extracting features from cloud image information, and still suffer from high prediction errors in scenarios with drastic weather changes. Therefore, this paper first preprocesses the captured ground-based cloud images to extract static features. Then, a recursive full-pair field optical flow algorithm is proposed to extract dynamic multi-features of the clouds. Finally, combining the advantages of CNN in local feature extraction and Autoformer in capturing the periodicity and trend of sequences, a combined photovoltaic power prediction model is constructed. Through case studies on a dataset categorized by weather type, the effectiveness of the proposed static and dynamic multi-feature fusion method for cloud images in reducing prediction errors in scenarios with frequent fluctuations in photovoltaic power output is verified. Summary of the Invention

[0007] To address the problem of insufficient feature extraction from cloud image information in existing technologies, leading to excessively high prediction errors, this application discloses the following technical solution:

[0008] A photovoltaic ultra-short-term power prediction method based on cloud map multi-feature extraction, the prediction method includes:

[0009] Acquire images of cloud changes and obtain a sequence of ground-based cloud images;

[0010] All cloud images in the ground cloud image sequence are preprocessed to obtain a preprocessed image sequence;

[0011] Based on the preprocessed image sequence, cloud map features are obtained;

[0012] Deep learning is performed on historical photovoltaic output power data and cloud map features to obtain a photovoltaic ultra-short-term power prediction model for photovoltaic ultra-short-term power prediction.

[0013] Optionally, acquiring images of cloud changes and obtaining a sequence of ground-based cloud images includes:

[0014] Acquire historical video footage of the cloud cluster to obtain images of its changing patterns;

[0015] All frame images of the cloud changes are acquired in chronological order to obtain cloud image frame images;

[0016] All cloud image frames are arranged in chronological order to obtain a ground cloud image sequence.

[0017] Optionally, the preprocessing of all cloud images in the ground-based cloud image sequence to obtain a preprocessed image sequence includes:

[0018] All cloud maps in the foundation cloud map image are acquired and cloud map distortion correction processing is performed to obtain the distortion-reduced image;

[0019] The distortion-reduced image is subjected to intraday difference reduction processing to obtain a difference-suppressed image;

[0020] The difference-suppressed image is subjected to strong light interference suppression processing to obtain a preprocessed image, which is then arranged in chronological order to obtain a preprocessed image sequence.

[0021] Optionally, the step of acquiring all cloud images in the foundation cloud image and performing cloud image distortion correction processing to obtain a distortion-reduced image includes:

[0022] Obtain the distortion center and distortion radius of the fisheye image during the acquisition process of the ground cloud image;

[0023] Based on the distortion center and distortion radius, the fisheye image is projected into the hemispherical model, and the position of each pixel in the fisheye image on the hemispherical surface is obtained on the hemispherical model.

[0024] The coordinate points of the hemispherical model are converted into hemispherical latitude and longitude coordinates, and the hemispherical latitude and longitude coordinates are converted into planar two-dimensional coordinates to obtain a planar two-dimensional target image;

[0025] The two-dimensional target image is subjected to a Gamma transformation to obtain the distortion-reduced image. The Gamma transformation equation is:

[0026] ,

[0027] Among them, I out I represents the image after distortion correction; I represents the two-dimensional target image in a plane. This represents the Gamma coefficient, which ranges from 0 to 1.

[0028] Optionally, performing intraday difference reduction processing on the distortion-reduced image to obtain a difference-suppressed image includes:

[0029] The distortion-reduced image is then subjected to histogram equalization to obtain a difference-suppressed image. The histogram equalization equation is as follows:

[0030] ,

[0031] Among them, s k r represents the result after the k-th layer equalization process; k r represents the pixel value of the k-th layer; j Represents the gray level of the j-th layer; n j T(r) represents the total number of pixels in the j-th layer; n represents the total number of pixels in the image; r represents the grayscale levels of the image; k represents the number of layers in the image; T(r) k ) indicates a pixel transformation function; P r (r) j ) represents the probability density of the gray level of the j-th layer; m represents the number of gray levels; j represents the layer index; k represents the total number of layer indices.

[0032] Optionally, the step of performing strong light interference suppression processing on the difference-suppressed image to obtain a preprocessed image and arranging it in chronological order to obtain a preprocessed image sequence includes:

[0033] To obtain the sun's position in the horizontal coordinate system, the sun's angle is determined by the following equation:

[0034]

[0035] in, Indicates the solar zenith angle; Indicates the solar azimuth angle; Indicates the solar declination angle; Indicates the geographical latitude and longitude of the imaging device; Indicates solar hour angle;

[0036] The image containing a clear sun outline is labeled with difference suppression image, and the pixel arrangement coordinates of the sun's center within the image are obtained. The equation for determining the pixel arrangement coordinates of the sun's center is:

[0037]

[0038] Where c represents the pixel coordinates of the sun's center within the image; Z represents the polynomial vector of the sun's zenith angle, denoted as... A represents the polynomial vector of the solar azimuth angle, denoted as: ;p ijThis represents the optimal coefficient obtained by fitting the sample data;

[0039] To obtain the adjustable red-green difference features within the difference-suppressed image, the equation for the adjustable red-green difference features is:

[0040] ,

[0041] Where ARGD represents the adjustable red-green difference feature; g represents the red channel weight coefficient; R represents the red channel gray value of the image pixel; and G represents the green channel gray value of the image pixel.

[0042] Adjust the red channel weight coefficients to enhance the pixel differences between clouds and space in the bright areas to obtain an image with enhanced differences;

[0043] The solar intensity and saturation differences of the enhanced image are obtained, and it is determined whether there is strong light interference in the image. The solar intensity equation is:

[0044] ,

[0045] Where SI represents the solar intensity value; I a This represents the grayscale value of the a-th pixel; 'a' represents the pixel grayscale index.

[0046] The equation for the difference in saturation is:

[0047] ,

[0048] Where SD represents the saturation difference value; n w n represents the number of pixels in the image after enhancing the difference; n1 represents the number of pixels in the solar region; S c S represents the pixel saturation of the c-th pixel in the image after enhancement; d This represents the pixel saturation of the d-th pixel in the solar region.

[0049] Optionally, obtaining cloud map features based on the preprocessed image sequence includes:

[0050] In the process of obtaining the preprocessed image sequence, all preprocessed images are converted to grayscale images based on the grayscale thresholding method to obtain binary images;

[0051] The probability of a pixel being located in the background region and the target region in the binary image is obtained. The probability equation for a pixel being located in the background region is as follows:

[0052] ,

[0053] in, The mean gray value of the background region in the binary image is represented by n; T represents the pixel gray value threshold; n represents the mean gray value of the background region in the binary image. eP represents the number of pixels with a grayscale value of e; e represents the pixel's grayscale value; e P represents the probability of a pixel with a grayscale value of e appearing; b This represents the probability that a pixel appears in the background of a binary image;

[0054] The probability equation for a pixel being located in the target region is:

[0055] ,

[0056] in, P represents the mean gray level of the target region in the binary image; L represents the gray level threshold of the cloud image after grayscale conversion; o This represents the probability that a pixel appears in the target portion of a binary image;

[0057] The inter-class variances of the background and target portions of the ground-based cloud image are obtained, and the inter-class variance equation is shown below:

[0058] ,

[0059] Where H represents the segmentation threshold; D(T) represents the inter-class variance;

[0060] Obtain the segmentation threshold corresponding to the maximum inter-class variance, divide the binary image, and calculate the cloud coverage.

[0061] Obtain the optical flow of all adjacent preprocessed images in the preprocessed image sequence. The constraint equation for the optical flow is:

[0062] ,

[0063] Where I(x, y, t) represents the intensity value of pixel (x, y) at time t; dt represents the time difference between adjacent preprocessed images; dx represents the displacement in the x-axis direction between adjacent preprocessed images; and dy represents the displacement in the y-axis direction between adjacent preprocessed images.

[0064] Based on the constraint equation of optical flow, the optical flow equation is obtained, which is:

[0065] ,

[0066] Among them, I x I y and I t Let v represent the derivatives of I(x, y, t) with respect to x, y, and t, respectively; x and v y These are the velocity components along the x-axis and y-axis, respectively;

[0067] Deep neural networks are used to learn all optical flows, and optical flows are acquired from all consecutive two-frame images to obtain dynamic features of the cloud clusters.

[0068] Optionally, the step of performing deep learning on historical photovoltaic output power data and cloud map features to obtain a photovoltaic ultra-short-term power prediction model for ultra-short-term power prediction includes:

[0069] Based on a convolutional neural network, local features are extracted from the photovoltaic feature matrix of the cloud map features to obtain local features;

[0070] The feature trends and periodic components in the local features are obtained to obtain a local feature change model;

[0071] Based on the local feature change model, local features are predicted to obtain predicted weather information;

[0072] Based on the K-means clustering method, the relationship between historical photovoltaic output power data and weather is obtained, and based on the predicted weather information, ultra-short-term photovoltaic power prediction is performed.

[0073] Optionally, the step of extracting local features from the photovoltaic feature matrix of the cloud map features based on a convolutional neural network to obtain local features includes:

[0074] The cloud map features are input into the input layer of the convolutional neural network, and the data is sent to convolutional layer 1.

[0075] After processing by convolutional layer 1, the processed data is sent to the pooling layer for data pooling.

[0076] The pooled data is sent to convolutional layer 2, and the processed data is sent to the fully connected layer to output local features.

[0077] The mathematical model for a convolutional layer is:

[0078] ,

[0079] Where O(x, y) represents the element with coordinates (x, y) in the output matrix O; I(x, y, z) and K(x, y, z) represent the elements with coordinates (x, y, z) in the input matrix I and the convolution kernel K, respectively; S is the stride of the convolution kernel; b is the bias term; f() represents the activation function; l, w, and d represent the coordinate values ​​on the x-axis, y-axis, and z-axis of the coordinate (x, y, z), respectively; and D, L', and W' are the maximum values ​​on the x-axis, y-axis, and z-axis, respectively.

[0080] Optionally, obtaining the feature trend and periodic components in the local features to obtain a local feature change model includes:

[0081] The local features are subjected to sequence decomposition to obtain a time series. The decomposition equation is as follows:

[0082] ,

[0083] ,

[0084] Among them, X t Indicates seasonal component; X s This represents the trend component extracted from the sequence; AvgPool() represents average pooling; Padding() represents the padding algorithm around the boundaries of the input data; X represents the input data;

[0085] The autocorrelation coefficient of the time series is obtained to determine the periodic dependence. The autocorrelation coefficient equation is as follows:

[0086] ,

[0087] in, Representing discrete time series X t and lagged time series The similarity of delays between them; Q represents the number of times the delayed operation is performed; u represents the delay between discrete time series.

[0088] The beneficial effects of this application include:

[0089] 1. Improved processing accuracy of cloud imagery. The technical solution of this application performs multi-step processing on the cloud imagery acquired from the ground to ensure that the pre-processed imagery improves data processing efficiency in terms of information carrying, processing, and data handling. This ensures the rationality of image processing and guarantees that the processing accuracy of the cloud imagery can be significantly improved in subsequent processing and applications.

[0090] 2. The accuracy of information in the cloud image is guaranteed. In the technical solution of this application, in the acquisition of information from the cloud image, on the one hand, the processing accuracy of the image can be guaranteed during the processing of the cloud image, thereby ensuring the basic accuracy of information acquisition; on the other hand, based on the image sequence, the continuity of information in the cloud image is obtained, and information is extracted from it, thereby ensuring that the accuracy of the obtained cloud image sequence information is improved.

[0091] 3. Improved prediction accuracy for photovoltaic systems. In the technical solution of this application, various types of information that can cause changes in photovoltaic processing data are extracted during the processing of all cloud images. The extracted information can then be correlated with the processing status of the photovoltaic system, allowing for the application of the specific processing status of the data, thereby significantly improving the prediction accuracy for photovoltaic systems. Attached Figure Description

[0092] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the embodiments of this application or the prior art will be briefly introduced below. Obviously, the following description is only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. The drawings are used to provide a further understanding of this disclosure and constitute a part of the specification. They are used together with the following detailed description to explain this disclosure, but do not constitute a limitation of this disclosure. In the drawings:

[0093] Figure 1 A flowchart of a photovoltaic ultra-short-term power prediction method based on cloud map multi-feature extraction provided in this application embodiment;

[0094] Figure 2 A flowchart of the latitude and longitude correction algorithm for the photovoltaic ultra-short-term power prediction method based on cloud map multi-feature extraction provided in the embodiments of this application;

[0095] Figure 3 A cloud cluster identification effect diagram of the photovoltaic ultra-short-term power prediction method based on cloud map multi-feature extraction provided in the embodiments of this application;

[0096] Figure 4 A cloud movement tracking effect diagram of the photovoltaic ultra-short-term power prediction method based on cloud map multi-feature extraction provided in the embodiments of this application;

[0097] Figure 5 A convolutional neural network structure diagram for the photovoltaic ultra-short-term power prediction method based on cloud map multi-feature extraction provided in the embodiments of this application;

[0098] Figure 6 A comparison of prediction results of different models under clear weather conditions for the photovoltaic ultra-short-term power prediction method based on cloud map multi-feature extraction provided in the embodiments of this application;

[0099] Figure 7 Error distribution of prediction results of different models under clear weather conditions for the photovoltaic ultra-short-term power prediction method based on cloud map multi-feature extraction provided in the embodiments of this application;

[0100] Figure 8A comparison of prediction results of different models under the clear-to-cloudy condition for the photovoltaic ultra-short-term power prediction method based on cloud map multi-feature extraction provided in the embodiments of this application;

[0101] Figure 9 Error distribution of prediction results of different models under sunny-to-cloudy conditions for the photovoltaic ultra-short-term power prediction method based on cloud map multi-feature extraction provided in the embodiments of this application;

[0102] Figure 10 A comparison of prediction results of different models under cloudy conditions for the photovoltaic ultra-short-term power prediction method based on cloud map multi-feature extraction provided in the embodiments of this application;

[0103] Figure 11 Error distribution of prediction results of different models under cloudy conditions for the photovoltaic ultra-short-term power prediction method based on cloud map multi-feature extraction provided in the embodiments of this application. Detailed Implementation

[0104] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application. Furthermore, in the embodiments of this application, "first," "second," etc., are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0105] While existing technologies have begun to predict ultra-short-term photovoltaic power based on cloud map features, these feature extraction methods suffer from insufficient feature extraction depth and low information processing accuracy. To address these issues, this application discloses a method for predicting ultra-short-term photovoltaic power based on multi-feature extraction from cloud maps, such as... Figure 1 The diagram shown is a flowchart of the photovoltaic ultra-short-term power prediction method based on cloud map multi-feature extraction provided in this application embodiment. Specifically...

[0106] S110. Obtain images of cloud changes and acquire a sequence of ground-based cloud images.

[0107] S120. Preprocess all cloud images in the ground cloud image sequence to obtain a preprocessed image sequence.

[0108] S130. Based on the preprocessed image sequence, obtain cloud map features.

[0109] S140. Perform deep learning on historical photovoltaic output power data and cloud map features to obtain a photovoltaic ultra-short-term power prediction model for photovoltaic ultra-short-term power prediction.

[0110] The following will provide a detailed explanation of all the steps above:

[0111] As described in step S110, the purpose of this step is that, during the operation of the photovoltaic system, both from the perspective of ease of information acquisition and the subsequent acquisition of dynamic characteristics, it is necessary to acquire and process the obtained information. Specifically:

[0112] S111. Acquire historical video footage of the cloud cluster to obtain images of the cloud cluster's changes.

[0113] S112. Obtain all frame images of the cloud changes in chronological order to obtain cloud image frame images.

[0114] S113. Arrange all cloud image frames in chronological order to obtain a ground cloud image sequence.

[0115] The process involves acquiring continuous aerial video footage, decomposing the images within the video, obtaining a time series, and then arranging the images according to the time series to obtain a ground-based cloud image sequence.

[0116] As described in step S120, the purpose of this step is that, in the image processing, the image used is from a wide-angle camera, and the resulting image is also a wide-angle image. Such images will have distortion, therefore, these distortions need to be processed before accurate image feature extraction results can be obtained. Specifically:

[0117] S121. Obtain all cloud maps in the foundation cloud map image and perform cloud map distortion correction processing to obtain the distortion-reduced image.

[0118] The purpose of this step is to process image distortions caused by the camera itself. Specifically:

[0119] S1211. Obtain the distortion center and distortion radius of the fisheye image during the acquisition process of the foundation cloud image.

[0120] S1212. Based on the distortion center and distortion radius, project the fisheye image into the hemispherical model, and obtain the position of each pixel point on the hemispherical surface of the fisheye image.

[0121] S1213. Convert the coordinate points of the hemispherical model into hemispherical latitude and longitude coordinates, and then convert the hemispherical latitude and longitude coordinates into planar two-dimensional coordinates to obtain a planar two-dimensional target image.

[0122] S1214. Perform a Gamma transformation on the two-dimensional target image to obtain the distortion-reduced image. The Gamma transformation equation is:

[0123] ,

[0124] Among them, I out I represents the image after distortion correction; I represents the two-dimensional target image in a plane. This represents the Gamma coefficient, which ranges from 0 to 1.

[0125] Fisheye cameras, used for cloud image capture, are widely employed due to their wide field of view and unique perspective effects. However, this wide-angle characteristic often leads to discrepancies between the size and position of the captured object and the actual object, especially at the image edges where distortion is more severe. Latitude and longitude distortion correction algorithms are widely used because they do not require external calibration. By remapping each pixel in the image, the wide-angle image captured by the fisheye lens is converted into a standard perspective image that more closely resembles human vision, improving the image's visual quality and geometric accuracy.

[0126] The implementation of the latitude and longitude fisheye correction algorithm consists of three main stages: data extraction, coordinate transformation, and correction completion, as shown in the figure. The specific steps are as follows: First, based on the geometric characteristics of the fisheye image, its distortion center and distortion radius are determined; this is the crucial starting point of the entire correction process. Then, a mapping relationship is established between the coordinate points of the fisheye image and the coordinate points of the hemispherical model. The fisheye image is projected onto the hemispherical model to determine the corresponding positions of each pixel in the image on the sphere. Next, the three-dimensional coordinate points on the sphere are converted into latitude and longitude coordinates of the hemispherical surface, and further, the points in three-dimensional space are mapped to two-dimensional coordinate points in the target planar image, completing the fisheye image distortion correction. Through this series of geometric mappings and coordinate transformations, curved lines in the fisheye image can be corrected to straight lines, generating a distortion-free planar image, thus ensuring accurate and effective correction of the fisheye image. Figure 2 The diagram shown is a flowchart of the latitude and longitude correction algorithm for the photovoltaic ultra-short-term power prediction method based on cloud map multi-feature extraction provided in this application embodiment.

[0127] S122. Perform intraday difference elimination processing on the distortion-reduced image to obtain a difference-suppressed image.

[0128] The purpose of this step is that, due to variations in light intensity, weather conditions, and observation angle, the pixel values ​​of ground-based cloud images typically fluctuate significantly over time. Histogram equalization is used to eliminate intraday differences in cloud image pixel values ​​caused by variations in light intensity. Specifically:

[0129] The probability density function of the original image is Pr(r), and the probability density function of the processed image is Ps(s). The two satisfy the following functional relationship:

[0130] (2),

[0131] The transformation function is known to be:

[0132] (3),

[0133] In the formula, ω represents the integral variable. Image processing is often expressed in discrete form. Assuming the original image has a total of n pixels and m gray levels, the result of the k-th layer equalization can be represented as s k As shown in the equation. T(r) k ) is the pixel transformation function, P r (r j ) represents the probability density of the j-th gray level, n j is the number of pixels at the j-th gray level.

[0134] The distortion-reduced image is then subjected to histogram equalization to obtain a difference-suppressed image. The histogram equalization equation is as follows:

[0135] (4),

[0136] Among them, s k r represents the result after the k-th layer equalization process; k r represents the pixel value of the k-th layer; j Represents the gray level of the j-th layer; n j T(r) represents the total number of pixels in the j-th layer; n represents the total number of pixels in the image; r represents the grayscale levels of the image; k represents the number of layers in the image; T(r) k ) indicates a pixel transformation function; P r (r) j ) represents the probability density of the gray level of the j-th layer; m represents the number of gray levels; j represents the layer index; k represents the total number of layer indices.

[0137] S123. Perform strong light interference suppression processing on the difference suppression image to obtain a preprocessed image and arrange it in chronological order to obtain a preprocessed image sequence.

[0138] The purpose of this step is to address the issue of strong light interference in image processing, particularly in ground-sky images where sunlight saturates and leads to false cloud detection. This is achieved by locating the coordinates of the sun's pixels to determine the presence of solar interference and suppress the impact of strong light on cloud detection. Specifically:

[0139] S1231. Obtain the position of the sun in the horizontal coordinate system to obtain the sun angle. The equation for determining the sun angle is:

[0140]

[0141] in, Indicates the solar zenith angle; Indicates the solar azimuth angle; Indicates the solar declination angle; Indicates the geographical latitude and longitude of the imaging device; It represents the solar hour angle.

[0142] S1232. Label the image containing a clear sun outline within the difference suppression image, and obtain the pixel arrangement coordinates of the sun's center within the image. The equation for determining the pixel arrangement coordinates of the sun's center is:

[0143]

[0144] Where c represents the pixel coordinates of the sun's center within the image; Z represents the polynomial vector of the sun's zenith angle, denoted as... A represents the polynomial vector of the solar azimuth angle, denoted as: ;p ij This represents the optimal coefficient obtained by fitting the sample data;

[0145] S1233. Obtain the adjustable red-green difference features within the difference-suppressed image. The equation for the adjustable red-green difference features is:

[0146] ,

[0147] Where ARGD represents the adjustable red-green difference feature; g represents the red channel weight coefficient; R represents the red channel gray value of the image pixel; and G represents the green channel gray value of the image pixel.

[0148] S1234. Adjust the red channel weight coefficient to enhance the pixel difference between clouds and space in the bright light area to obtain an image with enhanced difference.

[0149] S1235. Obtain the difference in solar intensity and saturation of the enhanced image, and determine whether there is strong light interference in the image. The solar intensity equation is:

[0150] ,

[0151] Where SI represents the solar intensity value; I a This represents the grayscale value of the a-th pixel; 'a' represents the pixel grayscale index.

[0152] S1236, the equation for the difference in saturation is:

[0153] ,

[0154] Where SD represents the saturation difference value; n wn represents the number of pixels in the image after enhancing the difference; n1 represents the number of pixels in the solar region; S c S represents the pixel saturation of the c-th pixel in the image after enhancement; d This represents the pixel saturation of the d-th pixel in the solar region.

[0155] The method determines whether there is strong light interference in the image by calculating two parameters: solar intensity (SI) and saturation difference (SD). When SI ≥ 180 or SD ≥ 0.1, the image is determined to have strong light interference. Finally, the pixel feature distortion problem caused by strong light interference is offset by adjusting the weight coefficients.

[0156] As described in step S130, the purpose of this step is to obtain cloud map features based on the preprocessed image after image preprocessing, and then obtain a photovoltaic ultra-short-term power prediction model based on the obtained features. Specifically:

[0157] S131. In the process of obtaining the preprocessed image sequence, all preprocessed images are converted to grayscale images based on the grayscale thresholding method to obtain binary images.

[0158] When dividing the cloud region and the sky region, grayscale thresholding is generally used to convert the grayscale image into a binary image. The cloud coverage parameter can be defined as the ratio of effective pixels to total pixels in the binary image.

[0159] (11)

[0160] In the formula, m and n correspond to the height and width of the image space dimension, respectively. B(k, p) follows a binarization mapping rule: it takes the value 1 when the pixel grayscale value is greater than or equal to the threshold T, and 0 otherwise. The choice of threshold becomes the key to segmentation. The maximum inter-class variance method, which can optimize the threshold selection, can achieve effective target segmentation in different types of images. When the grayscale level of the ground cloud map after grayscale conversion is L=256, and the number of pixels with grayscale value k is n k If there are N pixels, then use P k This represents the probability of a grayscale value of k occurring. After dividing the original cloud image into two types, background and target regions, using a threshold T, the probabilities of the two types occurring are obtained by summing the probabilities of different grayscale values ​​in the regions, respectively: P0. b P o .

[0161] S132. Obtain the probability that a pixel in the binary image is located in the background region and the target region. The probability equation for a pixel being located in the background region is:

[0162] (12),

[0163] in, The mean gray value of the background region in the binary image is represented by n; T represents the pixel gray value threshold; n represents the mean gray value of the background region in the binary image. e P represents the number of pixels with a grayscale value of e; e represents the pixel's grayscale value; e P represents the probability of a pixel with a grayscale value of e appearing; b This represents the probability that a pixel appears in the background of a binary image.

[0164] S133, The probability equation for a pixel being located in the target region is:

[0165] (13)

[0166] in, P represents the mean gray level of the target region in the binary image; L represents the gray level threshold of the cloud image after grayscale conversion; o This represents the probability that a pixel appears in the target portion of a binary image.

[0167] S134. Obtain the inter-class variance of the background and target portions of the ground-based cloud image. The inter-class variance equation is as follows:

[0168] (14)

[0169] Where H represents the segmentation threshold; D(T) represents the inter-class variance.

[0170] Among them, satisfying The segmentation threshold T is the optimal threshold for segmenting the background and target regions. Substituting the optimal threshold into the equation... The cloud cover at that moment was calculated.

[0171] S135. Obtain the segmentation threshold corresponding to the maximum inter-class variance, divide the binary image, and calculate the cloud coverage.

[0172] Among them, such as Figure 3 The image shown is a cloud cluster identification effect diagram of the photovoltaic ultra-short-term power prediction method based on cloud map multi-feature extraction provided in the embodiments of this application.

[0173] S136. Obtain the optical flow of all adjacent preprocessed images in the preprocessed image sequence. The constraint equation for the optical flow is:

[0174] (15)

[0175] Where I(x, y, t) represents the intensity value of pixel (x, y) at time t; dt represents the time difference between adjacent preprocessed images; dx represents the displacement in the x-axis direction of adjacent preprocessed images; and dy represents the displacement in the y-axis direction of adjacent preprocessed images.

[0176] Since cloud dissipation and formation are dynamic processes, it is necessary to extract the dynamic features of cloud clusters in ground-based cloud images. Optical flow refers to the motion characteristics of each pixel in a continuous image sequence, from frame t to frame t+1, consisting of the speed and direction of motion. If a pixel is located at position (x, y) at time t, after time dt, there are displacements of dx and dy between adjacent frames.

[0177] Where I(x, y, t) represents the intensity value of pixel (x, y) at time t. If the displacement during dt is small enough, the Taylor series can be used to expand the intensity at (x, y, t):

[0178] (16)

[0179] In the formula, R n (x, y, t) is a higher-order remainder term, approximately equal to zero and negligible.

[0180] S137. Based on the constraint equation of the optical flow, obtain the optical flow equation, which is:

[0181] (17)

[0182] Among them, I x I y and I t Let v represent the derivatives of I(x, y, t) with respect to x, y, and t, respectively; x and v y These are the velocity components along the x-axis and y-axis, respectively.

[0183] This step is a simplified rewrite of equation (16).

[0184] S138. Perform deep neural network learning on all optical flows and acquire optical flows from all two consecutive frames of images to obtain dynamic features of the cloud cluster.

[0185] To improve the accuracy of optical flow estimation, a Recurrent All-Pairs Field Transforms (RAFT) method based on deep neural networks was employed to extract the motion direction and speed of the cloud. RAFT refines the optical flow estimation through continuous iteration using a recurrent neural network, performing particularly well when dealing with small and fast-moving objects, while maintaining high computational efficiency and good generalization ability. The RAFT network consists of three modules. First, the feature encoder extracts features from two consecutive frames of images using multi-layer convolutions. Next, the correlation calculation module constructs a 4D correlation quantity by calculating the inner product between the feature vectors of the two frames, and performs average pooling operations on the last two dimensions to form feature representations at different scales, constructing a multi-scale correlation pyramid structure. Subsequently, relevant features are extracted through convolution operations for subsequent use. Finally, the optical flow decoder infers the final optical flow estimate using an iterative approach. In this process, RAFT introduces a gated recurrent unit to update the optical flow state; its output is processed through two convolutional layers to generate the optical flow estimate for the current iteration, and all iteration results are progressively superimposed to construct the final optical flow field.

[0186] To visually observe the cloud's motion, sparse motion vectors were drawn using green arrows in a 1024×1024 resolution image, with one arrow drawn every 50 pixels. The direction of each arrow represents the cloud's motion. The length of the arrow is proportional to the cloud's speed; the faster the cloud moves, the longer the arrow. Figure 4 The image shown is a cloud movement tracking effect diagram of the photovoltaic ultra-short-term power prediction method based on cloud map multi-feature extraction provided in the embodiments of this application.

[0187] As described in step S140, the purpose of this step is to achieve ultra-short-term photovoltaic power prediction. Specifically:

[0188] S141. Based on a convolutional neural network, local features are extracted from the photovoltaic feature matrix of the cloud map features to obtain local features.

[0189] The purpose of this step is to leverage the capabilities of Convolutional Neural Networks (CNNs), a major representative of deep learning models, through mechanisms such as local connectivity, weight sharing, and pooling operations. These mechanisms enable efficient feature extraction while significantly reducing the number of network parameters and data dimensionality, thereby improving the computational efficiency and analytical capabilities of CNNs. Specifically:

[0190] S1411. Input the cloud map features into the input layer of the convolutional neural network and send the data to convolutional layer 1.

[0191] like Figure 5The diagram shown is a convolutional neural network structure diagram of the photovoltaic ultra-short-term power prediction method based on cloud map multi-feature extraction provided in the embodiments of this application.

[0192] S1412. After processing by convolutional layer 1, the processed data is sent to the pooling layer for data pooling.

[0193] S1413. The pooled data is sent to convolutional layer 2, and the processed data is sent to the fully connected layer, and local features are output.

[0194] S1414, the mathematical model of the convolutional layer is:

[0195] ,

[0196] Where O(x, y) represents the element with coordinates (x, y) in the output matrix O; I(x, y, z) and K(x, y, z) represent the elements with coordinates (x, y, z) in the input matrix I and the convolution kernel K, respectively; S is the stride of the convolution kernel; b is the bias term; f() represents the activation function; l, w, and d represent the coordinate values ​​on the x-axis, y-axis, and z-axis of the coordinate (x, y, z), respectively; and D, L', and W' are the maximum values ​​on the x-axis, y-axis, and z-axis, respectively.

[0197] The convolutional layer is the core structure of a convolutional neural network, used to extract features at different scales from the input data. If the output of the convolution operation is a... The matrix O is used to extract rich features from the input data through convolutional layers, but these features may contain a large number of similar, redundant, or even noisy features. Pooling layers reduce the spatial dimension of the feature maps, thereby reducing the number of network parameters, which speeds up computation and suppresses overfitting while preserving key features.

[0198] Fully connected layers integrate the local features extracted by convolutional and pooling layers into global features, enabling classification and regression of the input data. Through backpropagation, deep features of the input data are learned, allowing for continuous optimization of the fully connected layer parameters.

[0199] S142. Obtain the feature trend and periodic components in the local features to obtain the local feature change model.

[0200] The purpose of this step is to address the widespread application of the attention mechanism in deep learning since its inception. While the Transformer effectively shortens the correlation distance between sequence elements using attention and boasts excellent parallel computing performance, its self-attention mechanism suffers from quadratic computational complexity, high memory consumption, and performance degradation when processing long sequences. The Autoformer, building upon the Transformer architecture, improves upon sequence decomposition and the self-attention mechanism by proposing a deep decomposition architecture and an autocorrelation mechanism. Specifically:

[0201] S1421. Perform sequence decomposition on the local features to obtain a time series. The decomposition equation is:

[0202] (19)

[0203] (20)

[0204] Among them, X t Indicates seasonal component; X s The trend component is extracted from the sequence; AvgPool() represents average pooling; Padding() represents the padding algorithm around the boundaries of the input data; X represents the input data.

[0205] The Autoformer model constructs a deep decomposition structure by embedding a sequence decomposition mechanism into the encoder and decoder modules. During the prediction phase, the model alternately decomposes and optimizes, gradually extracting trend and periodic components from the latent space to achieve a hierarchical and progressive sequence modeling process. The sequence decomposition unit uses an adjustable moving average mechanism to smooth the periodic fluctuations in the sequence, thereby highlighting the long-term trend. For an input sequence X of length L, it is processed according to equations (19) and (20).

[0206] S1422. Obtain the autocorrelation coefficient of the time series to obtain the periodic dependence. The autocorrelation coefficient equation is:

[0207] (twenty one),

[0208] in, Representing discrete time series X t and lagged time series The similarity of delays between them; Q represents the number of times the delayed operation is performed; u represents the delay between discrete time series.

[0209] Since time series data generally exhibits periodicity, curves in the same region will show certain similarities. Based on this, Autoformer introduces an autocorrelation mechanism to achieve efficient sequence-level connections, thereby improving the model's ability to extract time series information. By calculating the autocorrelation coefficient of the sequence, potential periodic dependency patterns are mined, and combined with time delay aggregation operations, structurally similar subsequences are effectively integrated, thereby optimizing information transmission and improving the accuracy and efficiency of prediction. For the input sequence X, there exist similar subsequences with a period of τ, which are processed according to equation (21).

[0210] S143. Based on the local feature change model, predict the local features to obtain predicted weather information.

[0211] S144. Based on the K-means clustering method, obtain the relationship between historical photovoltaic output power data and weather, and based on the predicted weather information, perform ultra-short-term photovoltaic power prediction.

[0212] Clustering algorithms are a typical unsupervised learning method that analyzes and summarizes the inherent patterns and feature distributions of unlabeled samples. This paper uses the K-means clustering method to classify the weather type for each day, meaning each sample can only be assigned to one cluster. By minimizing the Euclidean distance between a sample and the centroid of its cluster, the data is divided into a predetermined number of K clusters. This method is simple in structure, easy to implement, and has good scalability, making it suitable for processing large-scale data such as photovoltaic output power.

[0213] To more accurately segment historical samples with similar power generation patterns, historical photovoltaic (PV) output power data was used for weather-related classification. Compared to directly using meteorological parameters, this method better reflects the actual characteristics of power changes. By setting the cluster size to 3, the samples were divided into three different types: sunny, sunny turning cloudy, and overcast. Under sunny conditions, the PV power curve was relatively smooth with minimal fluctuations. Under cloudy conditions, PV power fluctuations were more drastic, mainly due to cloud cover reducing the solar irradiance received by the PV array. On overcast days, the overall PV output power was low, with drastic fluctuations and strong randomness, significantly increasing the difficulty of PV power prediction.

[0214] To better illustrate the beneficial effects of this application, actual data comparison and analysis methods were used, specifically:

[0215] 1. Instance Acquisition

[0216] To evaluate the performance of different models in ultra-short-term photovoltaic power forecasting, the experiment used the Stanford University SKIPP'D dataset, capturing data from 06:00 to 20:00 daily at a sampling frequency of 5 minutes, and compared the results under three weather conditions. The prediction results were evaluated using three commonly used error metrics: root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R-squared, R²).

[0217] Model 1: Standard Transformer model architecture.

[0218] Model 2: Removal of ground cloud features in the CNN-Autoformer-AS model.

[0219] Model 3: The ConvLSTM model extracts cloud map features directly from the original cloud map.

[0220] Model 4: The CNN-Autoformer model considers cloud image features and meteorological data.

[0221] 2. Results Analysis

[0222] (1) Sunny weather conditions

[0223] Under sunny weather conditions, the prediction results of all four models were quite satisfactory, such as Figure 6 The figure shows a comparison of prediction results for different models under clear weather conditions using the photovoltaic ultra-short-term power prediction method based on cloud map multi-feature extraction provided in this application embodiment. Under clear weather conditions, meteorological conditions are relatively stable, and the formation and dissipation of clouds have a relatively weak impact on irradiance. The photovoltaic output power exhibits a stable trend, and all models can effectively fit the photovoltaic power output curve. As can be seen from the magnified area in the figure, the model that integrates cloud map features is closer to the true value when dealing with slight power fluctuations, and the prediction curve is smoother and more accurate. This indicates that the extracted cloud map features can effectively improve the accuracy of ultra-short-term photovoltaic power prediction.

[0224] To more accurately evaluate the predictive performance of each model under clear weather conditions, the absolute error between the predicted and actual values ​​is as follows: Figure 7The figure shows the error distribution of prediction results for different models under clear weather conditions for the photovoltaic ultra-short-term power prediction method based on cloud image multi-feature extraction provided in this application embodiment. Under clear weather conditions, the median of the box for CNN-Autoformer is closest to zero, indicating that this model has the smallest prediction error and outperforms other comparative models. From the error distribution, although the ConvLSTM model can capture the spatiotemporal features of cloud images well, its error distribution range is wide, and its overall prediction performance is not as good as the model proposed in this paper. The CNN-Autoformer-AS model, which does not incorporate cloud image features, has the smallest box, indicating that the prediction results using only meteorological data and photovoltaic power data are relatively concentrated and have small fluctuations, failing to capture the slight fluctuations under clear weather conditions.

[0225] (2) Weather conditions from sunny to cloudy

[0226] To further evaluate the applicability of the CNN-Autoformer model under conditions of slight fluctuations in photovoltaic power, the following were selected: Figure 8 The image shown is a comparison of the prediction results of different models under sunny-to-cloudy weather conditions, validating the photovoltaic ultra-short-term power prediction method based on cloud map multi-feature extraction provided in this application embodiment. Under sunny-to-cloudy weather conditions, due to the large number and thickness of clouds in the sky, the intensity of solar radiation attenuates as it penetrates the clouds, resulting in reduced solar irradiance received by the photovoltaic array and consequently lower photovoltaic output power. When photovoltaic output power fluctuates, the prediction value of the CNN-Autoformer model best approximates the actual power change trend, while the CNN-Autoformer-AS model, which relies solely on meteorological and photovoltaic output data without incorporating cloud map information, differs significantly from the actual situation.

[0227] The distribution of absolute errors between model predictions and actual values ​​under sunny to partly cloudy weather conditions is as follows: Figure 9 The figure shows the error distribution of prediction results for different models under the clear-to-cloudy condition of the photovoltaic ultra-short-term power prediction method based on cloud image multi-feature extraction provided in the embodiments of this application. The error distribution of the CNN-Autoformer model is the most concentrated, significantly better than that of the Transformer model. That is, the time series decomposition module and autocorrelation mechanism in the CNN-Autoformer model have stronger capture ability and higher stability in dealing with power abrupt changes. The CNN-Autoformer-AS model, which does not introduce cloud image features, has the most dispersed error distribution and poor prediction performance. The error distribution of the ConvLSTM model is between that of the CNN-Autoformer model and the CNN-Autoformer-AS model, indicating that directly extracting cloud image features through the convolutional structure has a certain effect.

[0228] (3) Cloudy weather conditions

[0229] The models were validated using cloudy weather conditions, which exhibited the greatest variation in meteorological conditions. The prediction results are as follows: Figure 10 The figure shows a comparison of prediction results for different models under cloudy conditions using the photovoltaic ultra-short-term power prediction method based on cloud map multi-feature extraction provided in this application embodiment. Under cloudy weather conditions, although the prediction performance of each model is lower than that under sunny and sunny-to-cloudy weather conditions, the ConvLSTM model and the CNN-Autoformer model can better track the photovoltaic output power, while the Transformer and CNN-Autoformer-AS models show significant lag in the photovoltaic output power fluctuation range and struggle to respond promptly to output changes. The prediction performance of each model only improves as the photovoltaic power curve gradually stabilizes. Although the prediction accuracy of CNN-Autoformer decreases under abrupt weather conditions, it still maintains good trend fitting ability and tracking of the true value overall.

[0230] Error distribution of each model under cloudy weather conditions is as follows Figure 11 As shown, this diagram illustrates the error distribution of prediction results for different models under cloudy conditions using the photovoltaic ultra-short-term power prediction method based on cloud map multi-feature extraction provided in this application embodiment. The CNN-Autoformer model exhibits a more concentrated error distribution, with no obvious outliers outside the normal distribution curve. Although the Transformer model also incorporates cloud map features, its limitations in modeling complex temporal structures make it difficult to accurately capture the changing trends of photovoltaic power, resulting in the widest error distribution range. Comparing the error distributions of the CNN-Autoformer-AS model and the ConvLSTM model, the ConvLSTM model shows a more concentrated error distribution and improved prediction performance. This indicates that when photovoltaic power fluctuates significantly, both direct feature extraction from cloud maps and the use of cloud map features extracted in this paper can effectively improve prediction accuracy, verifying the importance of cloud map features under cloudy weather conditions.

[0231] The beneficial effects of this application include:

[0232] 1. Improved processing accuracy of cloud imagery. The technical solution of this application performs multi-step processing on the cloud imagery acquired from the ground to ensure that the pre-processed imagery improves data processing efficiency in terms of information carrying, processing, and data handling. This ensures the rationality of image processing and guarantees that the processing accuracy of the cloud imagery can be significantly improved in subsequent processing and applications.

[0233] 2. The accuracy of information in the cloud image is guaranteed. In the technical solution of this application, in the acquisition of information from the cloud image, on the one hand, the processing accuracy of the image can be guaranteed during the processing of the cloud image, thereby ensuring the basic accuracy of information acquisition; on the other hand, based on the image sequence, the continuity of information in the cloud image is obtained, and information is extracted from it, thereby ensuring that the accuracy of the obtained cloud image sequence information is improved.

[0234] 3. Improved prediction accuracy for photovoltaic systems. In the technical solution of this application, various types of information that can cause changes in photovoltaic processing data are extracted during the processing of all cloud images. The extracted information can then be correlated with the processing status of the photovoltaic system, allowing for the application of the specific processing status of the data, thereby significantly improving the prediction accuracy for photovoltaic systems.

[0235] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to computer program instructions. The aforementioned computer program can be stored in a non-volatile storage medium, and when executed, it performs the steps of the above method embodiments. Alternatively, if the integrated unit of the present invention is implemented as a software functional module and sold or used as an independent product, it can also be stored in a non-volatile storage medium. Based on this understanding, the technical solution of the embodiments of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a non-volatile storage medium and includes several instructions to cause an electronic device (which may be a personal computer, server, network device, etc.) to execute all or part of the methods described in the various embodiments of the present invention.

[0236] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A photovoltaic ultra-short-term power prediction method based on cloud map multi-feature extraction, characterized in that, The prediction method includes: Acquire images of cloud changes and obtain a sequence of ground-based cloud images; All cloud images in the ground cloud image sequence are preprocessed to obtain a preprocessed image sequence; Based on the preprocessed image sequence, cloud map features are obtained; Deep learning is performed on historical photovoltaic output power data and cloud map features to obtain a photovoltaic ultra-short-term power prediction model for photovoltaic ultra-short-term power prediction. The step of obtaining cloud map features based on the preprocessed image sequence includes: In the process of obtaining the preprocessed image sequence, all preprocessed images are converted to grayscale images based on the grayscale thresholding method to obtain binary images; The probability of a pixel being located in the background region and the target region in the binary image is obtained. The probability equation for a pixel being located in the background region is as follows: ， in, The mean gray value of the background region in the binary image is represented by n; T represents the pixel gray value threshold; n represents the mean gray value of the background region in the binary image. e P represents the number of pixels with a grayscale value of e; e represents the pixel's grayscale value; e P represents the probability of a pixel with a grayscale value of e appearing; b This represents the probability that a pixel appears in the background of a binary image; The probability equation for a pixel being located in the target region is: ， in, P represents the mean gray level of the target region in the binary image; L represents the gray level threshold of the cloud image after grayscale conversion; o This represents the probability that a pixel appears in the target portion of a binary image; The inter-class variances of the background and target portions of the ground-based cloud image are obtained, and the inter-class variance equation is shown below: ， Where H represents the segmentation threshold; D(H) represents the inter-class variance; Obtain the segmentation threshold corresponding to the maximum inter-class variance, divide the binary image, and calculate the cloud coverage. Obtain the optical flow of all adjacent preprocessed images in the preprocessed image sequence. The constraint equation for the optical flow is: ， Where I(x, y, t) represents the intensity value of pixel (x, y) at time t; dt represents the time difference between adjacent preprocessed images; dx represents the displacement in the x-axis direction between adjacent preprocessed images; and dy represents the displacement in the y-axis direction between adjacent preprocessed images. Based on the constraint equation of optical flow, the optical flow equation is obtained, which is: ， Among them, I x I y and I t Let v represent the derivatives of I(x, y, t) with respect to x, y, and t, respectively; x and v y These are the velocity components along the x-axis and y-axis, respectively; Deep neural networks are used to learn all optical flows, and optical flows are acquired from all consecutive two-frame images to obtain dynamic features of the cloud clusters.

2. The photovoltaic ultra-short-term power prediction method based on cloud map multi-feature extraction according to claim 1, characterized in that, The acquisition of cloud change images and the obtaining of ground-based cloud image sequences include: Acquire historical video footage of the cloud cluster to obtain images of its changing patterns; All frame images of the cloud changes are acquired in chronological order to obtain cloud image frame images; All cloud image frames are arranged in chronological order to obtain a ground cloud image sequence.

3. The photovoltaic ultra-short-term power prediction method based on cloud map multi-feature extraction according to claim 1, characterized in that, The preprocessing of all cloud images in the ground cloud image sequence to obtain a preprocessed image sequence includes: All cloud maps in the foundation cloud map image are acquired and cloud map distortion correction processing is performed to obtain the distortion-reduced image; The distortion-reduced image is subjected to intraday difference reduction processing to obtain a difference-suppressed image; The difference-suppressed image is subjected to strong light interference suppression processing to obtain a preprocessed image, which is then arranged in chronological order to obtain a preprocessed image sequence.

4. The photovoltaic ultra-short-term power prediction method based on cloud map multi-feature extraction according to claim 3, characterized in that, The process of acquiring all cloud images in the foundation cloud image and performing cloud image distortion correction processing to obtain a distortion-free image includes: Obtain the distortion center and distortion radius of the fisheye image during the acquisition process of the ground cloud image; Based on the distortion center and distortion radius, the fisheye image is projected into the hemispherical model, and the position of each pixel in the fisheye image on the hemispherical surface is obtained on the hemispherical model. The coordinate points of the hemispherical model are converted into hemispherical latitude and longitude coordinates, and the hemispherical latitude and longitude coordinates are converted into planar two-dimensional coordinates to obtain a planar two-dimensional target image; The two-dimensional target image is subjected to a Gamma transformation to obtain the distortion-reduced image. The Gamma transformation equation is: ， Among them, I out I represents the image after distortion correction; I represents the two-dimensional target image in a plane. This represents the Gamma coefficient, which ranges from 0 to 1.

5. The photovoltaic ultra-short-term power prediction method based on cloud map multi-feature extraction according to claim 3, characterized in that, The process of performing intraday difference reduction processing on the distortion-reduced image to obtain a difference-suppressed image includes: The distortion-reduced image is then subjected to histogram equalization to obtain a difference-suppressed image. The histogram equalization equation is as follows: ， Among them, s k r represents the result after the k-th layer equalization process; k r represents the pixel value of the k-th layer; j Represents the gray level of the j-th layer; n j T(r) represents the total number of pixels in the j-th layer; n represents the total number of pixels in the image; r represents the grayscale levels of the image; k represents the number of layers in the image; T(r) k ) indicates a pixel transformation function; P r (r) j ) represents the probability density of the gray level of the j-th layer; m represents the number of gray levels; j represents the layer index; k represents the total number of layer indices.

6. The photovoltaic ultra-short-term power prediction method based on cloud map multi-feature extraction according to claim 3, characterized in that, The process of performing strong light interference suppression processing on the difference-suppressed image to obtain a preprocessed image, and arranging it in chronological order to obtain a preprocessed image sequence, includes: To determine the sun's position in the horizontal coordinate system, we need to obtain the sun's angle. The equation for determining the sun's angle is: ， in, Indicates the solar zenith angle; Indicates the solar azimuth angle; Indicates the solar declination angle; Indicates the geographical latitude and longitude of the imaging device; Indicates solar hour angle; The image containing a clear sun outline is labeled with difference suppression image, and the pixel arrangement coordinates of the sun's center within the image are obtained. The equation for determining the pixel arrangement coordinates of the sun's center is: ， Where c represents the pixel coordinates of the sun's center within the image; Z represents the polynomial vector of the sun's zenith angle, denoted as... A represents the polynomial vector of the solar azimuth angle, denoted as: ;p ij This represents the optimal coefficient obtained by fitting the sample data; To obtain the adjustable red-green difference features within the difference-suppressed image, the equation for the adjustable red-green difference features is: ， Where ARGD represents the adjustable red-green difference feature; g represents the red channel weight coefficient; R represents the red channel gray value of the image pixel; and G represents the green channel gray value of the image pixel. Adjust the red channel weight coefficients to enhance the pixel differences between clouds and space in the bright areas to obtain an image with enhanced differences; The solar intensity and saturation differences of the enhanced image are obtained, and it is determined whether there is strong light interference in the image. The solar intensity equation is: ， Where SI represents the solar intensity value; I a This represents the grayscale value of the a-th pixel; a represents the pixel grayscale index; b represents the total number of pixel grayscale indexes. The equation for the difference in saturation is: ， Where SD represents the saturation difference value; n w n represents the number of pixels in the image after enhancing the difference; n1 represents the number of pixels in the solar region; S c S represents the pixel saturation of the c-th pixel in the image after enhancement; d This represents the pixel saturation of the d-th pixel in the solar region.

7. The photovoltaic ultra-short-term power prediction method based on cloud map multi-feature extraction according to claim 1, characterized in that, The process of performing deep learning on historical photovoltaic output power data and cloud map features to obtain a photovoltaic ultra-short-term power prediction model for ultra-short-term power prediction includes: Based on a convolutional neural network, local features are extracted from the photovoltaic feature matrix of the cloud map features to obtain local features; The feature trends and periodic components in the local features are obtained to obtain a local feature change model; Based on the local feature change model, local features are predicted to obtain predicted weather information; Based on the K-means clustering method, the relationship between historical photovoltaic output power data and weather is obtained, and based on the predicted weather information, ultra-short-term photovoltaic power prediction is performed.

8. The photovoltaic ultra-short-term power prediction method based on cloud map multi-feature extraction according to claim 7, characterized in that, The method of extracting local features from the photovoltaic feature matrix of the cloud map features based on a convolutional neural network to obtain local features includes: The cloud map features are input into the input layer of the convolutional neural network, and the data is sent to convolutional layer 1. After processing by convolutional layer 1, the processed data is sent to the pooling layer for data pooling. The pooled data is sent to convolutional layer 2, and the processed data is sent to the fully connected layer to output local features. The mathematical model for a convolutional layer is: ， Where O(x, y) represents the element with coordinates (x, y) in the output matrix O; I(x, y, z) and K(x, y, z) represent the elements with coordinates (x, y, z) in the input matrix I and the convolution kernel K, respectively; S is the stride of the convolution kernel; b is the bias term; f() represents the activation function; l, w, and d represent the coordinate values ​​on the x-axis, y-axis, and z-axis of the coordinate (x, y, z), respectively; and D, L', and W' are the maximum values ​​on the x-axis, y-axis, and z-axis, respectively.

9. The photovoltaic ultra-short-term power prediction method based on cloud map multi-feature extraction according to claim 7, characterized in that, The step of obtaining the feature trend and periodic components in the local features to obtain a local feature change model includes: The local features are subjected to sequence decomposition to obtain a time series. The decomposition equation is as follows: ， ， Among them, X t Indicates seasonal component; X s This represents the trend component extracted from the sequence; AvgPool() represents average pooling; Padding() represents the padding algorithm around the boundaries of the input data; X represents the input data; The autocorrelation coefficient of the time series is obtained to determine the periodic dependence. The autocorrelation coefficient equation is as follows: ， in, Representing discrete time series X t and lagged time series The similarity of delays between them; Q represents the number of times the delayed operation is performed; u represents the delay between discrete time series.

Images