Distributed photovoltaic short-term power generation prediction method, system, device and medium

Abstract

This invention relates to the field of photovoltaic power generation technology, and provides a method, system, device, and medium for short-term power generation prediction of distributed photovoltaic power. The method includes acquiring ground-based cloud map time-series data, historical power time-series data, and clustered distributed photovoltaic information of the target photovoltaic power station; extracting features from the ground-based cloud map time-series data to obtain cloud map feature time-series data; performing power prediction based on a first prediction model using the cloud map feature time-series data and historical power time-series data to obtain a first prediction result; performing power prediction based on a second prediction model using spatiotemporal coupling feature time-series data obtained from clustered distributed photovoltaic information, cloud coverage rate time-series data, and cloud change rate time-series data to obtain a second prediction result; and weightedly fusing the first and second prediction results to obtain the target prediction result. This invention can significantly improve the reliability, stability, and accuracy of short-term power prediction results under complex and abrupt weather conditions such as cloudy or rainy weather.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power generation technology, and in particular to a method, system, equipment and medium for predicting short-term power generation of distributed photovoltaic systems. Background Technology

[0002] With the large-scale grid connection of distributed photovoltaic (PV) power, its short-term power forecast results have become an important basis for grid optimization and dispatch, providing crucial support for the safe and stable operation of the grid. However, the actual output of distributed PV is significantly affected by factors such as sudden weather changes and dynamic cloud cover, leading to drastic short-term power fluctuations and increasing the difficulty of accurate short-term power forecasting.

[0003] While existing technologies recognize the impact of cloud information on the accuracy of short-term power forecasting, they only use shallow, static features of cloud images and do not consider the role of deep image features and dynamic trends of clouds. They fail to fully utilize cloud image information and do not consider the correlation of output between distributed photovoltaic power stations within a cluster. Therefore, they cannot effectively guarantee the reliability and accuracy of forecast results under complex weather scenarios such as cloudy or rainy weather. Summary of the Invention

[0004] The purpose of this invention is to provide a method for short-term power generation prediction of distributed photovoltaic power. Based on a multi-level cloud map feature analysis mechanism that extracts edge texture features and dynamic change features from ground-based cloud maps, combined with a time-coupled feature extraction and analysis mechanism between photovoltaic power stations based on dynamic change features, this method effectively integrates and fully utilizes multi-modal data. It can significantly improve the reliability, stability, and accuracy of short-term prediction results under complex weather scenarios such as cloudy and rainy weather, and provide high-precision and robust scheduling analysis support for the safe and stable operation of the power grid with a high proportion of renewable energy and the efficient consumption of renewable energy.

[0005] To achieve the above objectives, it is necessary to provide a method, system, computer equipment, and storage medium for predicting short-term power generation from distributed photovoltaic systems.

[0006] In a first aspect, embodiments of the present invention provide a method for predicting short-term power generation from distributed photovoltaic systems, the method comprising the following steps:

[0007] Acquire ground-based cloud map time-series data, historical power time-series data, and clustered distributed photovoltaic information of the target photovoltaic power station;

[0008] Feature extraction is performed on the ground-based cloud map time-series data to obtain cloud map feature time-series data; the cloud map feature time-series data includes edge feature time-series data, texture feature time-series data, cloud coverage time-series data, and cloud change rate time-series data.

[0009] Based on the cloud map feature time series data and the historical power time series data, power prediction is performed based on a pre-built first prediction model to obtain a first prediction result;

[0010] Based on the clustered distributed photovoltaic information, the cloud coverage time series data, and the cloud change rate time series data, spatiotemporal coupling characteristic time series data are obtained, and based on the spatiotemporal coupling characteristic time series data, power prediction is performed based on the pre-built second prediction model to obtain the second prediction result;

[0011] The first prediction result and the second prediction result are weighted and fused to obtain the target prediction result.

[0012] Furthermore, the step of extracting features from the ground-based cloud map time-series data to obtain cloud map feature time-series data includes:

[0013] Each foundation cloud map in the time series data of the foundation cloud map is preprocessed to obtain the time series data of the foundation cloud map to be analyzed.

[0014] Based on the dynamic scale oriented gradient histogram technique, edge features are extracted from the time series data of the ground cloud map to be analyzed to obtain the edge feature time series data; the gradient statistical scale in the dynamic scale oriented gradient histogram technique is adaptively determined based on gradient energy.

[0015] Based on the multi-directional weighted gray-level co-occurrence matrix technology, texture features are extracted from the time-series data of the ground cloud map to be analyzed to obtain the texture feature time-series data; the weights of the gray-level co-occurrence matrix in each direction of the multi-directional weighted gray-level co-occurrence matrix technology are dynamically allocated based on the main texture direction;

[0016] Based on the Rayleigh scattering principle, cloud feature analysis is performed on the ground-based cloud map time series data to obtain the cloud coverage rate time series data and the cloud change rate time series data.

[0017] Furthermore, the step of extracting edge features from the time-series data of the ground cloud map to be analyzed based on the dynamic scale oriented gradient histogram to obtain the edge feature time-series data includes:

[0018] Based on the candidate unit scales in the preset unit scale set, pixel gradient energy change analysis is performed on each ground cloud map to be analyzed in the time series data of the ground cloud map to be analyzed, so as to obtain the target scale of different pixel positions in each ground cloud map to be analyzed.

[0019] Based on the target scale of different pixel locations in each of the ground cloud images to be analyzed, the directional gradient histogram is calculated for the ground cloud images to be analyzed to obtain the corresponding image histogram features.

[0020] The image histogram features corresponding to all the ground cloud images to be analyzed are summarized to generate the edge feature time series data.

[0021] Furthermore, the step of extracting texture features from the time-series data of the ground cloud map to be analyzed based on the multi-directional weighted gray-level co-occurrence matrix technology to obtain the texture feature time-series data includes:

[0022] Based on the preset orientation angle set and preset pixel spacing, gray-level co-occurrence matrix calculations are performed on each ground cloud map to be analyzed in the time series data of the ground cloud map to be analyzed, so as to obtain the corresponding orientation gray-level co-occurrence matrix set;

[0023] Texture features are extracted from each direction gray-level co-occurrence matrix in the set of gray-level co-occurrence matrices of each ground cloud map to be analyzed, to obtain the corresponding texture feature set; the texture feature set includes texture features of each directional angle;

[0024] Structural tensor analysis is performed on each of the foundation cloud maps to be analyzed to obtain the corresponding global dominant direction angles. Then, angular distance similarity analysis is performed between the global dominant direction angles and each direction angle in the preset direction angle set to obtain the corresponding direction weight set.

[0025] Based on the directional weight set of each of the ground cloud images to be analyzed, the texture features of each directional angle in the texture feature set are weighted and fused to obtain the corresponding image texture features;

[0026] The image texture features corresponding to all the ground cloud maps to be analyzed are summarized to generate the time series data of the texture features.

[0027] Furthermore, the step of performing structural tensor analysis on each of the foundation cloud maps to be analyzed to obtain the corresponding global dominant orientation angle includes:

[0028] The image gradient fields of each of the ground cloud images to be analyzed are obtained respectively, and the corresponding pixel gradient outer product matrix is ​​calculated according to the gradient components of different pixel positions in the image gradient fields.

[0029] The corresponding elements in the gradient outer product matrix of all pixels in each of the ground cloud maps to be analyzed are summed to obtain the corresponding structural tensor matrix.

[0030] The structural tensor matrix of each of the ground cloud maps to be analyzed is decomposed into eigenvalues ​​to obtain the eigenvector corresponding to the smallest eigenvalue as the texture dominant eigenvector. The direction information of the texture dominant eigenvector is extracted to obtain the global dominant direction angle.

[0031] Furthermore, the step of performing cloud feature analysis on the ground-based cloud map time-series data based on the Rayleigh scattering principle to obtain the cloud coverage rate time-series data and the cloud change rate time-series data includes:

[0032] Based on the coordinates of the solar pixel in each of the ground cloud maps to be analyzed, the pixel brightness attenuation coefficient corresponding to each pixel in the ground cloud map is calculated, and pixel binarization is performed based on the brightness attenuation coefficient and the corresponding red-blue channel ratio to obtain the corresponding binarized grayscale cloud map.

[0033] Based on the accumulated binary grayscale values ​​of all pixels in the binary grayscale cloud map of each of the aforementioned foundation cloud maps and the size of the foundation cloud map, the corresponding cloud coverage rate is obtained, and the cloud coverage rates of all the aforementioned foundation cloud maps are summarized to generate the cloud coverage rate time series data.

[0034] The cloud coverage time series data is processed to obtain the corresponding cloud coverage change time series data, and the cloud change rate time series data is obtained based on the cloud coverage change time series data and the time interval of the time series data.

[0035] Furthermore, the cluster-distributed photovoltaic information includes the geographical location, power time-series data, and meteorological time-series data of all photovoltaic power stations within the cluster where the target photovoltaic power station is located;

[0036] The step of obtaining spatiotemporal coupling feature time series data based on the clustered distributed photovoltaic information, the cloud coverage time series data, and the cloud change rate time series data includes:

[0037] Cluster analysis is performed on the geographical location and power time series data of each photovoltaic power station in the clustered distributed photovoltaic information to obtain the target photovoltaic power station subgroup;

[0038] Based on the power time-series data and meteorological time-series data of all photovoltaic power stations within the target photovoltaic power station subgroup, construct the corresponding graph node feature matrix time-series data;

[0039] Based on the time-series data of the graph node feature matrix, the time-series data of cloud coverage, and the time-series data of cloud change rate, the corresponding edge weight time-series data are calculated; the edge weight time-series data includes the edge weights between photovoltaic power station pairs within the target photovoltaic power station subgroup.

[0040] Based on the edge weight time series data, a weighted graph adjacency matrix time series data is generated, and based on the weighted graph adjacency matrix time series data and the graph node feature matrix time series data, the spatiotemporal coupling feature time series data is generated.

[0041] Secondly, embodiments of the present invention provide a distributed photovoltaic short-term power generation prediction system, the system comprising:

[0042] The data acquisition module is used to acquire ground cloud map time-series data, historical power time-series data, and clustered distributed photovoltaic information of the target photovoltaic power station;

[0043] The feature extraction module is used to extract features from the ground-based cloud map time-series data to obtain cloud map feature time-series data; the cloud map feature time-series data includes edge feature time-series data, texture feature time-series data, cloud coverage time-series data, and cloud change rate time-series data.

[0044] The first prediction module is used to perform power prediction based on the cloud map feature time series data and the historical power time series data, and obtain a first prediction result;

[0045] The second prediction module is used to obtain spatiotemporal coupling feature time series data based on the cluster distributed photovoltaic information, the cloud coverage time series data and the cloud change rate time series data, and to perform power prediction based on the pre-built second prediction model according to the spatiotemporal coupling feature time series data to obtain the second prediction result.

[0046] The result generation module is used to perform weighted fusion of the first prediction result and the second prediction result to obtain the target prediction result.

[0047] Thirdly, embodiments of the present invention also provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the above-described method.

[0048] Fourthly, embodiments of the present invention also provide a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the above-described method.

[0049] This invention provides a method, system, device, and medium for predicting short-term power generation of distributed photovoltaic (PV) power. The method involves acquiring ground-based cloud map time-series data, historical power time-series data, and clustered distributed PV information of a target PV power plant; extracting features from the ground-based cloud map time-series data to obtain cloud map feature time-series data, including edge feature time-series data, texture feature time-series data, cloud coverage time-series data, and cloud change rate time-series data; performing power prediction based on a pre-constructed first prediction model using the cloud map feature time-series data and historical power time-series data to obtain a first prediction result; acquiring spatiotemporal coupling feature time-series data based on the clustered distributed PV information, cloud coverage time-series data, and cloud change rate time-series data; performing power prediction based on a pre-constructed second prediction model to obtain a second prediction result; and weightedly fusing the first and second prediction results to obtain the target prediction result. Compared with existing technologies, this distributed photovoltaic short-term power generation prediction method, based on a multi-level cloud map feature analysis mechanism that extracts edge texture features and dynamic change features from ground-based cloud maps, combined with a time-coupled feature extraction and analysis mechanism between photovoltaic power plants based on dynamic change features, achieves effective integration and full utilization of multi-modal data. It can significantly improve the reliability, stability and accuracy of short-term power prediction results under complex weather scenarios such as cloudy and rainy weather, and thus provide high-precision and robust scheduling analysis support for the safe and stable operation of the power grid with a high proportion of renewable energy and the efficient consumption of renewable energy. Attached Figure Description

[0050] Figure 1 This is a flowchart illustrating the distributed photovoltaic short-term power generation prediction method in an embodiment of the present invention.

[0051] Figure 2 This is a schematic diagram of the structure of the distributed photovoltaic short-term power generation prediction system in an embodiment of the present invention;

[0052] Figure 3 This is an internal structural diagram of the computer device in an embodiment of the present invention;

[0053] The attached figures are labeled as follows:

[0054] 1. Data acquisition module; 2. Feature extraction module; 3. First prediction module; 4. Second prediction module; 5. Result generation module. Detailed Implementation

[0055] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. Obviously, the embodiments described below are only part of the embodiments of this invention and are used to illustrate the invention, but are not intended to limit the scope of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0056] In one embodiment, such as Figure 1 As shown, a method for predicting short-term power generation of distributed photovoltaic power is provided, including the following steps:

[0057] S11. Obtain the ground-based cloud map time-series data, historical power time-series data, and clustered distributed photovoltaic information of the target photovoltaic power station. The target photovoltaic power station can be understood as the analysis object for short-term power generation prediction. The corresponding ground-based cloud map time-series data and historical power time-series data can be understood as data to be analyzed within a certain time range collected according to the prediction acquisition frequency (e.g., once every 5 minutes or once every 1 minute) based on the short-term power prediction requirements. The ground-based cloud map time-series data can be obtained through existing equipment such as all-sky imagers, and the historical power time-series data can be obtained based on existing relevant power grid monitoring systems, which will not be detailed here. Clustered distributed photovoltaic information can be understood as basic classification data of the smallest group of photovoltaic power stations that are strongly correlated with the power generation of the target photovoltaic power station. In this embodiment, the clustered distributed photovoltaic information preferably includes the geographical location, power time series data and meteorological time series data of all photovoltaic power stations in the cluster where the target photovoltaic power station is located. The power time series data of each photovoltaic power station in the cluster is consistent with the collection duration and time scale of the historical power time series data of the aforementioned target photovoltaic power station. The meteorological time series data can be understood as multi-dimensional time series data that is time-aligned with the historical power time series data, preferably including temperature time series data, air pressure time series data and humidity time series data, etc.

[0058] S12. Perform feature extraction on the ground cloud map time series data to obtain cloud map feature time series data; wherein, cloud map feature time series data can be understood as time series data of multi-level cloud map features obtained by performing in-depth static and dynamic feature analysis on the ground cloud map at each time point in the ground cloud map time series data; in this embodiment, the cloud map feature time series data is preferably set to include edge feature time series data, texture feature time series data, cloud coverage time series data, and cloud change rate time series data.

[0059] Specifically, the step of extracting features from the ground-based cloud map time-series data to obtain cloud map feature time-series data includes:

[0060] Each ground cloud image in the time series data of the ground cloud image is preprocessed to obtain the ground cloud image time series data to be analyzed. The preprocessing can be understood as performing operations such as image size adjustment, grayscale transformation and brightness correction. The specific processing steps can be determined based on the actual application requirements and are not specifically limited here.

[0061] Based on the dynamic-scale directional gradient histogram (HGZ) technique, edge features are extracted from the time-series data of the ground-based cloud image to be analyzed, resulting in the edge feature time-series data. The dynamic-scale HGZ technique can be understood as an improved HGZ calculation technique that adaptively selects the optimal statistical scale based on the local gradient energy of the cloud image, considering that traditional HGZ calculations use fixed-size units for gradient statistics, which are difficult to adapt to the complex structures of multi-scale cloud clusters (such as small-scale cumulus clouds and large-scale stratiform clouds) in ground-based cloud images. This significantly improves the sensitivity of edge features to abrupt changes in cloud formations. Specifically, the gradient statistical scale in the dynamic-scale HGZ technique is adaptively determined based on gradient energy. The steps of extracting edge features from the time-series data of the ground-based cloud image to be analyzed based on the dynamic-scale HGZ technique to obtain the edge feature time-series data include:

[0062] Based on the candidate unit scales in the preset unit scale set, pixel gradient energy change analysis is performed on each of the ground cloud maps to be analyzed in the time series data of the ground cloud maps to be analyzed, to obtain the target scale of different pixel positions in each of the ground cloud maps to be analyzed; wherein, each candidate unit scale in the preset unit scale set can be determined based on actual analysis needs, and the preset unit scale set can be... S The candidate unit scale (number of pixels) is selected using pixel size values ​​that increase sequentially, for example... The specific steps for obtaining the target scale at different pixel locations in each ground cloud image to be analyzed may include:

[0063] Assumption The image coordinates are The pixel value of each pixel can be obtained by performing a convolution operation on the ground cloud map to be analyzed using the gradient operator [-1,0,1] based on the following expression. exist gradient components on the axis (positive direction to the right) and using gradient operators Perform convolution operations on the ground cloud map to be analyzed to obtain each pixel. exist gradient components on the axis (upward is the positive direction) :

[0064]

[0065] ;

[0066] Then, the pixel points can be obtained. The gradient magnitude at point is as follows:

[0067]

[0068] In the formula, For pixels The corresponding gradient magnitude.

[0069] Assuming pixel Corresponding image local window The gradient energy within can be calculated using the following formula:

[0070]

[0071] In the formula, For image local windows The window side length reflects the statistical scale; For pixels Corresponding image local window Gradient energy within;

[0072] Based on the above formula for calculating gradient energy within a local window of an image, and using the following formula with a preset unit scale set... S The gradient energy change rate of each candidate unit is used as the window side length to analyze the gradient energy change rate of the image patch. The scale that maximizes the gradient energy change rate is taken as the target scale for that pixel region.

[0073]

[0074] In the formula, and They are respectively preset unit scale sets S The first in k The and the first k +1 candidate unit scale, ,and K For a set of preset unit scales S The total number of candidate unit scales in the data; and Each pixel With a window side length of and The gradient energy within a local window of the image.

[0075] Based on the target scale of different pixel locations in each of the ground cloud images to be analyzed, an oriented gradient histogram (HOG) is calculated for the ground cloud images to be analyzed to obtain the corresponding image histogram features. The process of obtaining the image histogram features may include: uniformly dividing the ground cloud image to be analyzed into grid cells of fixed size, such that each grid cell corresponds to a regular rectangular region in the image; for each grid cell, analyzing the distribution of all pixels corresponding to the target scale within it, and determining the dominant target scale of the grid cell (e.g., the target scale with the most pixels) through statistical methods; after determining the dominant target scale of each grid cell, an oriented gradient histogram (HOG) is calculated for each grid cell using its corresponding dominant target scale. The process involves calculating the histogram of oriented gradients (OORGs). For grid cells with a larger dominant target scale, a relatively loose statistical window is used to capture the overall edge features of large-scale cloud clusters. For grid cells with a smaller dominant target scale, a finer statistical window is used to capture the local details of thin or small cloud clusters. Within each grid cell, a standard OORG histogram calculation process is followed, including gradient calculation, orientation quantization, and magnitude-weighted statistics, to generate a local feature vector reflecting the edge distribution characteristics of the corresponding region. Adjacent grid cells are further combined into detection windows. Within the detection window, the feature vectors are subjected to contrast normalization to eliminate interference from illumination changes. All normalized window feature vectors are then concatenated and stitched together according to a preset spatial order to form a global edge feature descriptor of a unified dimension, thus obtaining the required image histogram features.

[0076] The grid-based dynamic scale partitioning strategy provided in this embodiment effectively preserves the spatial distribution characteristics of image edge structures by incorporating multi-scale HOG descriptors calculated through a local scale adaptive mechanism. While maintaining the consistency of the feature dimensions of traditional directional gradient histograms, it ensures the computational efficiency of feature extraction and effectively solves the feature distortion problem of fixed-scale HOG in thin cloud or fast-moving cloud cluster scenarios, thereby improving the ability to represent multi-scale cloud cluster structures.

[0077] The image histogram features corresponding to all the ground cloud images to be analyzed are summarized to generate the edge feature time series data.

[0078] This embodiment uses dynamic scale directional gradient histogram technology to quantify and analyze the gradient direction distribution of image sub-regions, which can effectively capture geometric structure information with significant features in ground-based cloud images. It can not only accurately characterize the linear edge features of cloud textures, but also analyze the geometric deformation law of cloud contours through gradient modulus distribution, effectively improving the accuracy of cloud image edge feature extraction, and thus providing a reliable guarantee for the accuracy of subsequent power generation prediction based on cloud image features.

[0079] Based on the multi-directional weighted gray-level co-occurrence matrix (GLCM) technique, texture features are extracted from the time-series data of the ground cloud image to be analyzed, resulting in the texture feature time-series data. The multi-directional weighted GLCM technique addresses the shortcomings of traditional gray-level co-occurrence matrix (GLCM) algorithms, which, when applying equal weights to the four directions (0°, 45°, 90°, and 135°), neglect the significant directional bias of cloud textures under the influence of atmospheric flow (e.g., cirrus clouds exhibit banding, stratus clouds exhibit isotropic properties), leading to a decrease in the discriminative power of the analyzed texture features. Therefore, an improved GLCM algorithm is proposed, which estimates the local principal texture direction using a structural tensor and dynamically allocates the fusion weights of the corresponding texture features in each direction of the GLCM. This achieves direction-sensitive texture representation, effectively characterizing the texture roughness, uniformity, and directionality of the cloud image. Specifically, the weights of the gray-level co-occurrence matrix in each direction are dynamically allocated based on the principal texture direction.

[0080] Specifically, the step of extracting texture features from the time-series data of the ground cloud map to be analyzed based on the multi-directional weighted gray-level co-occurrence matrix technology to obtain the texture feature time-series data includes:

[0081] Based on a preset set of orientation angles and a preset pixel spacing, gray-level co-occurrence matrices are calculated for each ground cloud image in the time series data to be analyzed, resulting in a corresponding set of orientation gray-level co-occurrence matrices. The preset set of orientation angles includes four orientation angles: 0°, 45°, 90°, and 135°. The preset pixel spacing is the distance between pixel pairs used when constructing the gray-level co-occurrence matrix, and can be set based on actual application requirements. The corresponding set of orientation gray-level co-occurrence matrices includes gray-level co-occurrence matrices calculated based on each orientation angle in the preset set of orientation angles. The specific calculation process of the gray-level co-occurrence matrix can be directly referenced from existing gray-level co-occurrence matrix algorithms, and will not be detailed here.

[0082] Texture features are extracted from each directional gray-level co-occurrence matrix in the set of directional gray-level co-occurrence matrices of each of the foundation cloud images to be analyzed, resulting in a corresponding texture feature set. The texture feature set includes texture features for each directional angle, and the texture features may include contrast, energy, entropy, and correlation: contrast reflects the overall situation of gray-level differences in the cloud image (the depth of grooves in the image texture); energy reflects the uniformity of gray-level distribution and the coarseness of the texture; entropy represents the complexity of the cloud image texture or the uncertainty of information; and correlation measures the linear relationship between pixel values ​​in the cloud image, reflecting the local gray-level correlation of the cloud image texture. It should be noted that the specific calculation process for texture feature extraction based on the directional gray-level co-occurrence matrix in this embodiment can be referenced from relevant existing technologies and will not be detailed here.

[0083] Structural tensor analysis is performed on each of the foundation cloud maps to be analyzed to obtain the corresponding global dominant orientation angle. Then, angular distance similarity analysis is performed between each of the global dominant orientation angles and each orientation angle in the preset orientation angle set to obtain the corresponding orientation weight set. The global dominant orientation angle can be understood as the main texture direction determined based on the cloud map structural tensor analysis. Specifically, the step of performing structural tensor analysis on each of the foundation cloud maps to be analyzed to obtain the corresponding global dominant orientation angle includes:

[0084] The image gradient fields of each of the foundation cloud images to be analyzed are obtained respectively, and the corresponding pixel gradient outer product matrix is ​​calculated based on the gradient components at different pixel positions in the image gradient field; wherein, the image gradient field is composed of the X-axis gradient component (horizontal direction) and Y-axis gradient component (vertical direction) of each pixel, and the specific calculation process can refer to existing technology; pixel The corresponding pixel gradient outer product matrix It can be represented as:

[0085]

[0086] in, and These are the pixel points calculated based on the Sobel operator. The X-axis gradient components and Y-axis gradient components; This represents a Gaussian-weighted local average. and Each pixel The square of the projection along the X-axis and the square of the projection along the Y-axis; For pixels The degree of correlation between horizontal and vertical changes.

[0087] The corresponding elements in the outer product matrices of the gradients of all pixels in each of the aforementioned ground cloud images to be analyzed are summed to obtain the corresponding structure tensor matrix. The structure tensor matrix can be understood as a matrix that captures the overall pattern of the gradient distribution of the entire image based on three statistical quantities: horizontal energy, vertical energy, and directional change correlation. The horizontal energy is the sum of the squared projections of the pixel gradient outer product matrices along the X-axis of all pixels; the vertical energy is the sum of the squared projections of the pixel gradient outer product matrices along the Y-axis of all pixels; and the directional change correlation is the sum of the correlation between horizontal and vertical changes in the pixel gradient outer product matrices of all pixels. The structure tensor matrix obtained through this aggregation method makes texture direction analysis more robust to noise and local variations, ensuring the scientific validity and reliability of the extracted main texture directions.

[0088] Eigenvalue decomposition is performed on the structural tensor matrix of each of the ground cloud images to be analyzed. The eigenvector corresponding to the smallest eigenvalue is obtained as the dominant texture eigenvector, and the direction information of the dominant texture eigenvector is extracted to obtain the global dominant orientation angle. Eigenvalue decomposition of the structural tensor matrix yields two eigenvalues ​​and their corresponding eigenvectors. Considering that in actual texture orientation analysis, the eigenvector corresponding to the smaller eigenvalue has the smallest gradient change, reflecting the continuous orientation of the texture, the eigenvector corresponding to the smaller eigenvalue is selected to represent the dominant orientation of the image texture. After determining the dominant texture eigenvector, the corresponding global dominant orientation angle can be extracted from this eigenvector, expressed as:

[0089]

[0090] In the formula, The texture-dominant feature vector; The radian angle corresponding to the global dominant orientation angle needs to be converted into an angle value within the range of 0° to 180° to obtain the global dominant orientation angle of the image, which represents the overall direction of the image texture.

[0091] After obtaining the global dominant direction angle through the above steps, angular distance similarity analysis can be performed between the global dominant direction angle and each other direction angle to obtain the corresponding direction weights, which can be expressed as:

[0092]

[0093] In the formula, It is a set of integers; For the first in the preset direction angle set d There are several orientation angles. When the preset orientation angle set is {0°, 45°, 90°, 135°}, ; for The corresponding directional weights; An integer representing a texture direction with a periodicity of 180 degrees; The hyperparameter for controlling the decay rate of the directional weight is preferably set to 15°; It is an exponential function.

[0094] The texture features of each directional angle in the texture feature set are weighted and fused according to the directional weight set of each of the ground cloud images to be analyzed, so as to obtain the corresponding image texture features; that is, each texture feature in the image texture features is a feature obtained by weighted averaging of the texture features corresponding to each directional angle and the directional weight; it should be noted that each directional weight in the directional weight set should be normalized before use, which will not be described in detail here.

[0095] The image texture features corresponding to all the ground cloud maps to be analyzed are summarized to generate the time series data of the texture features.

[0096] This embodiment is based on a dynamic weight allocation mechanism for directional gray-level co-occurrence matrix fusion, which takes into account the significant directional preference of cloud textures under the influence of atmospheric flow. This mechanism can effectively improve the accuracy and reliability of texture feature extraction, thereby enhancing the application value of texture features in power prediction.

[0097] Based on the Rayleigh scattering principle, cloud feature analysis is performed on the time-series data of the ground-based cloud map to obtain the time-series data of cloud coverage and cloud change rate. Rayleigh scattering refers to the phenomenon where light is scattered by particles much smaller than its wavelength when passing through a transparent medium. Considering practical applications, based on the Rayleigh scattering principle (atmospheric scattering physics mechanism), there are significant differences in the spectral response of visible light bands between cloud layers and clear sky scenes: blue light (about 450 nm) is scattered much more strongly than red light (about 650 nm) in clear sky due to its shorter wavelength, resulting in a lower red-blue ratio in clear sky areas. In contrast, clouds scatter light evenly across all wavelengths due to Mie scattering, resulting in similar brightness in the red and blue channels and a higher ratio. To improve the effectiveness of cloud coverage and cloud change rate analysis, this embodiment preferably uses the red-blue spectral ratio method to binarize the ground-based cloud map to suppress the brightness attenuation interference caused by changes in solar altitude angle.

[0098] Specifically, the steps of performing cloud feature analysis on the ground-based cloud map time-series data based on the Rayleigh scattering principle to obtain the cloud coverage rate time-series data and the cloud change rate time-series data include:

[0099] Based on the coordinates of the solar pixels in each of the ground cloud maps to be analyzed, the pixel brightness attenuation coefficient corresponding to each pixel in the ground cloud map is calculated. Then, pixel binarization is performed based on the brightness attenuation coefficient and the corresponding red-blue channel ratio to obtain the corresponding binarized grayscale cloud map. The solar pixel coordinates can be obtained based on relevant existing technologies. For example, the theoretical position of the sun in the image can be directly calculated using the known shooting time, the geographical location (latitude and longitude) of the camera, the camera orientation (azimuth and elevation angle), and the field of view. Alternatively, image processing technology can be used to detect the sun's position through features such as brightness and shape. If the sun is not obscured by clouds, the sun's position can be regarded as a very bright, approximately circular area. After converting the image to grayscale, the bright area is extracted using a brightness threshold. Then, morphological operations are performed on the bright area to obtain connected bright areas. The solar area is then selected based on the area and roundness of the outline of each bright area, and the center point of the solar area is used as the pixel coordinates of the sun.

[0100] In this embodiment, the pixel brightness attenuation coefficient corresponding to each pixel in the ground-based cloud map can be understood as a coefficient set based on the distance between the pixel and the sun pixel to suppress the brightness attenuation interference caused by changes in the sun's altitude angle. It can be expressed as:

[0101]

[0102] In the formula,

[0103]

[0104] in, The image coordinates of the sun's pixel; For pixels Distance to the sun pixel; and This represents the maximum distance between the pixel and the sun pixel and the corresponding brightness attenuation coefficient; For pixels The corresponding pixel brightness attenuation coefficient.

[0105] After obtaining the pixel brightness attenuation coefficient at each pixel location using the above method, the following formula can be used to obtain the corresponding pixel binarized value based on the red-blue channel ratio at the corresponding pixel location, thereby obtaining the corresponding binarized grayscale cloud map:

[0106]

[0107] In the formula, For pixels The binary grayscale value; and Each pixel The red channel pixel values ​​and blue channel pixel values; 0.95 is the initial red-blue channel ratio threshold.

[0108] This embodiment uses the Rayleigh scattering principle to construct the red-blue spectral ratio as a feature quantity for binarizing ground-based cloud images. This effectively solves the problem that traditional ground-based cloud image threshold segmentation methods do not consider the drastic changes in atmospheric scattering intensity during periods of low solar altitude angle, such as sunrise and sunset, which leads to a significant decrease in the overall brightness of clear sky areas. Directly setting fixed brightness / chromaticity thresholds in RGB or HSV space can easily misclassify clear sky as clouds or miss thin clouds. This embodiment can effectively eliminate the influence of overall light intensity scaling, making the segmentation threshold light invariant, thereby effectively improving the construction accuracy of binarized grayscale cloud images.

[0109] Based on the accumulated binary grayscale values ​​of all pixels in the binary grayscale cloud images of each ground-based cloud image and the size of the ground-based cloud image, the corresponding cloud coverage rate is obtained. The cloud coverage rates of all the ground-based cloud images are then summarized to generate time-series data of the cloud coverage rate. The cloud coverage rate of each ground-based cloud image can be understood as an indicator used to objectively measure the degree of cloud cover in the sky, and can be expressed as:

[0110]

[0111] In the formula, and The width and height of the binary grayscale cloud image; This refers to cloud cover.

[0112] The cloud coverage rate time series data is processed to obtain corresponding cloud coverage change time series data. Based on the cloud coverage change time series data and the time series data duration interval, the cloud change rate time series data is obtained. Here, the data at each moment in the cloud coverage change time series data is the difference between the cloud coverage rate at each moment and the cloud coverage rate at the corresponding previous moment in the cloud coverage rate time series data. The corresponding time series data duration interval can be understood as the duration interval between adjacent sampling moments in the cloud coverage rate time series data. Therefore, the cloud change rate at each moment in the cloud change rate time series data can be expressed as:

[0113]

[0114] In the formula, The duration interval for time series data; and These are the time series data of cloud coverage. t Time and Corresponding cloud coverage; for t The rate of change of cloud cover at any given time.

[0115] The cloud coverage rate obtained in this embodiment can quantitatively characterize the spatial blocking effect of atmospheric obstruction on solar radiation flux. Its numerical change is directly related to the effective irradiated area of ​​the photovoltaic array's receiving surface. Furthermore, the cloud change rate can characterize the cloud system displacement characteristics in continuous time-series satellite cloud images, which can be used to effectively capture the dynamic evolution law of cloud structure under the influence of atmospheric flow field, thereby providing effective analytical data for the reliable prediction of short-term power generation based on photovoltaic power plants.

[0116] S13. Based on the cloud map feature time series data and the historical power time series data, power prediction is performed based on the pre-constructed first prediction model to obtain a first prediction result. The first prediction model can be understood as a model capable of predicting power generation within a short-term timeframe based on the input cloud map feature time series data and historical power time series data. In principle, any deep learning network model capable of this function can be used, such as the traditional Autoformer model, an improved Autoformer model obtained by replacing the autocorrelation mechanism in the Autoformer model with the Nystrom self-Attention mechanism, or other neural network models, etc. No specific limitations are made here. It should be noted that in this embodiment, the first prediction model can be understood as a model obtained by training a selected network model for short-term power generation prediction based on historical cloud map feature time series data and corresponding historical power time series data collected from multiple distributed photovoltaic power stations. The specific training process can be implemented by referring to the training methods of the relevant selected network models, which will not be detailed here.

[0117] S14. Based on the clustered distributed photovoltaic information, the cloud coverage time series data, and the cloud change rate time series data, obtain spatiotemporal coupling feature time series data, and based on the spatiotemporal coupling feature time series data, perform power prediction based on the pre-constructed second prediction model to obtain a second prediction result; wherein, the spatiotemporal coupling feature time series data can be understood as the time series data of the fused features obtained by weighted fusion based on the power time series data and meteorological time series data of each distributed photovoltaic power station in the smallest group where the target photovoltaic power station is located, as well as the corresponding cloud dynamic change characteristics, and the feature fusion weights of each related photovoltaic power station of the target photovoltaic power station determined by graph topology dynamic analysis.

[0118] Specifically, the step of obtaining spatiotemporal coupling feature time-series data based on the clustered distributed photovoltaic information, the cloud coverage time-series data, and the cloud change rate time-series data includes:

[0119] Cluster analysis is performed on the geographical location and power time-series data of each photovoltaic power station in the clustered distributed photovoltaic information to obtain a target photovoltaic power station subgroup. This target photovoltaic power station subgroup can be understood as containing redundant related information among neighboring photovoltaic power stations. Directly using data from all neighboring photovoltaic power stations for distributed photovoltaic graph modeling would not only increase the complexity of power prediction but also affect prediction performance. However, by performing correlation analysis on neighboring photovoltaic power stations and leveraging the prediction performance of deep learning algorithms at each photovoltaic power station, the distributed photovoltaic cluster is divided into multiple sub-regional clusters with inherent consistency to achieve dimensionality reduction. The smallest group containing the target photovoltaic power station can be obtained through the following steps:

[0120] Pre-calculated based on the number of distributed photovoltaic power stations included in the cluster distributed photovoltaic information Set the number of clusters The range of values ​​for ;

[0121] Based on the geographical coordinates and historical power generation records of distributed photovoltaic power stations, the K-means clustering method is used to cluster them. Cluster analysis was performed on the photovoltaic power stations to obtain the final results. One initial cluster center;

[0122] use Training with power time series data and meteorological time series data of the initial cluster centers A BLSTM (Bi-directional Long Short-Term Memory) model with consistent hyperparameters and model structure (other deep learning algorithms can also be used); then, respectively in Used at a photovoltaic site Power prediction was performed using several different BLSTM models, and their prediction errors were calculated using the following formula:

[0123]

[0124] In the formula, and They are respectively t Moment Photovoltaic Power Station Using prediction models The predicted power and the actual power; This represents the total number of time points. For photovoltaic power station Using prediction models The prediction error;

[0125] based on The prediction errors of the MBLSTM model at different photovoltaic power plants are used to update the cluster center labels of photovoltaic sites based on the following formula:

[0126]

[0127] In the formula, Indicates photovoltaic power station After updating the category labels of all sites, the cluster centers are updated by calculating the average power of the power plants in each group. The maximum number of clusters;

[0128] When the preset iteration threshold is reached, the clustering process terminates, yielding the desired target photovoltaic power station subgroup. Each photovoltaic power station within this subgroup is then used as a graph node, with its power time-series data and meteorological time-series data serving as node features. Edge connections between nodes are constructed based on the spatial relationships between the photovoltaic power stations, creating a corresponding relational topology graph for subsequent analysis. It should be noted that the number of clusters can be selected based on the clustering value with the smallest error on the test set. If multiple clustering values ​​with similar errors exist, the smaller value should be chosen to reduce the complexity of the graph model construction.

[0129] Based on the power time-series data and meteorological time-series data of all photovoltaic power stations within the target photovoltaic power station subgroup, a corresponding graph node feature matrix time-series data is constructed. That is, the graph node feature matrix time-series data at different times is composed of the feature vectors of the corresponding nodes of all photovoltaic power stations within the target photovoltaic power station subgroup at different times, summarizing the feature information of each node in the graph. Furthermore, the feature vector of each node is composed of the power and various meteorological data at the corresponding time. Assuming there are N photovoltaic power stations within the target photovoltaic power station subgroup V, and the node feature dimension is F, for example, when the features include power, temperature, air pressure, and humidity, F=4, and the graph node feature matrix at time t... It can be represented as:

[0130]

[0131] in, , , and for t Time map nodes N The corresponding power, temperature, air pressure, and humidity.

[0132] Based on the time-series data of the graph node feature matrix, the cloud coverage rate, and the cloud change rate, the corresponding edge weight time-series data are calculated. The edge weight time-series data includes the edge weights between pairs of photovoltaic power stations within the target photovoltaic power station subgroup. In this embodiment, a pre-trained graph attention network is preferably used to obtain these weights. The specific process is as follows:

[0133]

[0134] In the formula, It is a parameter vector that can map a high-dimensional feature space to the real number space; This represents a learnable weight matrix used for feature transformation; This represents the operation of concatenating vectors; and The graph node feature matrices at time t are respectively The Middlei row and number j The eigenvectors of the row; The data consists of time series data on cloud coverage and time series data on cloud change rate. t The cloud change feature vector is composed of the cloud coverage rate and cloud change rate at each time point; then the attention coefficients (edge ​​weights) obtained by normalization are:

[0135]

[0136] In the formula, Represents a non-linear activation function; and They are respectively t Graph node feature matrix at time step The Middle u row and number v The eigenvectors of the row; V A set of graph nodes; Graph nodes in the graph node feature matrix i With graph nodes j Edge weights between them; This is the activation function.

[0137] Based on the edge weight time series data, weighted graph adjacency matrix time series data is generated, and based on the weighted graph adjacency matrix time series data and the graph node feature matrix time series data, spatiotemporal coupling feature time series data is generated. The weighted graph adjacency matrix at each time point in the weighted graph adjacency matrix time series data can be understood as a weighted graph adjacency matrix obtained by superimposing edge weights dynamically adjusted in real time based on cloud dynamics characteristics onto a static adjacency matrix constructed based on the correlation of power data between photovoltaic power stations within the target photovoltaic power station subgroup. In practical applications, the static adjacency matrix can be directly determined based on the absolute value of the Pearson correlation coefficient of the power time series data of each photovoltaic power station within the target photovoltaic power station subgroup. The edge relationship between two photovoltaic power stations in this matrix can be determined based on the following formula:

[0138]

[0139] In the formula, Static adjacency matrix China Photovoltaic Power Station Node and photovoltaic power station nodes The relationship between the nodes is indicated by a value of 1, which means that there is an edge connection between the two nodes, and a value of 0, which means that there is no edge connection between the two nodes. For photovoltaic power station nodes and photovoltaic power station nodes The degree of correlation between historical power data is expressed as the absolute value of the Pearson correlation coefficient. This is a threshold used to determine the relevance of an edge connection; N The total number of photovoltaic power stations within the target photovoltaic power station subgroup V.

[0140] By superimposing the edge weights corresponding to each time step onto the static adjacency matrices obtained through the above method, the required weighted graph adjacency matrix can be obtained. Then, based on the weighted graph adjacency matrix, the node features of the target photovoltaic power station at the corresponding time step are weighted and fused with the feature vectors of each node with which it has a correlation in the graph node feature matrix. This yields the spatiotemporal coupling features at the corresponding time step in the spatiotemporal coupling feature time series data, which can be used for subsequent power prediction. This information aggregation mechanism, based on the graph attention network's perception of the current cloud dynamic level to improve the reliability and rationality of the fusion weights, enables the target photovoltaic node to effectively absorb the feature information of neighboring nodes, significantly improving the accuracy of spatial feature aggregation, thereby effectively enhancing the expressive power of the target photovoltaic power station's fusion features. It should be noted that the weighted fusion process of the spatiotemporal coupling features in this embodiment can refer to the relevant implementations in existing graph attention networks, which will not be detailed here.

[0141] After obtaining the spatiotemporal coupling characteristic time series data using the aforementioned method, it can be input into a second prediction model capable of short-term power prediction based on this type of time series data to obtain the desired second prediction result. In principle, the second prediction model can be any network model capable of predictive analysis based on time series data. However, in order to ensure the accuracy and efficiency of data prediction while maintaining consistency with the aforementioned target photovoltaic power station subgroup analysis principle, this embodiment preferably uses a second prediction model constructed by training a BLSTM network. It should be noted that in the actual training process, it is only necessary to construct the relevant training dataset according to the aforementioned method for obtaining spatiotemporal coupling characteristic time series data, and train the model with stable network parameters using the existing BLSTM network training method. This will not be elaborated further here.

[0142] This embodiment constructs a distributed photovoltaic graph structure based on sub-region clustering. Considering the dynamic changes in cloud cover, it utilizes a spatiotemporal coupling feature extraction mechanism that dynamically learns spatial correlation weights such as geographical proximity and electrical coupling through a graph attention network. This mechanism not only ensures the adaptability of coupling feature extraction under different weather conditions and multi-site distribution scenarios, but also avoids interference from redundant information, thereby providing a reliable guarantee for the accuracy of power prediction based on spatiotemporal coupling features.

[0143] S15. The first prediction result and the second prediction result are weighted and fused to obtain the target prediction result. The target prediction result can be understood as the final prediction result obtained by fusion analysis of the first prediction result and the second prediction result based on the contribution of cloud map features and spatiotemporal coupling features to the short-term power generation prediction of photovoltaic power plants. The fusion weights used for the first prediction result and the second prediction result can be set based on actual application needs or determined through learning and training. No specific limitation is made here.

[0144] This invention, through providing time-series data of ground-based cloud maps, historical power time-series data, and clustered distributed photovoltaic information of a target photovoltaic power station, extracts features from the ground-based cloud map time-series data to obtain cloud map feature time-series data, including edge feature time-series data, texture feature time-series data, cloud coverage time-series data, and cloud change rate time-series data. Based on the cloud map feature time-series data and historical power time-series data, power prediction is performed using a pre-built first prediction model to obtain a first prediction result. Furthermore, based on the clustered distributed photovoltaic information, cloud coverage time-series data, and cloud change rate time-series data, spatiotemporal coupling feature time-series data is obtained, and power prediction is performed using a pre-built second prediction model to obtain... The distributed photovoltaic power generation prediction scheme, which yields the second prediction result and the target prediction result by weighted fusion of the first and second prediction results, can effectively integrate and fully utilize multimodal data based on a multi-level cloud map feature analysis mechanism that extracts edge texture features and dynamic change features from ground cloud maps, combined with a time-coupled feature extraction and analysis mechanism between photovoltaic power stations based on dynamic change features. This significantly improves the reliability, stability, and accuracy of short-term power prediction results under complex weather conditions such as cloudy and rainy weather, thereby providing high-precision and robust scheduling analysis support for the safe and stable operation of the power grid with a high proportion of new energy grid connection and the efficient consumption of new energy.

[0145] It should be noted that although the steps in the flowchart above are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise explicitly stated in this document, there is no strict order requirement for the execution of these steps, and they can be executed in other orders.

[0146] In one embodiment, such as Figure 2 As shown, a distributed photovoltaic short-term power generation prediction system is provided, the system comprising:

[0147] The data acquisition module is used to acquire ground cloud map time-series data, historical power time-series data, and clustered distributed photovoltaic information of the target photovoltaic power station;

[0148] The feature extraction module is used to extract features from the ground-based cloud map time-series data to obtain cloud map feature time-series data; the cloud map feature time-series data includes edge feature time-series data, texture feature time-series data, cloud coverage time-series data, and cloud change rate time-series data.

[0149] The first prediction module is used to perform power prediction based on the cloud map feature time series data and the historical power time series data, and obtain a first prediction result;

[0150] The second prediction module is used to obtain spatiotemporal coupling feature time series data based on the cluster distributed photovoltaic information, the cloud coverage time series data and the cloud change rate time series data, and to perform power prediction based on the pre-built second prediction model according to the spatiotemporal coupling feature time series data to obtain the second prediction result.

[0151] The result generation module is used to perform weighted fusion of the first prediction result and the second prediction result to obtain the target prediction result.

[0152] Specific limitations regarding the distributed photovoltaic short-term power generation prediction system can be found in the limitations of the distributed photovoltaic short-term power generation prediction method described above; the corresponding technical effects are equivalent and will not be repeated here. Each module in the aforementioned distributed photovoltaic short-term power generation prediction system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0153] Figure 3 An internal structural diagram of a computer device is shown in one embodiment. This computer device may specifically be a terminal or a server. Figure 3 As shown, the computer device includes a processor, memory, network interface, display, camera, and input device connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it can implement a method for predicting short-term distributed photovoltaic power generation. The display screen can be an LCD screen or an e-ink display screen. The input device can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device casing, or an external keyboard, touchpad, or mouse.

[0154] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present invention and does not constitute a limitation on the computer device to which the present invention is applied. Specific computing devices may include more or fewer components than those shown in the figure, or combine certain components, or have the same component arrangement.

[0155] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the method described above.

[0156] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0157] In summary, the distributed photovoltaic short-term power generation prediction method, system, equipment, and medium provided by this invention are based on a multi-level cloud map feature analysis mechanism that extracts edge texture features and dynamic change features from ground cloud maps. Combined with a time-coupled feature extraction and analysis mechanism between photovoltaic power plants based on dynamic change features, this invention achieves effective integration and full utilization of multi-modal data. It can significantly improve the reliability, stability, and accuracy of short-term power prediction results under complex weather scenarios such as cloudy and rainy weather, thereby providing high-precision and robust scheduling analysis support for the safe and stable operation of the power grid with a high proportion of renewable energy and the efficient consumption of renewable energy.

[0158] The various embodiments in this specification are described in a progressive manner. For directly identical or similar parts of the embodiments, refer to each other. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. It should be noted that the technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0159] The above-described embodiments are merely preferred embodiments of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various improvements and substitutions without departing from the principles of the present invention, and these improvements and substitutions should also be considered within the scope of protection of the present invention. Therefore, the scope of protection of this invention should be determined by the scope of the claims.

Claims

1. A distributed photovoltaic short-term power generation prediction method, characterized by, The method includes: Acquire ground-based cloud map time-series data, historical power time-series data, and clustered distributed photovoltaic information of the target photovoltaic power station; Feature extraction is performed on the ground-based cloud map time-series data to obtain cloud map feature time-series data; the cloud map feature time-series data includes edge feature time-series data, texture feature time-series data, cloud coverage time-series data, and cloud change rate time-series data. Based on the cloud map feature time series data and the historical power time series data, power prediction is performed based on a pre-built first prediction model to obtain a first prediction result; Based on the cluster-distributed photovoltaic information, the cloud coverage time-series data, and the cloud change rate time-series data, spatiotemporal coupling characteristic time-series data are obtained. Then, based on the spatiotemporal coupling characteristic time-series data, power prediction is performed using a pre-constructed second prediction model to obtain a second prediction result. The cluster-distributed photovoltaic information includes the geographical location, power time-series data, and meteorological time-series data of all photovoltaic power stations within the cluster where the target photovoltaic power station is located. The step of obtaining the spatiotemporal coupling characteristic time-series data based on the cluster-distributed photovoltaic information, the cloud coverage time-series data, and the cloud change rate time-series data includes: Cluster analysis is performed on the geographical location and power time series data of each photovoltaic power station in the clustered distributed photovoltaic information to obtain the target photovoltaic power station subgroup; Based on the power time-series data and meteorological time-series data of all photovoltaic power stations within the target photovoltaic power station subgroup, construct the corresponding graph node feature matrix time-series data; Based on the time-series data of the graph node feature matrix, the time-series data of cloud coverage, and the time-series data of cloud change rate, the corresponding edge weight time-series data are calculated; the edge weight time-series data includes the edge weights between photovoltaic power station pairs within the target photovoltaic power station subgroup. Based on the edge weight time series data, a weighted graph adjacency matrix time series data is generated, and based on the weighted graph adjacency matrix time series data and the graph node feature matrix time series data, the spatiotemporal coupling feature time series data is generated. The first prediction result and the second prediction result are weighted and fused to obtain the target prediction result.

2. The method for predicting short-term distributed photovoltaic power generation as described in claim 1, characterized in that, The step of extracting features from the ground-based cloud map time-series data to obtain cloud map feature time-series data includes: Each foundation cloud map in the time series data of the foundation cloud map is preprocessed to obtain the time series data of the foundation cloud map to be analyzed. Based on the dynamic scale oriented gradient histogram technique, edge features are extracted from the time series data of the ground cloud map to be analyzed to obtain the edge feature time series data; the gradient statistical scale in the dynamic scale oriented gradient histogram technique is adaptively determined based on gradient energy. Based on the multi-directional weighted gray-level co-occurrence matrix technology, texture features are extracted from the time-series data of the ground cloud map to be analyzed to obtain the texture feature time-series data; the weights of the gray-level co-occurrence matrix in each direction of the multi-directional weighted gray-level co-occurrence matrix technology are dynamically allocated based on the main texture direction; Based on the Rayleigh scattering principle, cloud feature analysis is performed on the ground-based cloud map time series data to obtain the cloud coverage rate time series data and the cloud change rate time series data.

3. The method for predicting short-term distributed photovoltaic power generation as described in claim 2, characterized in that, The step of extracting edge features from the time-series data of the ground cloud map to be analyzed based on the dynamic scale directional gradient histogram to obtain the edge feature time-series data includes: Based on the candidate unit scales in the preset unit scale set, pixel gradient energy change analysis is performed on each ground cloud map to be analyzed in the time series data of the ground cloud map to be analyzed, so as to obtain the target scale of different pixel positions in each ground cloud map to be analyzed. Based on the target scale of different pixel locations in each of the ground cloud images to be analyzed, the directional gradient histogram is calculated for the ground cloud images to be analyzed to obtain the corresponding image histogram features. The image histogram features corresponding to all the ground cloud images to be analyzed are summarized to generate the edge feature time series data.

4. The method for predicting short-term distributed photovoltaic power generation as described in claim 2, characterized in that, The step of extracting texture features from the time-series data of the ground cloud map to be analyzed based on the multi-directional weighted gray-level co-occurrence matrix technology to obtain the texture feature time-series data includes: Based on the preset orientation angle set and preset pixel spacing, gray-level co-occurrence matrix calculations are performed on each ground cloud map to be analyzed in the time series data of the ground cloud map to be analyzed, so as to obtain the corresponding orientation gray-level co-occurrence matrix set; Texture features are extracted from each direction gray-level co-occurrence matrix in the set of gray-level co-occurrence matrices of each of the ground cloud maps to be analyzed, to obtain the corresponding texture feature set; the texture feature set includes texture features of each directional angle; Structural tensor analysis is performed on each of the foundation cloud maps to be analyzed to obtain the corresponding global dominant orientation angles. Then, angular distance similarity analysis is performed between the global dominant orientation angles and each orientation angle in the preset orientation angle set to obtain the corresponding orientation weight set. Based on the directional weight set of each of the ground cloud images to be analyzed, the texture features of each directional angle in the texture feature set are weighted and fused to obtain the corresponding image texture features; The image texture features corresponding to all the ground cloud maps to be analyzed are summarized to generate the time series data of the texture features.

5. The method for predicting short-term distributed photovoltaic power generation as described in claim 4, characterized in that, The step of performing structural tensor analysis on each of the foundation cloud maps to be analyzed to obtain the corresponding global dominant orientation angle includes: The image gradient fields of each of the ground cloud images to be analyzed are obtained respectively, and the corresponding pixel gradient outer product matrix is ​​calculated according to the gradient components of different pixel positions in the image gradient fields. The corresponding elements in the gradient outer product matrix of all pixels in each of the ground cloud maps to be analyzed are summed to obtain the corresponding structural tensor matrix. The structural tensor matrix of each of the ground cloud maps to be analyzed is decomposed into eigenvalues. The eigenvector corresponding to the smallest eigenvalue is obtained as the texture dominant eigenvector. The direction information of the texture dominant eigenvector is extracted to obtain the global dominant direction angle.

6. The method for predicting short-term distributed photovoltaic power generation as described in claim 2, characterized in that, The steps of performing cloud feature analysis on the ground-based cloud map time-series data based on the Rayleigh scattering principle to obtain the cloud coverage rate time-series data and the cloud change rate time-series data include: Based on the coordinates of the solar pixel in each of the ground cloud maps to be analyzed, the pixel brightness attenuation coefficient corresponding to each pixel in the ground cloud map is calculated, and pixel binarization is performed based on the brightness attenuation coefficient and the corresponding red-blue channel ratio to obtain the corresponding binarized grayscale cloud map. Based on the accumulated binary grayscale values ​​of all pixels in the binary grayscale cloud map of each of the aforementioned foundation cloud maps and the size of the foundation cloud map, the corresponding cloud coverage rate is obtained, and the cloud coverage rates of all the aforementioned foundation cloud maps are summarized to generate the cloud coverage rate time series data. The cloud coverage time series data is processed to obtain the corresponding cloud coverage change time series data, and the cloud change rate time series data is obtained based on the cloud coverage change time series data and the time interval of the time series data.

7. A distributed photovoltaic short-term power generation prediction system, characterized in that, The system, employing the distributed photovoltaic short-term power generation prediction method as described in claim 1, comprises: The data acquisition module is used to acquire ground cloud map time-series data, historical power time-series data, and clustered distributed photovoltaic information of the target photovoltaic power station; The feature extraction module is used to extract features from the ground-based cloud map time-series data to obtain cloud map feature time-series data; the cloud map feature time-series data includes edge feature time-series data, texture feature time-series data, cloud coverage time-series data, and cloud change rate time-series data. The first prediction module is used to perform power prediction based on the cloud map feature time series data and the historical power time series data, and obtain a first prediction result; The second prediction module is used to obtain spatiotemporal coupling feature time series data based on the cluster distributed photovoltaic information, the cloud coverage time series data and the cloud change rate time series data, and to perform power prediction based on the pre-built second prediction model according to the spatiotemporal coupling feature time series data to obtain the second prediction result. The result generation module is used to perform weighted fusion of the first prediction result and the second prediction result to obtain the target prediction result.

8. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

Images