Ultra-short-term photovoltaic power prediction method based on multi-source cloud picture and convolutional vision Transform

Abstract

The invention discloses an ultra-short-term photovoltaic power prediction method, system, equipment and medium based on a multi-source cloud picture and convolutional vision Transform, and the method comprises the following steps: collecting meteorological information data, photovoltaic unit data, and foundation cloud picture and satellite cloud picture data, and carrying out the preprocessing of numerical input data; carrying out distortion correction on the foundation cloud picture, reconstructing satellite cloud picture space information in combination with a geometric correction technology, determining a satellite cloud picture concern area based on a solar azimuth angle and a cloud cluster motion vector, and carrying out resampling processing and gray scale standardization conversion; analyzing a motion vector field of adjacent time sequence cloud image pixel points to generate an optical flow graph; the continuous multi-source optical flow graph is coded through a convolutional vision Transform, and cloud graph features representing cloud cluster information are extracted; and fusing the multi-source cloud picture features, the meteorological data and the photovoltaic unit data, constructing a photovoltaic power prediction model based on DLinear, and performing photovoltaic power generation power prediction.

Description

Technical Field

[0001] This invention belongs to the field of photovoltaic prediction technology for new energy power generation and grid connection, specifically involving an ultra-short-term photovoltaic power prediction method, system, computer equipment, and storage medium based on multi-source cloud maps and convolutional visual Transformer. Background Technology

[0002] Photovoltaic power output is significantly affected by meteorological conditions (especially cloud dynamics). The rapid movement, shading effect, and morphological evolution of clouds can lead to drastic fluctuations in photovoltaic power, posing a severe challenge to grid dispatch and stable operation. However, with the development of remote sensing and sky imaging technologies, satellite cloud images and ground-based cloud images have been introduced into photovoltaic prediction modeling, providing more possibilities for improving the accuracy of photovoltaic power prediction. Existing photovoltaic power prediction methods mainly rely on numerical weather forecasts or single satellite cloud image information, which have the following limitations: First, traditional satellite cloud images have insufficient spatial resolution, making it difficult to accurately represent the details of local cloud movement; second, although ground-based cloud images can reflect high-precision cloud distribution, their field of view is limited and lacks effective coordination with satellite data; third, existing models lack the ability to spatiotemporally model the dynamic evolution characteristics of clouds. For example, traditional convolutional neural networks are unable to capture global features, while ordinary visual Transformers, due to the lack of local inductive bias, tend to ignore local motion details in cloud images. Therefore, there is an urgent need for a photovoltaic power generation prediction scheme that integrates high-dimensional features of multi-source cloud images, enhances cloud motion modeling capabilities, and takes into account meteorological and photovoltaic unit data. This will improve the accuracy of predictions, provide an important foundation for maintaining the stable and efficient operation of the new power system, and be crucial for the economic transactions of the electricity market. Summary of the Invention

[0003] The main objective of this invention is to overcome the shortcomings and deficiencies of the prior art and provide an ultra-short-term photovoltaic power prediction method, system, computer equipment, and storage medium based on multi-source cloud maps and convolutional visual Transformer, so as to provide a foundation for the utilization and grid connection of photovoltaic resources.

[0004] To achieve the above objectives, the present invention adopts the following technical solution:

[0005] The first aspect of this invention discloses an ultra-short-term photovoltaic power prediction method based on multi-source cloud maps and convolutional visual Transformer, the ultra-short-term photovoltaic power prediction method comprising the following steps:

[0006] S1. Collect meteorological information data, photovoltaic unit operation data, ground-based cloud map and satellite cloud map data. Among them, the meteorological information data includes temperature, irradiance and solar azimuth angle. The numerical input data is preprocessed.

[0007] S2. Perform distortion correction on the ground-based cloud image and reconstruct the satellite cloud image using geometric correction techniques; determine the key areas of interest in the reconstructed satellite cloud image based on the solar azimuth angle and cloud motion vector to achieve spatial feature dimensionality reduction, where the cloud motion vector refers to the maximum normal motion speed of the cloud; resample and perform grayscale standardization on the preprocessed ground-based cloud image and satellite cloud image to obtain a standardized multi-source cloud image dataset;

[0008] S3. Using a multi-source cloud image dataset, optical flow method is used to analyze the motion vector field of pixels in the multi-source cloud image at adjacent time points, obtain the motion direction and velocity information of pixels in the multi-source cloud image as time changes, generate an optical flow feature map that characterizes the dynamic evolution of the cloud image, and use the optical flow feature map as a multi-source optical flow map.

[0009] S4. Encode the above multi-source optical flow map using a convolutional visual Transformer to extract multi-source cloud map feature vectors that represent cloud motion and occlusion information. The convolutional visual Transformer consists of multiple layers of encoding blocks connected in sequence. The processed multi-source optical flow map is input into each layer of the convolutional Transformer in sequence. Each layer of the encoding block consists of a convolutional operation layer, a multi-head self-attention layer, a normalization layer, a feedforward neural network layer, and a pooling layer.

[0010] S5. Construct a photovoltaic power prediction model based on DLinear. This model includes an input layer, a decomposition layer, a trend term linear layer, a residual term linear layer, a merging layer, and an output layer connected in sequence. The input layer receives meteorological information data with a time resolution of t, photovoltaic unit operation data, and multi-source cloud image feature vectors from the past T hours, and maps them to a pre-specified unified embedding dimension to achieve multi-modal feature alignment. Then, they are spliced ​​along the time dimension to form fused features. The decomposition layer decomposes the fused features into trend terms and residual terms, which are input to the trend term linear layer and the residual term linear layer, respectively, to extract short-term trend features and periodic change features. The short-term trend features and periodic change features are sent to the merging layer for fusion to reconstruct the time series change trend. Finally, the output layer generates the prediction result, achieving high-precision photovoltaic power generation prediction.

[0011] Furthermore, considering that outliers can cause the model to learn incorrect patterns, severely affecting the reliability of the prediction model, step S1 aims to construct a high-quality, spatiotemporally consistent multi-source heterogeneous dataset through outlier removal and Newton interpolation completion, providing reliable input for subsequent modeling. The process is as follows:

[0012] S11. Collect meteorological information data, photovoltaic unit data, ground-based cloud map and satellite cloud map data, remove outliers and missing values ​​from the numerical input data, and obtain preprocessed meteorological information data and photovoltaic unit data;

[0013] S12. Newton interpolation method is used to fill in the missing meteorological data and photovoltaic unit data respectively to obtain complete numerical input data.

[0014] Furthermore, step S2 bridges the inherent differences between ground-based and satellite cloud images in terms of viewing angle, resolution, and geometric distortion. The scheme first corrects the distortion of the ground-based image and performs geometric reprojection on the satellite cloud image, aligning them pixel-level within the same coordinate frame. Then, it uses the solar azimuth angle and the "maximum normal cloud velocity" vector to delineate a dynamic region of interest, compressing the entire spatial information to key areas that may affect photovoltaic output, achieving spatial dimensionality reduction. Finally, through resampling and grayscale standardization, it eliminates scale and illumination differences, outputting a spatiotemporally consistent, low-redundancy multi-source cloud image dataset. This solves the problems of inaccurate cloud shadow recognition and high computational load caused by viewing angle deviation and high data dimensionality in existing technologies. The process is as follows:

[0015] S2.1. Perform distortion correction on the foundation cloud map, as follows:

[0016] S2.1.1 Query the parameters of the sky imager to obtain the first, second, and third radial distortion parameters k1, k2, and k3 and the first and second tangential distortion parameters p1 and p2 of the ground-based cloud image;

[0017] S2.1.2 Mapping distorted pixels in the ground cloud map to distort-free locations, the process is as follows:

[0018] , ,

[0019] In the formula, x dis y dis r represents the pixel coordinates in the distorted image, x and y represent the x and y coordinates of the pixel in the undistorted image, and r represents the pixel coordinates in the undistorted image. dis = This represents the distance from the distorted pixel to the center point of the image;

[0020] S2.1.3 Since distortion correction may cause pixel position discontinuities, bilinear interpolation is used to reconstruct the corrected image.

[0021] S2.2. Geometric correction techniques are used to reconstruct spatial information in satellite cloud images. Based on the solar elevation angle, azimuth angle, and cloud motion vectors, the region of interest in the satellite cloud image is determined. The process is as follows:

[0022] S2.2.1 First, calculate the horizontal and vertical coordinates L of the corrected cloud map pixels. x L y :

[0023] , ,

[0024] In the formula, C p The x-coordinate of a pixel in a satellite cloud image, C off C represents the offset of the horizontal axis. fac L represents the scaling factor of the x-axis. p The vertical coordinate of a pixel in a satellite cloud image, L off L represents the offset of the vertical axis. fac The scaling factor for the ordinate;

[0025] Then calculate the latitude and longitude L of each pixel in the satellite cloud image. on L at :

[0026] , ,

[0027] , , , ,

[0028] , ,

[0029] In the formula, h d D represents the altitude of the satellite. ea and D eb For the Earth's major and minor axes, Indicates direct path gain. Indicates northbound gain. This indicates the total altitude of the satellite to the monitoring point on the Earth's surface. Indicates eastward gain. Indicates vertical gain. λ represents the combined gain. mid The reference longitude is used to reconstruct the image based on the latitude and longitude of the pixels, resulting in a corrected satellite cloud image.

[0030] S2.2.2 Calculate the maximum latitude and longitude variation of cloud cluster movement:

[0031] , , ,

[0032] In the formula, d max v represents the maximum distance the cloud can travel. max The maximum speed of the cloud formation. The input feature time scale is R, where R is the Earth's radius. This represents the change in latitude. This represents the change in longitude.

[0033] Determine the boundaries of key areas of interest in the cloud map based on changes in latitude and longitude:

[0034] ,

[0035] In the formula, X0 is the latitude of the power station to be measured, and Y0 is the longitude of the power station to be measured. The minimum latitude of the region, The maximum latitude of the region, Minimum longitude of the region The maximum longitude of the region can be used to determine the maximum geographical range of satellite cloud image input through the region's boundaries;

[0036] S2.3. Perform resampling and grayscale normalization conversion on the preprocessed multi-source cloud image.

[0037] Furthermore, step S3 addresses the issue of the strong randomness of cloud movement and the difficulty of a single sensor in characterizing its short-term dynamics. It applies optical flow to a multi-source cloud image dataset, simultaneously calculating pixel-level motion vector fields from both ground-based and satellite perspectives. This yields information on the direction and velocity of the cloud as it evolves over time, generating a multi-source optical flow map in a unified coordinate system. By fusing dual-view optical flow features, it compensates for potential occlusion or blind spots in single observations, improving the robustness and accuracy of cloud movement estimation. This provides high spatiotemporal resolution dynamic input for subsequent cloud shadow arrival time prediction and photovoltaic power fluctuation early warning. The process is as follows:

[0038] On the cloud image at given adjacent time points, the relationship between the displacement of each cloud pixel and the binomial coefficient is derived through binomial transformation and constraint conditions. The displacement of each pixel along different directions is estimated, and optical flow maps along the X and Y axes are obtained. A binomial expansion is performed, and a binomial approximation is made in a local window with each pixel on the cloud image as the center.

[0039] ,

[0040] In the formula, r1, r2, r3, r4, r5, and r6 are the first, second, third, fourth, fifth, and sixth weight coefficients of the iterative operation, and m and n represent the horizontal and vertical coordinates of the cloud map pixels, which are obtained using the weighted least squares method.

[0041] Through the The expansion yields the relationship between the grayscale value and the coordinates of each pixel in the cloud map. The grayscale value of the pixel is then converted into a spatial coordinate representation, and the displacement vectors in the X and Y directions are obtained to obtain the optical flow map.

[0042] Furthermore, step S4 overcomes the shortcomings of traditional convolution in efficiently completing global image modeling, while pure Transformer is prone to losing local occlusion details. Therefore, a convolutional visual Transformer is introduced to encode continuous multi-source optical flow maps step by step: first, convolutional operation layers extract local information within each encoding block; then, multi-head self-attention layers capture global cloud displacement and occlusion relationships; normalization and feedforward network layers stabilize training and enhance nonlinear expression; pooling layers progressively downsample to expand the receptive field; finally, a compact multi-source cloud map feature vector is output, achieving a unified representation of local occlusion details and global motion semantics in the same embedding space, providing highly discriminative dynamic features for subsequent short-term cloud shadow coverage prediction. The process is as follows:

[0043] S4.1 In the convolutional visual Transformer, the convolutional operation layer embeds multi-source optical flow map sequences through convolutional operations to extract local features, as detailed below: In the formula, F conv The output feature map represents the convolution operation, Conv() represents the convolution function, O represents the multi-source optical flow map tensor, E represents the kernel size, S represents the stride of the convolution operation, and P represents padding. The convolution operation process is shown below:

[0044] In the formula, i and j represent the index positions of the convolution operation on the multi-source optical flow graph tensor, c represents the output channel, and F conv (i,j,c) represents the value of the output feature map of the convolution operation at position (i,j) and channel c, e and l represent the two-dimensional position information of each weight element in the convolution kernel, E(e,l) represents the weight of the convolution kernel at (e,l), b is the bias term, and O The part representing the multi-source optical flow map convolution operation flattens the embedded feature map into a sequence, as follows:

[0045] In the formula, F flat Flatten() is the flattened sequence;

[0046] S4.2 The operation process of the multi-head self-attention layer and normalization layer in the convolutional visual Transformer is as follows:

[0047] ,

[0048] ,

[0049] ,

[0050] ,

[0051] In the formula, Q represents the query vector, K represents the key vector, V represents the value vector, ConProj() represents the convolution projection operation, and W... Q W K W V Let d represent the learnable weight matrices of the query vector, query vector, and value vector, respectively, where A represents the attention weight matrix score, and d represents the value vector. k This represents the dimension of the key vector, MultiHead() represents the multi-head self-attention function, Concat() represents the concatenation function, and head... i Let W represent the i-th attention head, h represent the number of attention heads, and W represent the number of attention heads. o This represents the output weight matrix, and LayerNorm() represents the layer normalization function;

[0052] S4.3 The operation process of the feedforward neural network layer in the convolutional vision Transformer is as follows:

[0053] ,

[0054] ,

[0055] In the formula, FFN() represents the feedforward neural network function, W1 and W2 are the first and second feedforward weight matrices, respectively, and b1 and b2 are the first and second feedforward bias terms, respectively.

[0056] S4.4 In the convolutional visual Transformer, the pooling layer obtains multi-source cloud map feature vectors through pooling operations, as follows:

[0057] In the formula, X ne This represents the feature vector of a multi-source cloud map, and Pool() represents the pooling function.

[0058] Furthermore, step S5 addresses the challenge of traditional black-box models for photovoltaic power generation failing to simultaneously account for trends and transients. This design explicitly decouples multimodal inputs within the DLinear framework: first, multi-source features from the past T hours are mapped to the same embedding dimension through the input layer and time-aligned to form unified fused features; then, the decomposition layer breaks them down into trend and residual terms, which are fed into corresponding linear layers to capture short-term trend features and periodic change features; the merging layer then weights and reconstructs these two features, preserving macroscopic trends while amplifying abrupt changes caused by clouds; finally, the output layer outputs a high-precision power curve in one step, achieving simultaneous modeling of "trend + disturbance" with linear complexity, balancing computational efficiency and prediction accuracy in multi-cloud scenarios. The process includes:

[0059] S51. The input layer receives meteorological information data with a time resolution of t, photovoltaic unit operation data, and multi-source cloud image feature vectors from the past T hours, and maps them to a pre-specified unified embedding dimension to achieve multimodal feature alignment. Then, they are concatenated along the time dimension to form fused features, as follows: In the formula, X np Represents meteorological information data, X pv X represents the operating data of the photovoltaic unit. ne X represents the feature vector of a multi-source cloud map. 0 Indicates fused features; Embed() represents the mapping embedding function; Concat represents the concatenation function.

[0060] S52. The decomposition layer decomposes the fused features into trend terms and residual terms, as follows: In the formula, X R X represents the trend term. T Represents the residual term; Decompose() represents the moving average kernel decomposition function.

[0061] S53. Input the trend term and residual term into the trend term linear layer and residual term linear layer respectively, and extract short-term trend features and periodic change features respectively. The process is as follows: , ,

[0062] In the formula, H R H represents short-term trend characteristics. T W represents the characteristic of periodic change. R W T These represent the linear layer of the trend term and the linear layer of the residual term, respectively.

[0063] S54. The extracted short-term trend features and periodic change features are fed into the merging layer for fusion to reconstruct the time series change trend. Finally, the output layer generates the prediction result. The process is as follows: In the formula, Y represents the final prediction result, and FC() represents the fully connected function of the output layer.

[0064] A second aspect of this invention discloses a photovoltaic power generation prediction system based on multi-source cloud image information and a convolutional visual Transformer. Running the aforementioned photovoltaic power generation prediction method based on multi-source cloud image information and a convolutional visual Transformer, the ultra-short-term photovoltaic power prediction system includes:

[0065] The data acquisition and preprocessing module collects meteorological information data, photovoltaic unit operation data, ground-based cloud map data, and satellite cloud map data. Among them, the meteorological information data includes temperature, irradiance, and solar azimuth angle. The module preprocesses the numerical input data.

[0066] The multi-source cloud image processing module performs distortion correction on the ground-based cloud image and reconstructs the satellite cloud image using geometric correction techniques. Based on the solar azimuth angle and cloud cluster motion vector, it determines the key areas of interest in the reconstructed satellite cloud image, achieving spatial feature dimensionality reduction. Here, the cloud cluster motion vector refers to the maximum normal motion velocity of the cloud cluster. The preprocessed ground-based cloud image and satellite cloud image are resampled and grayscale standardized to obtain a standardized multi-source cloud image dataset.

[0067] The optical flow feature map generation module uses a multi-source cloud map dataset and employs the optical flow method to analyze the motion vector field of pixels in the multi-source cloud map at adjacent time points. It obtains the motion direction and velocity information of the pixels in the multi-source cloud map as time changes, generates an optical flow feature map that represents the dynamic evolution of the cloud map, and uses the optical flow feature map as a multi-source optical flow map.

[0068] The feature extraction module encodes the multi-source optical flow map using a convolutional visual Transformer to extract multi-source cloud map feature vectors that represent cloud motion and occlusion information. The convolutional visual Transformer consists of multiple layers of encoding blocks connected in sequence. The processed multi-source optical flow map is input into each encoding block of the convolutional visual Transformer in sequence. Each encoding block consists of a convolutional operation layer, a multi-head self-attention layer, a normalization layer, a feedforward neural network layer, and a pooling layer.

[0069] The power prediction module is used to construct a photovoltaic power prediction model based on DLinear. This model includes an input layer, a decomposition layer, a trend term linear layer, a residual term linear layer, a merging layer, and an output layer connected in sequence. The input layer receives meteorological information data with a time resolution of t, photovoltaic unit operation data, and multi-source cloud image feature vectors from the past T hours, and maps them to a pre-specified unified embedding dimension to achieve multimodal feature alignment. Then, they are concatenated along the time dimension to form a fused feature. The decomposition layer decomposes the fused feature into a trend term and a residual term, which are input to the trend term linear layer and the residual term linear layer, respectively, to extract short-term trend features and periodic change features. The short-term trend features and periodic change features are fed into the merging layer for fusion to reconstruct the time series trend. Finally, the output layer generates the prediction result, achieving high-precision photovoltaic power generation prediction.

[0070] A third aspect of the present invention discloses a computer storage medium storing a computer program that, when executed by a processor, implements any of the above-described methods for ultra-short-term photovoltaic power prediction based on multi-source cloud maps and convolutional visual Transformers.

[0071] A fourth aspect of the present invention discloses a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the instructions of any of the above-described ultra-short-term photovoltaic power prediction methods based on multi-source cloud maps and convolutional visual Transformers.

[0072] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0073] This invention reconstructs a high-quality meteorological-photovoltaic data sequence without missing data or abrupt changes through "anomaly detection + Newton interpolation." It then compresses ground-based-satellite cloud images to the key cloud shadow area using distortion-geometric dual correction and solar azimuth-cloud velocity vector registration, solving the problems of perspective bias and dimensionality curse. Subsequently, it generates a dual-source optical flow map using optical flow method, and a convolutional visual Transformer is used to amplify local cloud edge details and global motion dependencies layer by layer, outputting image features representing cloud distribution and motion information. Finally, within the DLinear framework, the fused sequence is explicitly decoupled into a trend-residual bilinear projection to efficiently capture both long-term photovoltaic variation characteristics and short-term photovoltaic variation characteristics caused by cloud shading. This achieves end-to-end, real-time, high-precision, and highly generalized ultra-short-term photovoltaic power prediction with zero artificial physical quantity definition. Attached Figure Description

[0074] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0075] Figure 1 This is a flowchart of the photovoltaic power generation prediction method based on multi-source cloud map information and convolutional visual Transformer disclosed in this invention;

[0076] Figure 2 This is a schematic diagram of the process by which the convolutional visual Transformer encodes the above-mentioned multi-source optical flow map in this invention;

[0077] Figure 3 This is a schematic diagram of the photovoltaic power prediction model based on DLinear in this invention;

[0078] Figure 4 This is a schematic diagram of the process for achieving high-precision photovoltaic power generation prediction in step S5 of embodiment 1 of the present invention;

[0079] Figure 5 This is the cloud image collected in step S1 of embodiment 1 of the present invention;

[0080] Figure 6 This is a comparison chart of photovoltaic power generation data before and after preprocessing in Embodiment 1 of the present invention;

[0081] Figure 7 This is a schematic diagram of the multi-source cloud image after resampling and grayscale normalization conversion in Embodiment 1 of the present invention;

[0082] Figure 8 This is the multi-source optical flow map generated in step S3 of embodiment 1 of the present invention;

[0083] Figure 9 This is a schematic diagram of the error evaluation results within the prediction range of each model in Embodiment 1 of the present invention;

[0084] Figure 10 This is a schematic diagram of the plotted curves of the photovoltaic predicted value and the actual value of the experimental site on a certain day in Embodiment 1 of the present invention;

[0085] Figure 11 This is a graph showing the error evaluation results for different input sequence lengths within the prediction range of each model in Embodiment 1 of the present invention;

[0086] Figure 12 This is a structural block diagram of the intelligent handwriting detection system for automatic homework correction in Embodiment 3 of the present invention;

[0087] Figure 13 This is a structural block diagram of the computer device in Embodiment 4 of the present invention. Detailed Implementation

[0088] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present 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 the present application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present application without creative effort are within the scope of protection of the present application.

[0089] In this application, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this application can be combined with other embodiments.

[0090] Example 1

[0091] Photovoltaic power generation forecasting is the foundation of energy management and optimized scheduling in new power systems, and its forecasting accuracy directly affects the overall operational performance of the system. However, the temporal characteristics within the photovoltaic output cycle and sudden weather changes can significantly impact its forecasting accuracy. Therefore, improving the accuracy of photovoltaic power generation forecasting is of great significance for guiding the orderly optimized scheduling of new power systems. The flowchart is shown in Figure 1. Specifically, it mainly includes collecting meteorological information data, photovoltaic unit data, ground-based cloud images, and satellite cloud image data; preprocessing the numerical input data; performing distortion correction on the ground-based cloud images; reconstructing the spatial information of the satellite cloud images using geometric correction techniques; determining the region of interest in the satellite cloud images based on the solar azimuth angle and cloud motion vectors; performing resampling and grayscale normalization conversion; analyzing the motion vector fields of adjacent temporal cloud image pixels to generate optical flow maps; encoding continuous multi-source optical flow maps using a convolutional vision Transformer to extract cloud image features representing cloud information; and fusing multi-source cloud image features, meteorological data, and photovoltaic unit data to construct a photovoltaic power forecasting model based on DLinear for photovoltaic power generation forecasting. The ultra-short-term photovoltaic power forecasting method disclosed in this embodiment is as follows:

[0092] S1. Collect meteorological information data, photovoltaic unit operation data, ground-based cloud map and satellite cloud map data. Among them, the meteorological information data includes temperature, irradiance and solar azimuth angle. The numerical input data is preprocessed.

[0093] In step S1, meteorological information data, photovoltaic unit operation data, ground-based cloud imagery, and satellite cloud imagery data are collected. The meteorological information data includes temperature, irradiance, and solar azimuth. Some of the raw cloud imagery and meteorological data are shown below. Figure 5 As shown in Table 1.

[0094] Table 1. Schematic diagram of some raw cloud image data and meteorological data

[0095]

[0096] Next, the numerical input data is preprocessed. Taking the preprocessing of photovoltaic power generation data as an example, the results are as follows: Figure 6 As shown. Figure 6 This is a comparison chart of photovoltaic power generation data before and after preprocessing. Figure 6 The data on the left side of the middle page contains actual photovoltaic power generation data with missing values. Figure 6 The right side of the image shows the actual photovoltaic power generation data after a series of preprocessing steps, including interpolation of missing values.

[0097] S2. Distortion correction is applied to the ground-based cloud image, and the satellite cloud image is reconstructed using geometric correction techniques. Based on the solar azimuth angle and cloud motion vectors, the key areas of interest in the reconstructed satellite cloud image are determined to achieve spatial feature dimensionality reduction. Here, the cloud motion vector refers to the maximum normal velocity of the cloud cluster. The preprocessed ground-based and satellite cloud images are resampled and subjected to grayscale standardization to obtain a standardized multi-source cloud image dataset. The specific process of step S2 in this embodiment is as follows:

[0098] S2.1. Perform distortion correction on the foundation cloud map, as follows:

[0099] S2.1.1 Query the parameters of the sky imager to obtain the first, second, and third radial distortion parameters k1, k2, and k3 and the first and second tangential distortion parameters p1 and p2 of the ground-based cloud image;

[0100] S2.1.2 Mapping distorted pixels in the ground cloud map to distort-free locations, the process is as follows:

[0101] , ,

[0102] In the formula, x dis y dis r represents the pixel coordinates in the distorted image, x and y represent the x and y coordinates of the pixel in the undistorted image, and r represents the pixel coordinates in the undistorted image. dis = This represents the distance from the distorted pixel to the center point of the image;

[0103] S2.1.3 Since distortion correction may cause pixel position discontinuities, bilinear interpolation is used to reconstruct the corrected image.

[0104] S2.2. Geometric correction techniques are used to reconstruct spatial information in satellite cloud images. Based on the solar elevation angle, azimuth angle, and cloud motion vectors, the region of interest in the satellite cloud image is determined. The process is as follows:

[0105] S2.2.1 First, calculate the horizontal and vertical coordinates L of the corrected cloud map pixels. x L y :

[0106] , ,

[0107] In the formula, C p The x-coordinate of a pixel in a satellite cloud image, C off C represents the offset of the horizontal axis. fac L represents the scaling factor of the x-axis. p The vertical coordinate of a pixel in a satellite cloud image, L off L represents the offset of the vertical axis. facThe scaling factor for the ordinate;

[0108] Then calculate the latitude and longitude L of each pixel in the satellite cloud image. on L at :

[0109] , ,

[0110] , , , ,

[0111] , ,

[0112] In the formula, h d D represents the altitude of the satellite. ea and D eb For the Earth's major and minor axes, Indicates direct path gain. Indicates northbound gain. This indicates the total altitude of the satellite to the monitoring point on the Earth's surface. Indicates eastward gain. Indicates vertical gain. λ represents the combined gain. mid The reference longitude is used to reconstruct the image based on the latitude and longitude of the pixels, resulting in a corrected satellite cloud image.

[0113] S2.2.2 Calculate the maximum latitude and longitude variation of cloud cluster movement:

[0114] , , ,

[0115] In the formula, d max v represents the maximum distance the cloud can travel. max The maximum speed of the cloud formation. The input feature time scale is R, where R is the Earth's radius. This represents the change in latitude. This represents the change in longitude.

[0116] Determine the boundaries of key areas of interest in the cloud map based on changes in latitude and longitude: ,

[0117] In the formula, X0 is the latitude of the power station to be measured, and Y0 is the longitude of the power station to be measured. The minimum latitude of the region, The maximum latitude of the region, Minimum longitude of the region The maximum longitude of the region can be used to determine the maximum geographical range of satellite cloud image input through the region's boundaries;

[0118] S2.3. Perform resampling and grayscale normalization conversion on the preprocessed multi-source cloud image.

[0119] In step S2, distortion correction is performed on the ground-based cloud image, and the satellite cloud image is reconstructed by combining geometric correction technology. Based on the solar azimuth angle and cloud motion vector, the key areas of interest in the reconstructed satellite cloud image are determined to achieve spatial feature dimensionality reduction. Here, the cloud motion vector refers to the maximum normal motion speed of the cloud. The preprocessed ground-based cloud image and satellite cloud image are resampled and grayscale standardized to obtain a standardized multi-source cloud image dataset.

[0120] Based on the ground cloud image distortion correction, the image after correction, resampling, and grayscale normalization conversion is as follows: Figure 7 As shown.

[0121] S3. Using a multi-source cloud image dataset, optical flow method is used to analyze the motion vector field of pixels in the multi-source cloud image at adjacent time points, obtain the motion direction and velocity information of pixels in the multi-source cloud image as time changes, generate an optical flow feature map that characterizes the dynamic evolution of the cloud image, and use the optical flow feature map as a multi-source optical flow map.

[0122] In step S3, using a multi-source cloud image dataset, the motion vector field of pixels in the multi-source cloud image at adjacent time points is analyzed using the optical flow method. This obtains the motion direction and velocity information of the pixels in the multi-source cloud image as they change over time, generating an optical flow feature map characterizing the dynamic evolution of the cloud image. This optical flow feature map is then used as the multi-source optical flow map, and the resulting optical flow map is shown below. Figure 8 As shown.

[0123] S4. Encode the multi-source optical flow map using a convolutional visual Transformer to extract multi-source cloud map feature vectors representing cloud motion and occlusion information. The convolutional visual Transformer consists of multiple sequentially connected encoding blocks. The processed multi-source optical flow map is sequentially input into each encoding block of the convolutional visual Transformer. Each encoding block consists of a convolutional operation layer, a multi-head self-attention layer, a normalization layer, a feedforward neural network layer, and a pooling layer. Step S4 is as follows:

[0124] S4.1 In the convolutional visual Transformer, the convolutional operation layer embeds multi-source optical flow map sequences through convolutional operations to extract local features, as detailed below: In the formula, F convThe output feature map represents the convolution operation, Conv() represents the convolution function, O represents the multi-source optical flow map tensor, E represents the kernel size, S represents the stride of the convolution operation, and P represents padding. The convolution operation process is shown below:

[0125] ,

[0126] In the formula, i and j represent the index positions of the convolution operation on the multi-source optical flow graph tensor, c represents the output channel, and F conv (i,j,c) represents the value of the output feature map of the convolution operation at position (i,j) and channel c, e and l represent the two-dimensional position information of each weight element in the convolution kernel, E(e,l) represents the weight of the convolution kernel at (e,l), b is the bias term, and O The part representing the multi-source optical flow map convolution operation flattens the embedded feature map into a sequence, as follows: ,

[0127] In the formula, F flat Flatten() is the flattened sequence;

[0128] S4.2 The operation process of the multi-head self-attention layer and normalization layer in the convolutional visual Transformer is as follows:

[0129] , ,

[0130] ,

[0131] ,

[0132] In the formula, Q represents the query vector, K represents the key vector, V represents the value vector, ConProj() represents the convolution projection operation, and W... Q W K W V Let d represent the learnable weight matrices of the query vector, query vector, and value vector, respectively, where A represents the attention weight matrix score, and d represents the value vector. k This represents the dimension of the key vector, MultiHead() represents the multi-head self-attention function, Concat() represents the concatenation function, and head... i Let W represent the i-th attention head, h represent the number of attention heads, and W represent the number of attention heads. o This represents the output weight matrix, and LayerNorm() represents the layer normalization function;

[0133] S4.3 The operation process of the feedforward neural network layer in the convolutional vision Transformer is as follows:

[0134] ,

[0135] ,

[0136] In the formula, FFN() represents the feedforward neural network function, W1 and W2 are the first and second feedforward weight matrices, respectively, and b1 and b2 are the first and second feedforward bias terms, respectively.

[0137] S4.4 In the convolutional visual Transformer, the pooling layer obtains multi-source cloud map feature vectors through pooling operations, as follows:

[0138] In the formula, X ne This represents the feature vector of a multi-source cloud map, and Pool() represents the pooling function.

[0139] S5. Construct a photovoltaic power prediction model based on DLinear. This model includes an input layer, a decomposition layer, a trend term linear layer, a residual term linear layer, a merging layer, and an output layer connected in sequence. The input layer receives meteorological information data with a time resolution of t, photovoltaic unit operation data, and multi-source cloud image feature vectors from the past T hours, and maps them to a pre-specified unified embedding dimension to achieve multi-modal feature alignment. Then, they are spliced ​​along the time dimension to form fused features. The decomposition layer decomposes the fused features into trend terms and residual terms, which are input to the trend term linear layer and the residual term linear layer, respectively, to extract short-term trend features and periodic change features. The short-term trend features and periodic change features are sent to the merging layer for fusion to reconstruct the time series change trend. Finally, the output layer generates the prediction result, achieving high-precision photovoltaic power generation prediction.

[0140] To verify the superiority of the method, a public dataset was selected, and a high-performance server with an Intel(R) Xeon(R) Platinum8350C CPU @ 2.60 GHz, 42-GB RAM, and an NVIDIA GeForce RTX 3090 (24GB) was chosen as the test platform for the experiment. The programming languages ​​were PyTorch 2.1.0 and Python 3.8, and all deep learning models were trained in parallel on the GPU.

[0141] Furthermore, advanced benchmark models PatchTST, LightTS, DLinear, LSTM-Transformer, and ResNet were selected for comparative analysis. The error evaluation results within the prediction range of each model are as follows: Figure 9 As shown. From Figure 9As can be seen from this, the photovoltaic power prediction method disclosed in this embodiment yields prediction results with the smallest mean absolute error (MAE) and normalized root mean square error (NRMSE) compared to other benchmark models such as PatchTST, LightTS, DLinear, LSTM-Transformer, and ResNet.

[0142] The curves showing the predicted and actual photovoltaic values ​​at the experimental site on a certain day are as follows: Figure 10 As shown, from Figure 10 As can be seen from this, the photovoltaic power prediction value obtained by the photovoltaic power prediction method disclosed in this embodiment is closer to the actual value than other benchmark models such as PatchTST, LightTS, DLinear, LSTM-Transformer and ResNet.

[0143] Since the length of the input sequence is crucial to the model's predictive performance, four input sequence lengths (24, 36, 48, and 96) were selected for experiments to verify their impact. A grid method was then used to validate the experiments on a dataset. The error evaluation results for different input sequence lengths within the prediction range of each model are shown below. Figure 11 As shown.

[0144] In summary, the ultra-short-term photovoltaic power prediction method based on multi-source cloud maps and convolutional visual Transformer disclosed in this embodiment verifies that the present invention achieves high-efficiency synchronous capture of long-term photovoltaic variation characteristics and short-term photovoltaic variation characteristics caused by cloud shading by comparing the error evaluation results within the prediction range of each model, the curve comparison between the photovoltaic predicted value and the actual value of the experimental site on a certain day, and the error evaluation results of different input sequence lengths within the prediction range of each model.

[0145] Example 2

[0146] like Figure 12 As shown, this embodiment provides a photovoltaic power generation prediction system based on multi-source cloud image information and convolutional visual Transformer. The photovoltaic power generation prediction system includes: a data acquisition and preprocessing module 1201, a multi-source cloud image processing module 1202, an optical flow feature map generation module 1203, a feature extraction module 1204, and a power prediction module 1205. The specific functions of each module are as follows:

[0147] The data acquisition and preprocessing module 1201 collects meteorological information data, photovoltaic unit operation data, ground-based cloud map and satellite cloud map data. Among them, the meteorological information data includes temperature, irradiance and solar azimuth angle. The module preprocesses the numerical input data.

[0148] The multi-source cloud image processing module 1202 performs distortion correction on the ground-based cloud image and reconstructs the satellite cloud image by combining geometric correction technology; it determines the key areas of interest in the reconstructed satellite cloud image based on the solar azimuth angle and cloud cluster motion vector, thereby achieving spatial feature dimensionality reduction. Here, the cloud cluster motion vector refers to the maximum normal motion speed of the cloud cluster; it then resamples and performs grayscale standardization on the preprocessed ground-based cloud image and satellite cloud image to obtain a standardized multi-source cloud image dataset.

[0149] The optical flow feature map generation module 1203 uses a multi-source cloud map dataset and employs the optical flow method to analyze the motion vector field of pixels in the multi-source cloud map at adjacent time points. It obtains the motion direction and velocity information of the pixels in the multi-source cloud map as time changes, generates an optical flow feature map that characterizes the dynamic evolution of the cloud map, and uses the optical flow feature map as a multi-source optical flow map.

[0150] The feature extraction module 1204 encodes the multi-source optical flow map using a convolutional visual Transformer to extract multi-source cloud map feature vectors that represent cloud motion and occlusion information. The convolutional visual Transformer consists of multiple layers of encoding blocks connected in sequence. The processed multi-source optical flow map is sequentially input into each encoding block of the convolutional visual Transformer. Each encoding block consists of a convolutional operation layer, a multi-head self-attention layer, a normalization layer, a feedforward neural network layer, and a pooling layer.

[0151] The power prediction module 1205 is used to construct a photovoltaic power prediction model based on DLinear. This model includes an input layer, a decomposition layer, a trend term linear layer, a residual term linear layer, a merging layer, and an output layer connected in sequence. The input layer receives meteorological information data with a time resolution of t, photovoltaic unit operation data, and multi-source cloud image feature vectors from the past T hours, and maps them to a pre-specified unified embedding dimension to achieve multi-modal feature alignment. Then, they are spliced ​​along the time dimension to form fused features. The decomposition layer decomposes the fused features into trend terms and residual terms, which are input to the trend term linear layer and the residual term linear layer, respectively, to extract short-term trend features and periodic change features. The short-term trend features and periodic change features are sent to the merging layer for fusion to reconstruct the time series change trend. Finally, the output layer generates the prediction result, achieving high-precision photovoltaic power generation prediction.

[0152] Example 3

[0153] This embodiment provides a computer device, which can be a computer, such as... Figure 13 As shown, the processor 1302, memory, input device 1303, display 1304, and network interface 1305 are connected via system bus 1301. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium 1306 and internal memory 1307. The non-volatile storage medium 1306 stores the operating system, computer programs, and database. The internal memory 1307 provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. When the processor 1302 executes the computer programs stored in the memory, it implements the photovoltaic power generation prediction method based on multi-source cloud map information and convolutional visual Transformer proposed in Embodiment 1, which includes the following steps:

[0154] S1. Collect meteorological information data, photovoltaic unit operation data, ground-based cloud map and satellite cloud map data;

[0155] S2. Perform distortion correction on the ground-based cloud image and reconstruct the satellite cloud image using geometric correction techniques; determine the key areas of interest in the reconstructed satellite cloud image based on the solar azimuth angle and cloud motion vector to achieve spatial feature dimensionality reduction; resample and perform grayscale standardization on the preprocessed ground-based cloud image and satellite cloud image to obtain a standardized multi-source cloud image dataset.

[0156] S3. Using a multi-source cloud image dataset, optical flow method is used to analyze the motion vector field of pixels in the multi-source cloud image at adjacent time points, obtain the motion direction and velocity information of pixels in the multi-source cloud image as time changes, generate an optical flow feature map that characterizes the dynamic evolution of the cloud image, and use the optical flow feature map as a multi-source optical flow map.

[0157] S4. Encode the above multi-source optical flow map using a convolutional visual Transformer to extract multi-source cloud map feature vectors that represent cloud motion and occlusion information;

[0158] S5. Construct a photovoltaic power prediction model based on DLinear. This model includes an input layer, a decomposition layer, a trend term linear layer, a residual term linear layer, a merging layer, and an output layer connected in sequence. The input layer receives meteorological information data with a time resolution of t, photovoltaic unit operation data, and multi-source cloud image feature vectors from the past T hours, and maps them to a pre-specified unified embedding dimension to achieve multi-modal feature alignment. Then, they are spliced ​​along the time dimension to form fused features. The decomposition layer decomposes the fused features into trend terms and residual terms, which are input to the trend term linear layer and the residual term linear layer, respectively, to extract short-term trend features and periodic change features. The short-term trend features and periodic change features are sent to the merging layer for fusion to reconstruct the time series change trend. Finally, the output layer generates the prediction result, achieving high-precision photovoltaic power generation prediction.

[0159] Example 4

[0160] This embodiment provides a storage medium, which is a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the photovoltaic power generation prediction method based on multi-source cloud image information and convolutional visual Transformer proposed in Embodiment 1 above, including the following steps:

[0161] S1. Collect meteorological information data, photovoltaic unit operation data, ground-based cloud map and satellite cloud map data;

[0162] S2. Perform distortion correction on the ground-based cloud image and reconstruct the satellite cloud image using geometric correction techniques; determine the key areas of interest in the reconstructed satellite cloud image based on the solar azimuth angle and cloud motion vector to achieve spatial feature dimensionality reduction; resample and perform grayscale standardization on the preprocessed ground-based cloud image and satellite cloud image to obtain a standardized multi-source cloud image dataset.

[0163] S3. Using a multi-source cloud image dataset, optical flow method is used to analyze the motion vector field of pixels in the multi-source cloud image at adjacent time points, obtain the motion direction and velocity information of pixels in the multi-source cloud image as time changes, generate an optical flow feature map that characterizes the dynamic evolution of the cloud image, and use the optical flow feature map as a multi-source optical flow map.

[0164] S4. Encode the above multi-source optical flow map using a convolutional visual Transformer to extract multi-source cloud map feature vectors that represent cloud motion and occlusion information;

[0165] S5. Construct a photovoltaic power prediction model based on DLinear. This model includes an input layer, a decomposition layer, a trend term linear layer, a residual term linear layer, a merging layer, and an output layer connected in sequence. The input layer receives meteorological information data with a time resolution of t, photovoltaic unit operation data, and multi-source cloud image feature vectors from the past T hours, and maps them to a pre-specified unified embedding dimension to achieve multi-modal feature alignment. Then, they are spliced ​​along the time dimension to form fused features. The decomposition layer decomposes the fused features into trend terms and residual terms, which are input to the trend term linear layer and the residual term linear layer, respectively, to extract short-term trend features and periodic change features. The short-term trend features and periodic change features are sent to the merging layer for fusion to reconstruct the time series change trend. Finally, the output layer generates the prediction result, achieving high-precision photovoltaic power generation prediction.

[0166] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0167] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A method for predicting ultra-short-term photovoltaic power based on multi-source cloud maps and convolutional visual Transformer, characterized in that, The ultra-short-term photovoltaic power prediction method includes the following steps: S1. Collect meteorological information data, photovoltaic unit operation data, ground-based cloud map and satellite cloud map data. Among them, the meteorological information data includes temperature, irradiance and solar azimuth angle. The numerical input data is preprocessed. S2. Perform distortion correction on the ground-based cloud image and reconstruct the satellite cloud image using geometric correction techniques; determine the key areas of interest in the reconstructed satellite cloud image based on the solar azimuth angle and cloud motion vector to achieve spatial feature dimensionality reduction, where the cloud motion vector refers to the maximum normal motion speed of the cloud; resample and perform grayscale standardization on the preprocessed ground-based cloud image and satellite cloud image to obtain a standardized multi-source cloud image dataset; S3. Using a multi-source cloud image dataset, optical flow method is used to analyze the motion vector field of pixels in the multi-source cloud image at adjacent time points, obtain the motion direction and velocity information of pixels in the multi-source cloud image as time changes, generate an optical flow feature map that characterizes the dynamic evolution of the cloud image, and use the optical flow feature map as a multi-source optical flow map. S4. The multi-source optical flow map is encoded by the convolutional visual Transformer to extract the multi-source cloud map feature vector representing cloud motion and occlusion information. The convolutional visual Transformer consists of multiple layers of encoding blocks connected in sequence. The processed multi-source optical flow map is input into each layer of the convolutional visual Transformer in sequence. Each layer of the encoding block consists of a convolutional operation layer, a multi-head self-attention layer, a normalization layer, a feedforward neural network layer, and a pooling layer. S5. Construct a photovoltaic power prediction model based on DLinear. This model includes an input layer, a decomposition layer, a trend term linear layer, a residual term linear layer, a merging layer, and an output layer connected in sequence. The input layer receives meteorological information data with a time resolution of t, photovoltaic unit operation data, and multi-source cloud image feature vectors from the past T hours, and maps them to a pre-specified unified embedding dimension to achieve multi-modal feature alignment. Then, they are spliced ​​along the time dimension to form fused features. The decomposition layer decomposes the fused features into trend terms and residual terms, which are input to the trend term linear layer and the residual term linear layer, respectively, to extract short-term trend features and periodic change features. The short-term trend features and periodic change features are sent to the merging layer for fusion to reconstruct the time series change trend. Finally, the output layer generates the prediction result, achieving high-precision photovoltaic power generation prediction.

2. The photovoltaic power generation prediction method based on multi-source cloud map information and convolutional visual Transformer according to claim 1, characterized in that, The process of step S1 is as follows: S11. Collect meteorological information data, photovoltaic unit data, ground-based cloud map and satellite cloud map data, remove outliers and missing values ​​from the numerical input data, and obtain preprocessed meteorological information data and photovoltaic unit data; S12. Newton interpolation method is used to fill in the missing meteorological data and photovoltaic unit data respectively to obtain complete numerical input data.

3. The photovoltaic power generation prediction method based on multi-source cloud map information and convolutional visual Transformer according to claim 1, characterized in that, The process of step S2 is as follows: S2.1 Perform distortion correction on the foundation cloud map; S2.

2. Reconstruct spatial information in satellite cloud images using geometric correction techniques. Determine the region of interest in the satellite cloud image based on the solar elevation angle, azimuth angle, and cloud motion vector. The process is as follows: S2.2.1 First, calculate the horizontal and vertical coordinates L of the pixels in the corrected cloud image. x L y : , , In the formula, C p The x-coordinate of a pixel in a satellite cloud image, C off C represents the offset of the horizontal axis. fac L represents the scaling factor of the x-axis. p The vertical coordinate of a pixel in a satellite cloud image, L off L represents the offset of the vertical axis. fac The scaling factor for the ordinate; Then calculate the latitude and longitude L of each pixel in the satellite cloud image. on L at : ， ， ， ， ， ， ， ， In the formula, h d D represents the altitude of the satellite. ea and D eb For the Earth's major and minor axes, Indicates direct path gain. Indicates northbound gain. This indicates the total altitude of the satellite to the monitoring point on the Earth's surface. Indicates eastward gain. Indicates vertical gain. This represents the synthesis gain, which is used to reconstruct the image based on the latitude and longitude of each pixel, resulting in a corrected satellite cloud image. S2.2.2 Calculate the maximum latitude and longitude variation of cloud cluster movement: ， ， ， In the formula, d max v represents the maximum distance the cloud can travel. max The maximum speed of the cloud formation. The input feature time scale is R, where R is the Earth's radius. This represents the change in latitude. This represents the change in longitude. Determine the boundaries of key areas of interest in the cloud map based on changes in latitude and longitude: ， In the formula, X0 is the latitude of the power station to be measured, and Y0 is the longitude of the power station to be measured. The minimum latitude of the region, The maximum latitude of the region, Minimum longitude of the region The maximum longitude of the region can be used to determine the maximum geographical range of satellite cloud image input through the region's boundaries; S2.

3. Perform resampling and grayscale normalization conversion on the preprocessed multi-source cloud image.

4. The photovoltaic power generation prediction method based on multi-source cloud map information and convolutional visual Transformer according to claim 1, characterized in that, The process of step S3 is as follows: On the cloud image at given adjacent time points, the relationship between the displacement of each cloud pixel and the binomial coefficient is derived through binomial transformation and constraint conditions. The displacement of each pixel along different directions is estimated, and optical flow maps along the X and Y axes are obtained. A binomial expansion is performed, and a binomial approximation is made in a local window with each pixel on the cloud image as the center. ， In the formula, r1, r2, r3, r4, r5, and r6 are the first, second, third, fourth, fifth, and sixth weight coefficients of the iterative operation, and m and n represent the horizontal and vertical coordinates of the cloud map pixels, which are obtained using the weighted least squares method. Through the The expansion yields the relationship between the grayscale value and the coordinates of each pixel in the cloud map. The grayscale value of the pixel is then converted into a spatial coordinate representation, and the displacement vectors in the X and Y directions are obtained to obtain the optical flow map.

5. The photovoltaic power generation prediction method based on multi-source cloud map information and convolutional visual Transformer according to claim 1, characterized in that, The process of step S4 is as follows: S4.1 In the convolutional visual Transformer, the convolutional operation layer embeds multi-source optical flow map sequences through convolutional operations to extract local features, as detailed below: In the formula, F conv The output feature map represents the convolution operation, Conv() represents the convolution function, O represents the multi-source optical flow map tensor, E represents the kernel size, S represents the stride of the convolution operation, and P represents padding. The convolution operation process is shown below: In the formula, i and j represent the index positions of the convolution operation on the multi-source optical flow graph tensor, c represents the output channel, and F conv (i,j,c) represents the value of the output feature map of the convolution operation at position (i,j) and channel c, e and l represent the two-dimensional position information of each weight element in the convolution kernel, E(e,l) represents the weight of the convolution kernel at (e,l), b is the bias term, and O The part representing the multi-source optical flow map convolution operation flattens the embedded feature map into a sequence, as follows: , In the formula, F flat Flatten() is the flattened sequence; S4.2 The operation process of the multi-head self-attention layer and normalization layer in the convolutional visual Transformer is as follows: ， ， ， ， In the formula, Q represents the query vector, K represents the key vector, V represents the value vector, ConProj() represents the convolution projection operation, and W... Q W K W V Let d represent the learnable weight matrices of the query vector, query vector, and value vector, respectively, where A represents the attention weight matrix score, and d represents the value vector. k This represents the dimension of the key vector, MultiHead() represents the multi-head self-attention function, Concat() represents the concatenation function, and head... i Let W represent the i-th attention head, h represent the number of attention heads, and W represent the number of attention heads. o This represents the output weight matrix, and LayerNorm() represents the layer normalization function; S4.3 The operation process of the feedforward neural network layer in the convolutional vision Transformer is as follows: ， ， In the formula, FFN() represents the feedforward neural network function, W1 and W2 are the first and second feedforward weight matrices, respectively, and b1 and b2 are the first and second feedforward bias terms, respectively. S4.4 In the convolutional visual Transformer, the pooling layer obtains multi-source cloud map feature vectors through pooling operations, as follows: , In the formula, X ne This represents the feature vector of a multi-source cloud map, and Pool() represents the pooling function.

6. The photovoltaic power generation prediction method based on multi-source cloud map information and convolutional visual Transformer according to claim 1, characterized in that, Step S5 includes: S51. The input layer receives meteorological information data with a time resolution of t, photovoltaic unit operation data, and multi-source cloud image feature vectors from the past T hours, and maps them to a pre-specified unified embedding dimension to achieve multimodal feature alignment. Then, they are concatenated along the time dimension to form fused features, as follows: , In the formula, X np Represents meteorological information data, X pv X represents the operating data of the photovoltaic unit. ne X represents the feature vector of a multi-source cloud map. 0 Indicates fused features; Embed() represents the mapping embedding function; Concat represents the concatenation function. S52. The decomposition layer decomposes the fused features into trend terms and residual terms, as follows: , In the formula, X R X represents the trend term. T Represents the residual term; Decompose() represents the moving average kernel decomposition function. S53. Input the trend term and residual term into the trend term linear layer and residual term linear layer respectively, and extract short-term trend features and periodic change features respectively. The process is as follows: , , In the formula, H R H represents short-term trend characteristics. T W represents the characteristic of periodic change. R W T These represent the linear layer of the trend term and the linear layer of the residual term, respectively. S54. The extracted short-term trend features and periodic change features are fed into the merging layer for fusion to reconstruct the time series change trend. Finally, the output layer generates the prediction result. The process is as follows: , In the formula, Y represents the final prediction result, and FC() represents the fully connected function of the output layer.

7. A photovoltaic power generation prediction system based on multi-source cloud image information and convolutional visual Transformer, running the photovoltaic power generation prediction method based on multi-source cloud image information and convolutional visual Transformer according to any one of claims 1-6, characterized in that, The ultra-short-term photovoltaic power prediction system includes: The data acquisition and preprocessing module collects meteorological information data, photovoltaic unit operation data, ground-based cloud map data, and satellite cloud map data. Among them, the meteorological information data includes temperature, irradiance, and solar azimuth angle. The module preprocesses the numerical input data. The multi-source cloud image processing module performs distortion correction on the ground-based cloud image and reconstructs the satellite cloud image using geometric correction techniques. Based on the solar azimuth angle and cloud cluster motion vector, it determines the key areas of interest in the reconstructed satellite cloud image, achieving spatial feature dimensionality reduction. Here, the cloud cluster motion vector refers to the maximum normal motion velocity of the cloud cluster. The preprocessed ground-based cloud image and satellite cloud image are resampled and grayscale standardized to obtain a standardized multi-source cloud image dataset. The optical flow feature map generation module uses a multi-source cloud map dataset and employs the optical flow method to analyze the motion vector field of pixels in the multi-source cloud map at adjacent time points. It obtains the motion direction and velocity information of the pixels in the multi-source cloud map as time changes, generates an optical flow feature map that represents the dynamic evolution of the cloud map, and uses the optical flow feature map as a multi-source optical flow map. The feature extraction module encodes the multi-source optical flow map using a convolutional visual Transformer to extract multi-source cloud map feature vectors that represent cloud motion and occlusion information. The convolutional visual Transformer consists of multiple layers of encoding blocks connected in sequence. The processed multi-source optical flow map is input into each encoding block of the convolutional visual Transformer in sequence. Each encoding block consists of a convolutional operation layer, a multi-head self-attention layer, a normalization layer, a feedforward neural network layer, and a pooling layer. The power prediction module is used to construct a photovoltaic power prediction model based on DLinear. This model includes an input layer, a decomposition layer, a trend term linear layer, a residual term linear layer, a merging layer, and an output layer connected in sequence. The input layer receives meteorological information data with a time resolution of t, photovoltaic unit operation data, and multi-source cloud image feature vectors from the past T hours, and maps them to a pre-specified unified embedding dimension to achieve multimodal feature alignment. Then, they are concatenated along the time dimension to form a fused feature. The decomposition layer decomposes the fused feature into a trend term and a residual term, which are input to the trend term linear layer and the residual term linear layer, respectively, to extract short-term trend features and periodic change features. The short-term trend features and periodic change features are fed into the merging layer for fusion to reconstruct the time series trend. Finally, the output layer generates the prediction result, achieving high-precision photovoltaic power generation prediction.

8. A computer storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements any of the ultra-short-term photovoltaic power prediction methods based on multi-source cloud maps and convolutional visual Transformers as described in claims 1-6.

9. 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 instructions for any of the ultra-short-term photovoltaic power prediction methods based on multi-source cloud maps and convolutional visual Transformers as described in claims 1-6.

Images