Photovoltaic power generation prediction method and device

Abstract

The invention provides a photovoltaic power generation prediction method and device, and the method comprises the steps: collecting an atmospheric motion vector diagram (including a lower-layer wind field, a middle-layer wind field and an upper-layer wind field), and determining a target region and an adjacent search region. The method comprises the following steps: extracting cloud factors from a search area, and converting the cloud factors into cloud influence factors of each wind field on the cloud amount change of a target area; carrying out support vector machine (SVM) regression training by taking the cloud influence factor and the cloud factor as inputs and taking the cloud cover and the irradiance prediction value of the target area as outputs; and evaluating the prediction performance of the trained SVM and then deploying the SVM to predict photovoltaic power generation. Through extraction and analysis of an atmospheric motion vector diagram (AMV) and combination of SVM regression training, accurate prediction of the cloud cover and irradiance of a target area is realized, and then prediction of photovoltaic power generation is realized.

Description

Technical Field

[0001] The present application relates to the field of photovoltaic power generation prediction, and in particular to a method and device for photovoltaic power generation prediction. Background Art

[0002] Due to the overuse of fossil fuels, the world is facing serious natural resource depletion and climate change challenges. To meet these challenges, it has become a top priority to adopt more alternative energy sources, namely "renewable energy", to replace fossil fuels. In recent years, the penetration of renewable energy into the main power grid has been gradually increasing, and it is expected that this penetration will grow at a faster rate by 2030. Large-scale photovoltaic power stations (PV power stations) have been widely used around the world as an important part of renewable energy. Countries such as Germany, China and the United States have installed a large number of large-scale PV power stations.

[0003] However, as the number of large-scale PV power plants continues to increase, the proportion of photovoltaic power generation in the total power grid is also increasing. The power generation of these PV power plants is widely affected by meteorological conditions, especially the number and movement of clouds, which can cause large fluctuations in photovoltaic power generation. Therefore, large-scale PV operators need a more accurate way to predict how much energy their PV power plants can generate in order to conduct effective photovoltaic transactions in the energy market. Summary of the invention

[0004] The purpose of this application is to overcome the defects existing in the above-mentioned prior art and provide a method and device for photovoltaic power generation prediction.

[0005] The present application provides a method for photovoltaic power generation prediction, comprising:

[0006] Collecting an atmospheric motion vector diagram, wherein the atmospheric motion vector diagram includes a lower-layer wind field, a middle-layer wind field, and an upper-layer wind field;

[0007] Determining a target area for photovoltaic power generation prediction and a plurality of search areas adjacent to the target area;

[0008] Extracting cloud factors from the search area, and converting the cloud factors into cloud impact factors of the lower-level wind field, the middle-level wind field, and the upper-level wind field respectively affecting the cloud amount change in the target area;

[0009] Using the cloud impact factor and the cloud factor as input data and the predicted values ​​of cloud cover and irradiance in the target area as output data, support vector machine regression training is performed;

[0010] The trained support vector machine is deployed after evaluating the prediction performance of the cloud cover and irradiance prediction values ​​of the target area to predict photovoltaic power generation.

[0011] Optionally, after collecting the atmospheric motion vector map, extracting image features of the lower-layer wind field, the middle-layer wind field and the upper-layer wind field includes:

[0012] Corresponding to the lower-level wind field, the middle-level wind field and the upper-level wind field, red, green and blue pixels are separated from the atmospheric motion vector diagram and stored in respective color libraries;

[0013] Based on the atmospheric motion vector map of each monochrome pixel, a preselected pixel is randomly selected;

[0014] Calculating the distance between the selected pre-selected pixel and other pixels, taking the adjacent pixel with the smallest distance as the next pre-selected pixel, and storing it in a sub-library of the monochrome pixel color library;

[0015] Calculate the center coordinates of the pixels in the sub-library, and repeat the previous step by taking the next pre-selected pixel as the pre-selected pixel;

[0016] If the distance between the pre-selected pixel and the next selected pixel is greater than or equal to a preset threshold, or the next pre-selected pixel is farther from the center coordinate than the pre-selected pixel, stop repeating the previous step.

[0017] Optionally, determining a target area for photovoltaic power generation prediction and a plurality of search areas adjacent to the target area includes:

[0018] Specify a target area;

[0019] Determine a plurality of search areas adjacent to the target area according to the size and shape of the target area;

[0020] The coordinates, width and height of the target area and the search area are determined.

[0021] Optionally, the trained support vector machine is evaluated for prediction performance of cloud cover and irradiance prediction values ​​of a target area, including:

[0022] The performance of the trained SVM was evaluated using root mean square error, mean relative error, and / or coefficient of determination.

[0023] Optionally, deploying a trained support vector machine for photovoltaic power generation prediction includes:

[0024] Integrate the trained support vector machine into the photovoltaic power generation prediction system;

[0025] Receive satellite image data in real time, and then use support vector machine to predict the cloud cover and irradiance of the target area, and then predict photovoltaic power generation.

[0026] The present application also provides a photovoltaic power generation prediction device, comprising:

[0027] An acquisition module collects an atmospheric motion vector map, wherein the atmospheric motion vector map includes a lower-layer wind field, a middle-layer wind field, and an upper-layer wind field;

[0028] A setting module is provided to determine a target area for photovoltaic power generation prediction and a plurality of search areas adjacent to the target area;

[0029] An extraction module extracts cloud factors from the search area and converts the cloud factors into cloud influence factors of the lower wind field, the middle wind field and the upper wind field respectively affecting the cloud amount change in the target area;

[0030] A training module, using the cloud impact factor and the cloud factor as input data and the cloud amount and irradiance prediction value of the target area as output data, to perform support vector machine regression training;

[0031] The prediction module is deployed after evaluating the prediction performance of the trained support vector machine for the predicted values ​​of cloud cover and irradiance in the target area to predict photovoltaic power generation.

[0032] Optionally, after the acquisition module acquires the atmospheric motion vector map, the image features of the lower wind field, the middle wind field and the upper wind field are extracted, including:

[0033] Corresponding to the lower-level wind field, the middle-level wind field and the upper-level wind field, red, green and blue pixels are separated from the atmospheric motion vector diagram and stored in respective color libraries;

[0034] Based on the atmospheric motion vector map of each monochrome pixel, a preselected pixel is randomly selected;

[0035] Calculating the distance between the selected pre-selected pixel and other pixels, taking the adjacent pixel with the smallest distance as the next pre-selected pixel, and storing it in a sub-library of the monochrome pixel color library;

[0036] Calculate the center coordinates of the pixels in the sub-library, and repeat the previous step by taking the next pre-selected pixel as the pre-selected pixel;

[0037] If the distance between the pre-selected pixel and the next selected pixel is greater than or equal to a preset threshold, or the next pre-selected pixel is farther from the center coordinate than the pre-selected pixel, stop repeating the previous step.

[0038] Optionally, the setting module determines a target area for photovoltaic power generation prediction, and a plurality of search areas adjacent to the target area, including:

[0039] Specify a target area;

[0040] Determine a plurality of search areas adjacent to the target area according to the size and shape of the target area;

[0041] The coordinates, width and height of the target area and the search area are determined.

[0042] Optionally, the prediction module evaluates the prediction performance of the cloud cover and irradiance prediction values ​​of the target area on the trained support vector machine, including:

[0043] The performance of the trained SVM was evaluated using root mean square error, mean relative error, and / or coefficient of determination.

[0044] Optionally, the prediction module deploys a trained support vector machine to perform photovoltaic power generation prediction, including:

[0045] Integrate the trained support vector machine into the photovoltaic power generation prediction system;

[0046] Receive satellite image data in real time, and then use support vector machine to predict the cloud cover and irradiance of the target area, and then predict photovoltaic power generation.

[0047] The beneficial effects of this application are:

[0048] The present application provides a method for photovoltaic power generation prediction, including: collecting an atmospheric motion vector diagram, the atmospheric motion vector diagram including a lower wind field, a middle wind field, and an upper wind field; determining a target area for photovoltaic power generation prediction, and a plurality of search areas adjacent to the target area; extracting cloud factors from the search area, and converting the cloud factors into cloud influence factors that affect the cloud cover change in the target area respectively by the lower wind field, the middle wind field, and the upper wind field; using the cloud influence factors and the cloud factors as input data, and the predicted values ​​of cloud cover and irradiance in the target area as output data, to perform support vector machine regression training; and evaluating the prediction performance of the predicted values ​​of cloud cover and irradiance in the target area for the trained support vector machine, and then deploying it to predict photovoltaic power generation. The present application realizes accurate prediction of cloud cover and irradiance in the target area by extracting and analyzing the atmospheric motion vector diagram (AMV) and combining it with support vector machine (SVM) regression training, thereby realizing accurate prediction of photovoltaic power generation. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 It is a schematic diagram of the process of photovoltaic power generation prediction in this application;

[0050] Figure 2 A schematic diagram of a single AMV vector in this application;

[0051] Figure 3 A schematic diagram of the target area and search area of ​​the AMV satellite image in this application;

[0052] Figure 4 Comparison chart of cloud cover measured and predicted by the model in this application;

[0053] Figure 5 This is a comparison chart of the irradiance measured and predicted by the model in this application. DETAILED DESCRIPTION

[0054] The exemplary embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that various forms of implementing the present disclosure should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided in order to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0055] Please refer to Figure 1 As shown, the present application provides a method for photovoltaic power generation prediction, the steps of which include:

[0056] S101, collecting an atmospheric motion vector map, wherein the atmospheric motion vector map includes a lower-level wind field, a middle-level wind field, and an upper-level wind field.

[0057] Collect available satellite imagery.

[0058] The satellite images used in this application include atmospheric motion vector images, cloud analysis images, and irradiance images.

[0059] The atmospheric motion vector image (AMV) image consists of red, green and blue colors, representing the lower, middle and upper wind fields respectively.

[0060] Cloud analysis images include cloud quantity, cloud shape and cloud thickness images.

[0061] Irradiance images are based on the amount of light reflected from the ground and can be used together with module temperature to predict photovoltaic (PV) power generation.

[0062] Taking the atmospheric motion vector image as an example, after the pixels of the AMV satellite image are saved in the corresponding library, the AMV image is further extracted to obtain red, green and blue AMV images.

[0063] Taking the extraction of blue AMV as an example, the specific process is as follows:

[0064] Extract each single blue AMV from the blue AMV repository. Randomly select a pixel from the blue AMV repository and calculate the distance between the selected pixel and other pixels in the blue AMV repository. After the distance calculation is completed, store the neighboring pixel with the smallest distance into a sub-repository of a single blue AMV and designate that pixel as the next selection point.

[0065] Each time a new pixel is stored in a single blue AMV sub-bank, the center coordinates of the pixel in the single blue AMV sub-bank are calculated.

[0066] Then, the same process continues by considering the following two constraints:

[0067] 1. The distance between the pre-selected pixel and the next selected pixel should be less than the preset threshold;

[0068] 2. The next selected pixel is closer to the center coordinate when it is smaller than the preset threshold.

[0069] If no other pixels in the blue AMV library satisfy these two constraints, the extraction of a single blue AMV is completed, and the sub-library of a single blue AMV is completed.

[0070] Please refer to Figure 2 As shown in Figure 1, the head, body, and tail pixels are found from the sub-library of a single blue AMV. In this step, the two farthest pixels are taken from the sub-library to assign the head and tail, and their midpoints are calculated (the purple part in the figure).

[0071] Except for the two selected pixels, the pixel farthest from the midpoint in the sub-library is assigned as the body of a single blue AMV. Then, in the distance comparison between the whole body and the two selected pixels, the pixel with the shorter distance is assigned as the head and the pixel with the longer distance is assigned as the tail.

[0072] A wind direction vector d=(dx, dy) is calculated from the body to the tail, and the wind direction angle is calculated as θ.

[0073] Continue the above process for the red and blue sub-libraries.

[0074] Count the total number of pixels in the sub-library of a single blue AMV, then calculate the midpoint between the head and body pixels. After that, calculate the distance between the midpoint and each pixel in the sub-library of a single blue AMV.

[0075] If the number of pixels within the distance threshold from the midpoint is greater than 4 pixels, the wind speed is calculated starting from 25m / s, otherwise the wind speed is calculated starting from 5m / s. Finally, the remaining wind speed bars (long bars 5m / s, short bars 2m / s) are considered to calculate the wind speed.

[0076] S102: Determine a target area for photovoltaic power generation prediction and a plurality of search areas adjacent to the target area.

[0077] Specify a target area and multiple search areas.

[0078] The target area is where you want to predict photovoltaic power generation. The target area has a target coordinate (x target ,y target ), width w target , height h target .

[0079] The search area is the adjacent area of ​​the target area. The i-th search area has search coordinates (x search,i ,y search,i ), width w search , height h search .

[0080] S103: extract cloud factors from the search area, and convert the cloud factors into cloud impact factors of the lower-level wind field, the middle-level wind field, and the upper-level wind field that respectively affect the cloud amount change in the target area.

[0081] Characterize the cloud impact factor of each search area on the target area.

[0082] Since there are three colors of AMV, k = R, G, or B. According to the color representation and the index of the search area, the wind direction, speed and angle of a single AMV with color k in the i-th search area are represented as d i,k ={dx(i,k),dy(i,k)},V i,k (m / s), and θ i,k .

[0083] Assume that the movement of clouds roughly follows the wind direction and speed.

[0084] However, since it can be inferred that the speeds of the cloud and wind are different, the optimal ratio of the cloud speed to the wind speed will be found. α* represents the optimal ratio.

[0085] Another assumption: the formation and disappearance of new clouds are not considered.

[0086] like Figure 3 As shown, the overall AMV image of the target area (magenta box) and the search area (cyan box) is displayed.

[0087] In the i-th search area with color k, a set of average AMV parameters is calculated through multiple iterations, expressed as

[0088] With this set of parameters Ω i,k , we can predict the movement of the ith search area after t minutes. The center coordinates of the ith search area are updated as follows:

[0089]

[0090] Among them, (x i,k (t), y i,k (t)) and ε (m) are the newly moved center coordinates of the i-th search area with color k after t minutes and the ground resolution of 1 pixel, respectively.

[0091] In addition, if w target =w search and htarget =h search , then the sharing amount between the target area and the mobile search area is calculated as:

[0092]

[0093] Where A=|x i,k (t)-x target |,B=|y i,k (t)-y target |.

[0094] Cloud factors such as the number of clouds and cloud shapes are calculated in both the target area and the search area. i and S i Represent the average cloud amount and shape of the ith region respectively. The three wind fields and cloud shapes (1 to 10) of R, G, and B should match in height. The matching rules are R = {1, 2}, G = {3, 4}, B = {5, ..., 10}. According to the matching rules, only one of the shared quantities is activated, that is, if S i =1, then ζ i,R (t) is activated, but ζ i,G (t) = ζ i,B (t)=0.

[0095] Finally, the cloud impact factor of the i-th search area on the target area after t minutes can be calculated as follows:

[0096]

[0097] The total cloud impact factor on the target area after t minutes is calculated as

[0098]

[0099] Where I is the total number of search areas.

[0100] S104, performing support vector machine regression training with the cloud impact factor and the cloud factor as input data and the predicted cloud amount and irradiance of the target area as output data.

[0101] Train the support vector machine for regression.

[0102] Based on the support vector machine (SVM) regression scheme, a large amount of historical satellite image data is used to train the photovoltaic power generation prediction model. Before solving the training and testing process, some parameters are defined to define a given set of N input and output data pairs as D = {(x1, y1), (x2, y2), ..., (x N ,y N )}, X={x1, x2, …, x N} and Y = {y1, y2, ..., y N} represent the collection of input and output data respectively.

[0103] In X, x i =[x i,1 ,x i,2 , …x i,K ] T is the i-th input column vector, with a total of K variables. Using D training ={(x1,y1),(x2,y2),…,(x M ,y M )}, indicating that the total number of training input and output data is M.

[0104] The remaining (NM) input and output data are used for testing, and the test data set is denoted as D test ={(x M+1 ,y M+1 ), (x M+2 ,y M+2 ),…,(x N ,y N )}.

[0105] There is a complex nonlinear relationship between the input and output data. This complex nonlinear relationship can be expressed by y i =g(x i ), therefore, the main goal of machine learning (ML) is to find the most similar function f(x i )≈g(x i In this application, we will use SVM regression to find a function f(x i ).

[0106] In the SVM regression model, the regression function is expressed as

[0107] f(x i )=w T ·x i +b

[0108] where x i , w∈R K , b∈R, where w T ·x i For w and x i The inner product of .

[0109] Finding the optimal w is the main goal of SVM regression. The optimization problem of finding the optimal w is formulated as:

[0110]

[0111] subject to:y i -w T ·x i-b≤ε+ξ i ,Vi,

[0112]

[0113] Among them (ξ i ,ξ i * ), ε and C are slack variables, is a precision parameter, and is a constant that determines the trade-off between the flatness of f and the tolerance of ε.

[0114] Using a Lagrangian function and a dual optimization problem, the optimal coefficient vector is obtained:

[0115]

[0116] where β i and β i * is the Lagrange multiplier.

[0117] For nonlinear SVM regression model, kernel function method is introduced. Kernel Φ(·) transforms K-dimensional input data vector x i Expanded to a K+D-dimensional input data vector, Φ(x i )∈R K+D .

[0118] Therefore, the nonlinear SVM regression function is expressed as:

[0119] f(x i )=w T ·Φ(x i )+b

[0120] Among them, Φ(x i ), w∈R K+D , b∈R, the optimal coefficient vector is:

[0121]

[0122] Finally, the optimal nonlinear SVM regression function is expressed as:

[0123]

[0124] The radial basis function (RBF) kernel is:

[0125] k(x i ,x)=exp{-||x i -x|| / 2σ 2},

[0126] There are two prediction targets. The first is to predict the cloud cover in the target area in the future. The other is to predict the irradiance in the target area in the future.

[0127] Depending on the prediction target, the input and output datasets need to be configured differently for ML training.

[0128] In the cloud cover prediction, the input data vector x C is configured as a set of current cloud cover in the search area C and the total cloud impact factor Γ(t) for the t-minute forecast horizon. Therefore, the input column vector is configured as:

[0129] x C =[C,Γ(t)] T

[0130] The dimension of the vector is I+1, where I is the total number of search regions.

[0131] Set the cloud cover of the target area after t minutes to a single output value, y C =C target (t). Cloud cover can be used to determine the clear sky index η, which is defined as the irradiance Ψ over the target area. target (t) and clear sky irradiance Ψ clear (t) ratio. Cloud cover C target (t) is approximately linearly related to the clear sky index η, and the formula is:

[0132]

[0133] Therefore, the irradiance of the target area after t minutes can be predicted as:

[0134]

[0135] In the irradiance prediction, the sun altitude angle ω is additionally inserted, which is calculated at a specific date and time, and the current irradiance Ψ of the target area is target (0) Insert into the predicted cloud cover x explained above C The input data vector of , because the solar altitude angle and current irradiance are the reference data for predicting the irradiance value.

[0136] Therefore, the input column vector of irradiance for prediction is configured as:

[0137] x ψ =[C,Γ(t),ω,Φ target (0)] T ,

[0138] Its dimension is I+3, and it sets the irradiance of the target area after T minutes to a single output value, y Ψ =Ψ target (T).

[0139] S105, evaluating the prediction performance of the trained support vector machine for the predicted values ​​of cloud cover and irradiance in the target area and then deploying them to predict photovoltaic power generation.

[0140] Finally, according to the prediction accuracy indicators, such as root mean square error (RMSE), mean relative error (MRE) and determination coefficient R 2 , the performance of the proposed SVM-based cloud cover and irradiance prediction models are evaluated respectively, and the definitions of these indicators are as follows:

[0141]

[0142] in

[0143]

[0144] By inputting the cloud cover and irradiance data of the day to be predicted into the trained prediction model, we can obtain the cloud cover prediction in minutes based on the support vector machine model and the irradiance prediction value in minutes based on the support vector machine model for the predicted day.

[0145] The cloud cover and irradiance prediction results output by the prediction model are as follows: Figure 4 , Figure 5 As shown, Figure 4 Comparison graph of measured and predicted cloud cover based on the proposed SVM-based model; Figure 5 The comparison of measured and predicted irradiance based on the proposed support vector machine-based model is shown in Table 1. The cloud cover prediction results for different prediction time ranges with a time scale of minutes are shown in Table 1:

[0146] Table 1 Cloud cover prediction results for different prediction time ranges

[0147] Range (minutes) / Indicator RMSE MRE(%) R2(0～1) 15 5.7368 5.2853 0.9745 30 8.4465 7.6510 0.9444 60 10.8875 9.5643 0.9410 90 14.9760 13.2469 0.8261 120 16.8574 17.1598 0.7976 150 17.6563 17.9959 0.7697 180 19.4637 18.1018 0.7009 240 23.8976 25.0497 0.6123 300 24.5989 26.5645 0.5759

[0148] The irradiance prediction results for different prediction time ranges with minutes as the time scale are shown in Table 2:

[0149] Table 2 Irradiance prediction results for different prediction time ranges

[0150]

[0151]

[0152] The above are only specific embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for photovoltaic power generation prediction, characterized in that: include: Collecting an atmospheric motion vector diagram, wherein the atmospheric motion vector diagram includes a lower-layer wind field, a middle-layer wind field, and an upper-layer wind field; Determining a target area for photovoltaic power generation prediction and a plurality of search areas adjacent to the target area; Extracting cloud factors from the search area, and converting the cloud factors into cloud impact factors of the lower-level wind field, the middle-level wind field, and the upper-level wind field respectively affecting the cloud amount change in the target area; Using the cloud impact factor and the cloud factor as input data and the predicted values ​​of cloud cover and irradiance in the target area as output data, support vector machine regression training is performed; The trained support vector machine is deployed after evaluating the prediction performance of the cloud cover and irradiance prediction values ​​of the target area to predict photovoltaic power generation.

2. The method for photovoltaic power generation prediction according to claim 1, characterized in that: After collecting the atmospheric motion vector map, the image features of the lower wind field, the middle wind field and the upper wind field are extracted, including: Corresponding to the lower-level wind field, the middle-level wind field and the upper-level wind field, red, green and blue pixels are separated from the atmospheric motion vector diagram and stored in respective color libraries; Based on the atmospheric motion vector map of each monochrome pixel, a preselected pixel is randomly selected; Calculating the distance between the selected pre-selected pixel and other pixels, taking the adjacent pixel with the smallest distance as the next pre-selected pixel, and storing it in a sub-library of the monochrome pixel color library; Calculate the center coordinates of the pixels in the sub-library, and repeat the previous step by taking the next pre-selected pixel as the pre-selected pixel; If the distance between the pre-selected pixel and the next selected pixel is greater than or equal to a preset threshold, or the next pre-selected pixel is farther from the center coordinate than the pre-selected pixel, stop repeating the previous step.

3. The method for photovoltaic power generation prediction according to claim 1, characterized in that: Determine a target area for photovoltaic power generation prediction, and multiple search areas adjacent to the target area, including: Specify a target area; Determine a plurality of search areas adjacent to the target area according to the size and shape of the target area; The coordinates, width and height of the target area and the search area are determined.

4. The method for photovoltaic power generation prediction according to claim 1, characterized in that: The trained support vector machine is used to evaluate the prediction performance of the cloud cover and irradiance prediction values ​​of the target area, including: The performance of the trained SVM was evaluated using root mean square error, mean relative error, and / or coefficient of determination.

5. The method for photovoltaic power generation prediction according to claim 1, characterized in that: Deployment of trained support vector machines for photovoltaic power generation prediction, including: Integrate the trained support vector machine into the photovoltaic power generation prediction system; Receive satellite image data in real time, and then use support vector machine to predict the cloud cover and irradiance of the target area, and then predict photovoltaic power generation.

6. A photovoltaic power generation prediction device, characterized in that: include: An acquisition module collects an atmospheric motion vector map, wherein the atmospheric motion vector map includes a lower-layer wind field, a middle-layer wind field, and an upper-layer wind field; A setting module is provided to determine a target area for photovoltaic power generation prediction and a plurality of search areas adjacent to the target area; An extraction module extracts cloud factors from the search area and converts the cloud factors into cloud influence factors of the lower wind field, the middle wind field and the upper wind field respectively affecting the cloud amount change in the target area; A training module, using the cloud impact factor and the cloud factor as input data and the cloud amount and irradiance prediction value of the target area as output data, to perform support vector machine regression training; The prediction module is deployed after evaluating the prediction performance of the trained support vector machine for the predicted values ​​of cloud cover and irradiance in the target area to predict photovoltaic power generation.

7. The photovoltaic power generation prediction device according to claim 6, characterized in that: After the acquisition module collects the atmospheric motion vector map, the image features of the lower wind field, the middle wind field and the upper wind field are extracted, including: Corresponding to the lower-level wind field, the middle-level wind field and the upper-level wind field, red, green and blue pixels are separated from the atmospheric motion vector diagram and stored in respective color libraries; Based on the atmospheric motion vector map of each monochrome pixel, a preselected pixel is randomly selected; Calculating the distance between the selected pre-selected pixel and other pixels, taking the adjacent pixel with the smallest distance as the next pre-selected pixel, and storing it in a sub-library of the monochrome pixel color library; Calculate the center coordinates of the pixels in the sub-library, and repeat the previous step by taking the next pre-selected pixel as the pre-selected pixel; If the distance between the pre-selected pixel and the next selected pixel is greater than or equal to a preset threshold, or the next pre-selected pixel is farther from the center coordinate than the pre-selected pixel, stop repeating the previous step.

8. The photovoltaic power generation prediction device according to claim 6, characterized in that: The setting module determines a target area for photovoltaic power generation prediction and a plurality of search areas adjacent to the target area, including: Specify a target area; Determine a plurality of search areas adjacent to the target area according to the size and shape of the target area; The coordinates, width and height of the target area and the search area are determined.

9. The photovoltaic power generation prediction device according to claim 6, characterized in that: The prediction module evaluates the prediction performance of the cloud cover and irradiance prediction values ​​of the target area on the trained support vector machine, including: The performance of the trained SVM was evaluated using root mean square error, mean relative error, and / or coefficient of determination.

10. The photovoltaic power generation prediction device according to claim 6, characterized in that: The prediction module deploys a trained support vector machine to predict photovoltaic power generation, including: Integrate the trained support vector machine into the photovoltaic power generation prediction system; Receive satellite image data in real time, and then use support vector machine to predict the cloud cover and irradiance of the target area, and then predict photovoltaic power generation.

Images