A photovoltaic power short-term prediction method

Abstract

The application discloses a photovoltaic power short-term prediction method, comprising the following steps: calculating the location of the shading area of different height clouds when the solar radiation at the photovoltaic station is shaded at the prediction time point; detecting whether each height cloud layer at the prediction time point moves to the corresponding height layer shading area, if the detection result is yes, marking the corresponding height cloud layer as a candidate cloud cluster, and generating a scanning area for scanning the candidate cloud cluster; tracking the cloud layer height of the candidate cloud cluster in the scanning area through the radar, and constructing a solar shading field at the prediction time point based on the tracked cloud layer height; inputting the solar shading field and multi-modal perception data into a prediction model, and outputting the power output of the photovoltaic station at each prediction time point. The solar shading field is constructed, multi-source observation data is uniformly mapped to the solar shading field facing the photovoltaic shading risk, shading features are guided to be fused in time, space and height based on the solar shading field, and a power prediction value in a future time limit is predicted.

Description

Technical Field

[0001] This invention belongs to the field of new energy technology, and more specifically, this invention relates to a method for short-term prediction of photovoltaic power. Background Technology

[0002] As the installed capacity of photovoltaics in new energy power systems continues to increase, photovoltaic power prediction plays an increasingly prominent role in grid dispatching, load balancing, energy storage control, power trading, and power fluctuation early warning.

[0003] By modeling models such as autoregressive models, support vector machines, random forests, and neural networks using a large amount of historical photovoltaic power, future photovoltaic power is predicted. However, the above photovoltaic power prediction mainly relies on the variation pattern of historical power and does not consider the fluctuation of photovoltaic power caused by changes in solar irradiance due to cloud cover. When clouds block direct solar irradiance, photovoltaic power may rise or fall rapidly within a few minutes to tens of minutes. Therefore, the existing photovoltaic power prediction has the problem of insufficient prediction accuracy. Summary of the Invention

[0004] This invention provides a method for short-term photovoltaic power prediction, aiming to solve at least one of the above-mentioned problems.

[0005] This invention is implemented as follows: a method for short-term prediction of photovoltaic power, the method being as follows:

[0006] (1) Calculate the area of ​​cloud cover at different altitudes when solar radiation is blocked at the photovoltaic site at the predicted time point;

[0007] (2) Detect whether the cloud layer at the current height at the predicted time point has moved to the occlusion area of ​​the corresponding height layer. If the detection result is yes, mark the cloud layer at the corresponding height as a candidate cloud cluster and generate a scanning area for scanning the candidate cloud cluster.

[0008] (3) The cloud height of candidate cloud clusters is tracked by radar in the scanning area, and the solar shading field at the predicted time point is constructed based on the tracked cloud height.

[0009] (4) Input the solar shading field and multimodal sensing data into the prediction model, and the prediction model outputs the power output of the photovoltaic site at each prediction time point.

[0010] Furthermore, the detection method for whether the cloud layers at each altitude have moved to the occlusion area of ​​the corresponding altitude layer is as follows:

[0011] The current movement direction and speed of cloud layers at each altitude are obtained to form the current motion information of cloud layers at each altitude. Based on the current motion information of cloud layers at each altitude, the estimated position of cloud layers at each altitude at the predicted time point is predicted. If the estimated position of cloud layers at each altitude is located within the corresponding occlusion area, it is determined that the current cloud layers at each altitude have moved to the occlusion area of ​​the corresponding altitude layer.

[0012] Furthermore, the cloud height within the scanning area is tracked using radars A, B, and C. Radar B is used to detect the cloud height directly above the current location and sends the detection data to radars A and C. Radar A is used to detect the cloud height in the near field ahead of the scanning area, and radar C is used to detect the cloud height in the far field ahead of the scanning area.

[0013] Furthermore, the specific process of constructing the solar shading field is as follows:

[0014] Calculate the prediction time point The height at which solar radiation is blocked at the lower photovoltaic site is The center of the cloud cover area In the center of the obstruction Build a mesh in the nearby occluded areas and put the built mesh into the mesh set. In the grid set It is expressed as follows:

[0015] ;

[0016] ;

[0017] in, Indicates the predicted time point The height at which solar radiation is blocked at the lower photovoltaic site is The j-th grid in the area obscured by the cloud, This represents the coordinates of the center of the j-th grid. This represents the distance from the center of the j-th grid to the center of the occlusion. The radius of the obscured area; Indicates the predicted time point The solar altitude angle below; Indicates the predicted time point The azimuth of the sun below.

[0018] Furthermore, the prediction model includes: a shared image patch projector, a numerical shared projector, a feature fusion unit connected to the shared image patch projector and the numerical shared projector, an attention weight module connected to the feature fusion unit, a memory update module connected to the attention weight module, and a prediction module connected to the memory update module, wherein...

[0019] A shared image patch projector is used to map input satellite images and fisheye images into M-dimensional image features;

[0020] Numerical sharing projectors are used to map various types of input sensory data into N-dimensional data features;

[0021] The feature fusion unit is used to aggregate the data features extracted by the numerical sharing projector and the image features extracted by the shared image patch projector into the solar shading field;

[0022] The attention weight module is used to extract fused features through an attention mechanism. Extract and fuse spatial, cloud height, and prediction time-of-flight occlusion features;

[0023] The memory update module is used to fuse the occlusion memory features from the previous moment with the current occlusion features and update the occlusion features at the current moment.

[0024] The prediction module is used for prediction based on occlusion features. Predicting photovoltaic sites at the predicted time point Photovoltaic power .

[0025] Furthermore, the aggregation process is as follows:

[0026] Calculate the contribution weights of various data features and image features to each grid in the solar shading field. Belonging to the aforementioned data features and image features, feature c corresponds to the grid. Contribution weight The specific calculation formula is as follows:

[0027] ;

[0028] in, Representation of features At the predicted time point The estimated location, Representation of features At the current time t, , Representation of features At the predicted time point The estimated altitude, the altitude at the current time t, , These represent scaling factors for position and height, respectively.

[0029] Each feature Grids converging into the solar shading field The specific aggregation formula is as follows:

[0030] ;

[0031] Each grid within all sun-shaded areas Fusion characteristics Fusion characteristics of the solar shading field .

[0032] Furthermore, the attention weight module consists of three levels of attention modules connected in series, each used for feature fusion. The spatial attention module, height attention module, and timeliness attention module are extracted based on the feature distribution in terms of space, height, and timeliness. The input and output residuals of each attention module are concatenated and then input into the normalization layer LayerNorm.

[0033] Furthermore, the shared image patch projector consists of a two-dimensional convolutional layer with a kernel size of 16×16 and a stride of 16, a normalization layer, and an activation layer.

[0034] Furthermore, the numerical sharing projector consists of a linear layer, a normalization layer, and an activation layer.

[0035] Furthermore, the prediction module includes a first fully connected layer and a second fully connected layer, wherein the first fully connected layer is used for predicting photovoltaic power and the second connected layer is used for predicting the trend of photovoltaic power change.

[0036] This invention constructs a solar shading field by taking the future position of the sun, cloud height, and extrapolated position of cloud clusters. It maps multi-source observation data to a solar shading field that addresses the risk of photovoltaic shading. Based on the solar shading field, it guides the fusion of shading characteristics in time, space, and height to predict future power forecasts and power change states. Attached Figure Description

[0037] Figure 1 A flowchart of the photovoltaic power short-term prediction method provided in an embodiment of the present invention;

[0038] Figure 2 This is a schematic diagram illustrating the formation of shading areas at different height levels provided in an embodiment of the present invention;

[0039] Figure 3 This is a schematic diagram illustrating the formation of a scanning region for candidate cloud scanning, provided in an embodiment of the present invention. Detailed Implementation

[0040] The specific embodiments of the present invention will be further described in detail below with reference to the accompanying drawings, so as to help those skilled in the art to have a more complete, accurate and in-depth understanding of the inventive concept and technical solution of the present invention.

[0041] Figure 1A flowchart of a photovoltaic power short-term prediction method provided in an embodiment of the present invention is shown below:

[0042] (1) Calculate the area of ​​cloud cover at different altitudes when solar radiation is blocked at the photovoltaic site at the predicted time point;

[0043] In this embodiment of the invention, the shading area is a three-dimensional inclined tubular region extending in the opposite direction of solar incidence from the photovoltaic site. Different cloud heights correspond to different shading areas. If a cloud at a certain height enters the shading area at that height, it may block direct solar radiation reaching the photovoltaic site. Figure 2 As shown.

[0044] The central location of the photovoltaic site is The current time is The predicted time point is ,and Obtain the predicted time point solar altitude angle at time The horizontal unit vector corresponding to the solar azimuth angle And the current cloud height h, calculate the cloud height h at the predicted time point. When a photovoltaic power station is obstructed, its ground projection obstruction point Approximately:

[0045] ;

[0046] The corresponding horizontal distance is: ;

[0047] Based on the above formulas, it can be seen that the higher the cloud height, the farther the shading point is from the photovoltaic site; the lower the solar altitude angle, the farther the shading point is from the photovoltaic site; and at low solar altitude angles in the early morning and late evening, the impact of far-field clouds on photovoltaic power is enhanced. Obtain the smallest circumcircle covering the entire photovoltaic site, and take the diameter Dsite of the smallest circumcircle. Then the equivalent radius of the photovoltaic site is: Rsite = Dsite / 2. The equivalent radius of the photovoltaic site is used as the shading radius.

[0048] In this embodiment of the invention, radar B detects multiple cloud layers at different altitudes and places the detected cloud layer data into a cloud layer set. Among them, the data of the i-th cloud layer , This represents the height of the cloud base of the i-th cloud layer; This represents the height of the cloud top of the i-th cloud layer; This represents the height of the i-th cloud layer; This represents the thickness of the i-th cloud layer; This represents the backscattering intensity or extinction-related characteristics of the i-th cloud layer.

[0049] For each cloud layer in the cloud set, calculate the occlusion area of ​​the cloud layer at the corresponding height at the predicted time point. .

[0050] (2) Obtain the current movement direction and speed of cloud layers at each altitude to form the current motion information of cloud layers at each altitude;

[0051] In this embodiment of the invention, optical flow is used to track fisheye images of the entire sky to obtain the current movement direction and speed of clouds at different layers, thus forming motion information of different cloud layers. , means as follows:

[0052] .

[0053] (3) Detect whether the cloud layers at the current height at the predicted time point have moved to the occlusion area of ​​the corresponding height layer. If the detection result is yes, mark the corresponding height cloud layer as a candidate cloud cluster and generate a scanning area for scanning the candidate cloud cluster.

[0054] In this embodiment of the invention, the estimated position of each cloud layer at a given altitude is predicted based on the current motion information of each cloud layer at that altitude at the predicted time point. If the estimated position of each cloud layer at that altitude is located within the corresponding occlusion area, then the corresponding cloud layer at that altitude is marked as a candidate cloud cluster. The estimated position of each cloud layer at the predicted time point is expressed as follows:

[0055] ;

[0056] in, This represents the position of the i-th cloud layer at the current time t. The i-th cloud layer at the predicted time point The estimated location, if it satisfies: If , then the i-th cloud layer is identified as a candidate cloud cluster.

[0057] Predicted location Extending forward along the direction of motion, it forms a scanning region of i cloud layers, such as... Figure 3 As shown.

[0058] (4) Track the cloud height of candidate cloud clusters in the scanning area using radar, and construct the solar shading field at the predicted time point based on the tracked cloud height;

[0059] In this embodiment of the invention, cloud height is tracked within the scanning area using radar A, radar B, and radar C. Radar B is used to detect the cloud height directly above the current location and sends the detection data to radar A and radar C. Radar A is used to detect the cloud height in the near field in front of the scanning area, and radar C is used to detect the cloud height in the far field in front of the scanning area.

[0060] Radar A alternately performs solar azimuth RHI scans and sector PPI scans of the scanned area. The solar azimuth RHI scan is primarily used to determine if there are any clouds obstructing the current sunlight path, while the sector PPI scan determines whether there are any obstructing clouds near the sunlight path in the next half hour. Solar Azimuth RHI Scan: RHI Azimuth A z Around the solar azimuth angle A s Settings: A z =A s And A z =A s ±10 ∘ Add A if necessary. z= A s ±20∘ is used to determine whether there are clouds in the sun's incident path and whether the cloud boundaries are close to the sun's path. The PPI of the scanned area sector: The PPI sector center points to the current scanned area; different sector widths can be used under different weather conditions.

[0061]

[0062] PPI elevation angle selectable: 3 ∘ 5 ∘ 8 ∘ ,12 ∘ 20 ∘ 35 ∘ 60 ∘ .

[0063] Radar B employs zenith observation and near-field multi-elevation angle observation. In zenith observation mode, the radar points directly upwards at E1=90°. ∘ Primarily used to acquire the vertical structure of clouds directly overhead, including cloud base height, cloud top height, cloud thickness, multi-layer cloud structure, and cloud intensity. Near-field multi-elevation angle observation: By scanning the near-field range at multiple low elevation angles, it obtains the horizontal distribution and intensity of clouds at different altitudes, supplementing near-surface information not covered by zenith observation. Recommended elevation angles include: 5... ∘ 10 ∘ 20 ∘ 45 ∘ 90 ∘

[0064] Radar C performs far-field scanning in the direction of upstream incoming clouds, including low elevation angle PPI in the direction of upstream main incoming clouds, RHI of upstream candidate clouds, PPI of upstream scanning area, upstream cloud boundary tracking scan, and low-frequency background cloud field update scan, in order to observe in advance the clouds that may affect the photovoltaic site in the next 30 to 60 minutes.

[0065] In this embodiment of the invention, the process of constructing the solar shading field is as follows:

[0066] Calculate the prediction time point The height at which solar radiation is blocked at the lower photovoltaic site is The center of the cloud cover area The specific calculation formula is as follows:

[0067] ;

[0068] At the center of the obstruction Build a local mesh nearby and add the built mesh to the mesh set. In the grid set It is expressed as follows:

[0069] ;

[0070] ;

[0071] in, Indicates the predicted time point The height at which solar radiation is blocked at the lower photovoltaic site is The j-th grid in the area obscured by the cloud, This represents the coordinates of the center of the j-th grid. This represents the distance from the center of the j-th grid to the center of the occlusion. The radius of the obscured area; Indicates the predicted time point The solar altitude angle below; Indicates the predicted time point The azimuth of the sun below.

[0072] (5) Input the solar shading field and the current multimodal sensing data into the prediction model, and the prediction model outputs the power output of the photovoltaic site at each prediction time point.

[0073] In this embodiment of the invention, the prediction model includes: a shared image patch projector, a numerical shared projector, a feature fusion unit connected to the shared image patch projector and the numerical shared projector, an attention weight module connected to the feature fusion unit, a memory update module connected to the attention weight module, and a prediction module connected to the memory update module.

[0074] (51) Shared image patch projector, used to map input satellite images and fisheye images into M-dimensional image features;

[0075] The acquired historical satellite images and fisheye images are input into the shared image patch projector. The shared image patch projector extracts the image features of the satellite images and fisheye images respectively. The shared image patch projector consists of a two-dimensional convolutional layer with a kernel size of 16×16 and a stride of 16, a normalization layer, and an activation layer.

[0076] Acquire the sunflower satellite image data of the photovoltaic site before the current time. The sunflower satellite image is preferably a continuous B03 band image block in a preset area above the site, with a region size of 128×128. The sunflower satellite image is mainly used to reflect the evolution of the cloud field at the regional scale, including whether the cloud cluster is approaching the site from upstream, whether the cloud system boundary is close to the site, and whether the cloud cluster is expanding or dissipating.

[0077] Acquire fisheye image data collected by the all-sky imager deployed at the photovoltaic site. The original fisheye image data size can be 4000×3000. Preprocess the fisheye image data, including scaling, cropping, distortion correction or normalization. The all-sky fisheye image is mainly used to reflect the local sky conditions of the photovoltaic site, including solar neighborhood shading, cloud boundary advancement, local cloud cover and local cloud movement trends.

[0078] Acquire ground meteorological data for photovoltaic sites, including minute-level or near-minute-level meteorological elements such as temperature, humidity, wind speed, wind direction, and air pressure;

[0079] Obtain historical power data of photovoltaic sites, including historical power data preferably at 15-minute intervals, but higher frequency power data can also be used;

[0080] Obtain numerical weather forecast wind field data for the target area, including information such as wind speed and direction at different altitudes or pressure levels.

[0081] (52) Numerical sharing projector, used to map various types of input sensory data into N-dimensional data features;

[0082] The collected historical radar data, surface meteorological data, photovoltaic power data, and NWP wind field data are input into the numerical sharing projector. The numerical sharing projector extracts the data features of the radar data, surface meteorological data, historical power data, and NWP wind field data respectively. The numerical sharing projector consists of a linear layer, a normalization layer, and an activation layer.

[0083] (53) Feature fusion unit, used to aggregate the data features extracted by the numerical sharing projector and the image features extracted by the shared image patch projector into the solar occlusion field. The aggregation process is as follows:

[0084] (531) Calculate the contribution weights of various data features and image features to each grid in the solar shading field. Belonging to the above data features and image features, feature c corresponds to the grid. Contribution weight The specific calculation formula is as follows:

[0085] ;

[0086] in, Representation of features At the predicted time point The estimated location, Representation of features At the current time t, , Representation of features At the predicted time point The estimated altitude, the altitude at the current time t, , These represent the scaling factors for position and height, respectively.

[0087] (532) Each feature Grids converging into the solar shading field The specific aggregation formula is as follows:

[0088] ;

[0089] Each grid within all sun-shaded areas Fusion characteristics Fusion characteristics of the solar shading field .

[0090] (54) Attention weight module, used to extract fused features through an attention mechanism. Extract and fuse spatial, cloud height, and prediction time-of-flight occlusion features;

[0091] The attention weight module consists of three levels of attention modules connected in series, which are used for feature fusion. The spatial attention module, height attention module, and timeliness attention module are used to extract features distributed in space, height, and timeliness. The input and output residuals of each attention module are concatenated and then input into the normalization layer LayerNorm.

[0092] Spatial Attention Module space It is expressed as follows:

[0093] ;

[0094] ;

[0095] Attention module height It is expressed as follows:

[0096] ;

[0097] ;

[0098] Time-sensitive attention module τ It is expressed as follows:

[0099] ;

[0100] .

[0101] in, , , The output feature maps are for the spatial attention module, the height attention module, and the time-sensitive attention module, respectively.

[0102] (55) Memory update module, used to update the occlusion memory features of the previous time step. Compared with current occlusion features Perform fusion and update the occlusion features at the current moment. At the same time, it serves as the occlusion memory at the current moment, used to reduce the impact of single-period observation noise and enhance the model's ability to track continuous cloud occlusion processes.

[0103] Because the cloud occlusion process is continuous, the cloud may gradually approach the sun's path, enter the occlusion area, obscure the sun, and eventually leave the occlusion area over multiple scanning cycles. To characterize this process, the occlusion characteristics at the current moment are used as a basis. and the memory features of the previous moment Update the occlusion features at the current moment And the obscured memory of the present moment The update process is as follows:

[0104] ;

[0105] ;

[0106] ;

[0107] ;

[0108] in, To update the door; To reset the door; To obscure memories in the present, Indicates candidate memories, This indicates element-wise multiplication. , , Indicates the weight value. This represents the Sigmoid activation function.

[0109] (56) The prediction module is based on occlusion features Predicting photovoltaic sites at the predicted time point Photovoltaic power .

[0110] The prediction module includes a first fully connected layer and a second fully connected layer. The first fully connected layer is used for predicting photovoltaic power, and the second connected layer is used for predicting the trend of photovoltaic power change, including three types of trends: rising, stable, and falling.

[0111] This invention constructs a solar shading field by taking the future position of the sun, cloud height, and extrapolated position of cloud clusters. It maps multi-source observation data to a solar shading field that addresses the risk of photovoltaic shading. Based on the solar shading field, it guides the fusion of shading characteristics in time, space, and height to predict future power forecasts and power change states.

[0112] The present invention has been described by way of example. Obviously, the specific implementation of the present invention is not limited to the above-described manner. Any non-substantial improvements made using the inventive concept and technical solution of the present invention, or the direct application of the inventive concept and technical solution of the present invention to other occasions without modification, are all within the protection scope of the present invention.

Claims

1. A method for short-term prediction of photovoltaic power, characterized in that, The method is as follows: (1) Calculate the area of ​​cloud cover at different altitudes when solar radiation is blocked at the photovoltaic site at the predicted time point; (2) Detect whether the cloud layer at the current height at the predicted time point has moved to the occlusion area of ​​the corresponding height layer. If the detection result is yes, mark the cloud layer at the corresponding height as a candidate cloud cluster and generate a scanning area for scanning the candidate cloud cluster. (3) The cloud height of candidate cloud clusters is tracked by radar in the scanning area, and the solar shading field at the predicted time point is constructed based on the tracked cloud height. (4) Input the solar shading field and multimodal sensing data into the prediction model, and the prediction model outputs the power output of the photovoltaic site at each prediction time point.

2. The photovoltaic power short-term prediction method as described in claim 1, characterized in that, The method for detecting whether cloud layers at each altitude have moved into the occlusion area of ​​the corresponding altitude layer is as follows: The current movement direction and speed of cloud layers at each altitude are obtained to form the current motion information of cloud layers at each altitude. Based on the current motion information of cloud layers at each altitude, the estimated position of cloud layers at each altitude at the predicted time point is predicted. If the estimated position of cloud layers at each altitude is located within the corresponding occlusion area, it is determined that the current cloud layers at each altitude have moved to the occlusion area of ​​the corresponding altitude layer.

3. The photovoltaic power short-term prediction method as described in claim 1, characterized in that, The cloud height within the scanning area is tracked using radars A, B, and C. Radar B is used to detect the cloud height directly above the current location and sends the detection data to radars A and C. Radar A is used to detect the cloud height in the near field ahead of the scanning area, and radar C is used to detect the cloud height in the far field ahead of the scanning area.

4. The photovoltaic power short-term prediction method as described in claim 1, characterized in that, The process of constructing a solar shading field is as follows: Calculate the prediction time point The height at which solar radiation is blocked at the lower photovoltaic site is The center of the area obscured by the cloud In the center of the obstruction Build a mesh in the nearby occluded areas and put the built mesh into the mesh set. In the grid set It is expressed as follows: ； ； in, Indicates the predicted time point The height at which solar radiation is blocked at the lower photovoltaic site is The j-th grid in the area obscured by the cloud, This represents the coordinates of the center of the j-th grid. This represents the distance from the center of the j-th grid to the center of the occlusion. The radius of the shading area; Indicates the predicted time point The solar altitude angle below; Indicates the predicted time point The azimuth of the sun below.

5. The photovoltaic power short-term prediction method as described in claim 1, characterized in that, The prediction model includes: a shared image patch projector, a numerical shared projector, a feature fusion unit connected to the shared image patch projector and the numerical shared projector, an attention weight module connected to the feature fusion unit, a memory update module connected to the attention weight module, and a prediction module connected to the memory update module. A shared image patch projector is used to map input satellite images and fisheye images into M-dimensional image features; Numerical sharing projectors are used to map various types of input sensory data into N-dimensional data features; The feature fusion unit is used to aggregate the data features extracted by the numerical sharing projector and the image features extracted by the shared image patch projector into the solar shading field; The attention weight module is used to extract fused features through an attention mechanism. Extract and fuse spatial, cloud height, and prediction time-of-flight occlusion features; The memory update module is used to fuse the occlusion memory features from the previous moment with the current occlusion features and update the occlusion features at the current moment. The prediction module is used for prediction based on occlusion features. Predicting photovoltaic sites at the predicted time point Photovoltaic power .

6. The photovoltaic power short-term prediction method as described in claim 5, characterized in that, The polymerization process is as follows: Calculate the contribution weights of various data features and image features to each grid in the solar shading field. Belonging to the aforementioned data features and image features, feature c corresponds to the grid. Contribution weight The specific calculation formula is as follows: ； in, Representation of features At the predicted time point The estimated location, Representation of features At the current time t, , Representation of features At the predicted time point The estimated altitude, the altitude at the current time t, , These represent scaling factors for position and height, respectively. Each feature Grids converging into the solar shading field The specific aggregation formula is as follows: ； Each grid within all sun-shaded areas Fusion characteristics Fusion characteristics of the solar shading field .

7. The photovoltaic power short-term prediction method as described in claim 5, characterized in that, The attention weight module consists of three levels of attention modules connected in series, each used for feature fusion. The spatial attention module, height attention module, and timeliness attention module are extracted based on the feature distribution in terms of space, height, and timeliness. The input and output residuals of each attention module are concatenated and then input into the normalization layer LayerNorm.

8. The photovoltaic power short-term prediction method as described in claim 5, characterized in that, The shared image patch projector consists of a two-dimensional convolutional layer with a kernel size of 16×16 and a stride of 16, a normalization layer, and an activation layer.

9. The photovoltaic power short-term prediction method as described in claim 5, characterized in that, The numerical sharing projector consists of a linear layer, a normalization layer, and an activation layer.

10. The photovoltaic power short-term prediction method as described in claim 5, characterized in that, The prediction module includes a first fully connected layer and a second fully connected layer. The first fully connected layer is used for predicting photovoltaic power, and the second connected layer is used for predicting the trend of photovoltaic power change.

Images