An ultra-short-term photovoltaic power output prediction method, device, equipment and medium

Abstract

The present application relates to the field of photovoltaic technology, and proposes an ultra-short-term photovoltaic output prediction method, device, equipment and medium. The method comprises: acquiring multiple real-time ground-based cloud images for a to-be-predicted photovoltaic station in a current period; predicting the ground-based cloud images in a future period by a pre-trained cloud image prediction model according to each real-time ground-based cloud image to generate multiple predicted ground-based cloud images; performing cloud image feature extraction processing according to each predicted ground-based cloud image to obtain multiple modal features corresponding to each predicted ground-based cloud image; and performing photovoltaic output prediction by a pre-trained photovoltaic power prediction model according to each predicted ground-based cloud image and the multiple modal features corresponding to each predicted ground-based cloud image to obtain a photovoltaic output prediction value in the future period. The present application can significantly enhance the prediction accuracy and robustness under sudden weather changes and improve the accuracy of photovoltaic output prediction.

Description

Technical Field

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

[0002] In current short-term photovoltaic (PV) output forecasting (such as hourly or minute-level forecasts), there is often a problem that the forecast results lag behind actual changes. This delay mainly stems from the high dependence of traditional forecasting methods on historical data, which lacks the ability to respond to rapid fluctuations in PV output, especially under extreme conditions such as sudden weather changes.

[0003] In recent years, short-term photovoltaic (PV) output prediction methods based on ground-based cloud maps have been widely studied and show promise for mitigating the uncertainties of PV power generation. However, cloud movement is highly random, and current methods struggle to accurately model and predict the future evolution of cloud maps. Furthermore, current methods suffer from biases in cloud map feature extraction, leading to insufficient accuracy in establishing the correlation between PV output and cloud maps, thus affecting the final prediction results. In summary, current prediction methods have low accuracy and struggle to achieve precise predictions of PV output. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method, apparatus, equipment and medium for predicting ultra-short-term photovoltaic power output, so as to solve the above-mentioned technical problem.

[0005] The technical solution of this invention to solve the above-mentioned technical problems is as follows: A method for predicting ultra-short-term photovoltaic power output, comprising: acquiring multiple real-time ground-based cloud maps of the photovoltaic station to be predicted in the current period; predicting the ground-based cloud maps in the future period based on each real-time ground-based cloud map using a pre-trained cloud map prediction model, generating multiple predicted ground-based cloud maps; performing cloud map feature extraction processing on each predicted ground-based cloud map to obtain the multimodal features corresponding to each predicted ground-based cloud map; and predicting the photovoltaic power output in the future period using a pre-trained photovoltaic power prediction model based on each predicted ground-based cloud map and the multimodal features corresponding to each predicted ground-based cloud map, thereby obtaining the predicted photovoltaic power output value in the future period.

[0006] The beneficial effects of this invention are as follows: This method first predicts ground-based cloud images for future periods using a pre-trained cloud image prediction model. By combining local motion features and global evolution trends, it effectively captures the dynamic changes of clouds and generates realistic predicted ground-based cloud images, achieving high-precision cloud image prediction. Then, cloud image feature extraction processing is performed to obtain multimodal features, which are used to predict photovoltaic power output. Through multimodal features, the prediction accuracy and robustness under sudden weather changes are significantly enhanced, improving the accuracy of photovoltaic power output prediction.

[0007] Based on the above technical solution, the present invention can be further improved as follows.

[0008] Furthermore, the pre-trained cloud image prediction model is obtained through the following methods: acquiring a historical foundation cloud image set for the photovoltaic station to be predicted; preprocessing the historical foundation cloud image set to obtain a processed historical foundation cloud image set; constructing samples based on the processed historical foundation cloud image set to generate training samples; and training a preset generative neural network based on the training samples using a simulated annealing strategy and a teacher-forced strategy to obtain the pre-trained cloud image prediction model.

[0009] Furthermore, the preset generative neural network includes interconnected stacked ConvLSTM encoder layers and single-layer ConvLSTM decoders. The stacked ConvLSTM encoder layers are multi-layer network structures formed by stacking multiple ConvLSTM encoders based on a multi-layer state propagation mechanism.

[0010] Furthermore, for each predicted ground-based cloud image, cloud image feature extraction processing is performed based on the predicted ground-based cloud image to obtain the multimodal features corresponding to the predicted ground-based cloud image. This includes: performing local contrast enhancement processing on the predicted ground-based cloud image through contrast-limited adaptive histogram equalization to generate an enhanced cloud image corresponding to the predicted ground-based cloud image; performing grayscale processing on the predicted ground-based cloud image to generate a grayscale cloud image corresponding to the predicted ground-based cloud image; performing edge detection processing on the grayscale cloud image corresponding to the predicted ground-based cloud image using the Sobel operator to generate an edge intensity map corresponding to the predicted ground-based cloud image; calculating cloud cover based on the edge intensity map corresponding to the predicted ground-based cloud image; calculating the actual radiance value corresponding to the predicted ground-based cloud image under the cloud cover based on the cloud cover; and obtaining the multimodal features corresponding to the predicted ground-based cloud image based on the enhanced cloud image, the edge intensity map, and the actual radiance value.

[0011] Further, the step of calculating cloud cover based on the edge intensity map corresponding to the predicted ground cloud map includes: determining the pixel category corresponding to each pixel in the edge intensity map corresponding to the predicted ground cloud map using the Otsu algorithm, wherein the pixel category is either a cloud area pixel or a non-cloud area pixel; and calculating the proportion of pixels with the corresponding pixel category as cloud area pixels in a preset effective field of view based on the pixel category corresponding to each pixel to obtain the cloud cover.

[0012] Further, the step of calculating the actual radiation value corresponding to the predicted ground-based cloud map under the cloud cover includes: calculating the theoretical total radiation and solar apex angle corresponding to the photovoltaic station to be predicted under cloudless conditions using the Ineichen clear sky model; calculating the optical thickness coefficient based on the cloud cover; calculating the atmospheric transmittance based on the optical thickness coefficient and the solar apex angle; correcting the atmospheric transmittance based on a preset adjustment factor to obtain the corrected atmospheric transmittance; calculating the scattered radiation value based on the preset adjustment factor and the theoretical total radiation; and obtaining the actual radiation value corresponding to the predicted ground-based cloud map through weighted calculation based on the scattered radiation value, the theoretical total radiation, and the corrected atmospheric transmittance.

[0013] Furthermore, the step of predicting photovoltaic output based on each predicted ground-based cloud map and its corresponding multimodal features, using a pre-trained photovoltaic power prediction model, to obtain the predicted photovoltaic output value for the future period includes: performing feature extraction processing on each predicted ground-based cloud map, its corresponding enhanced cloud map, and its corresponding edge intensity map to obtain a first feature vector for each predicted ground-based cloud map; performing feature extraction processing on the actual radiation value for each predicted ground-based cloud map to obtain a second feature vector for each predicted ground-based cloud map; fusing the first feature vector and the second feature vector for each predicted ground-based cloud map to obtain a fused feature for each predicted ground-based cloud map; and generating the predicted photovoltaic output value for the future period based on the fused features for each predicted ground-based cloud map.

[0014] To address the aforementioned technical problems, the present invention also provides an ultra-short-term photovoltaic power output prediction device, comprising:

[0015] The image acquisition module is used to acquire multiple real-time ground cloud images of the photovoltaic station to be predicted during the current time period;

[0016] The image prediction module is used to predict the ground cloud map for future periods based on each real-time ground cloud map and a pre-trained cloud map prediction model, generating multiple predicted ground cloud maps.

[0017] The feature extraction module is used to perform cloud map feature extraction processing on each predicted foundation cloud map to obtain the multimodal features corresponding to each predicted foundation cloud map;

[0018] The photovoltaic power output prediction module is used to predict photovoltaic power output based on each predicted ground cloud map and the corresponding multimodal features, and to obtain the predicted photovoltaic power output value for the future period through a pre-trained photovoltaic power prediction model.

[0019] To address the aforementioned technical problems, the present invention also provides an electronic 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 ultra-short-term photovoltaic power output prediction method as described above.

[0020] To address the aforementioned technical problems, the present invention also provides a non-transitory computer-readable storage medium storing computer instructions for causing a computer to execute the ultra-short-term photovoltaic power output prediction method as described above. Attached Figure Description

[0021] Figure 1 This is a flowchart of an ultra-short-term photovoltaic power output prediction method according to the present invention;

[0022] Figure 2 This is a schematic diagram of the ground-based cloud map prediction results of the ultra-short-term photovoltaic power output prediction method of the present invention;

[0023] Figure 3 This is a schematic diagram of the photovoltaic output prediction results of the ultra-short-term photovoltaic output prediction method of the present invention;

[0024] Figure 4 This is a schematic diagram of an ultra-short-term photovoltaic power output prediction device according to the present invention;

[0025] Figure 5 This is a schematic diagram of an electronic device according to the present invention. Detailed Implementation

[0026] The principles and features of the present invention are described below. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0027] Example 1

[0028] like Figure 1 As shown, this embodiment provides a method for predicting ultra-short-term photovoltaic power output, including:

[0029] S101. Acquire multiple real-time ground-based cloud images for the photovoltaic station to be predicted during the current time period. The current time period represents the time period formed by the first time interval preceding the current moment. During the current time period, one real-time ground-based cloud image is acquired every first time interval, i.e., each real-time ground-based cloud image corresponds to one time moment. At least two real-time ground-based cloud images are acquired during the current time period. Specifically, the observation data of the all-sky imager located at the observation point is acquired to generate real-time ground-based cloud images.

[0030] S102. Based on each real-time ground cloud map, use a pre-trained cloud map prediction model to predict the ground cloud map for future periods and generate multiple predicted ground cloud maps.

[0031] The future time period represents the time segment formed by the first duration following the current moment. The number of generated predicted ground-based cloud images is equal to the number of input real-time ground-based cloud images. For example, after inputting 5 frames of real-time ground-based cloud images into a pre-trained cloud image prediction model, predicted ground-based cloud images for the next 5 frames are obtained. Figure 2 As shown, Input1, Input2, Input3, Input4, and Input5 are the 5 real-time ground cloud maps of the input, GT1, GT2, GT3, GT4, and GT5 are the corresponding 5 real ground cloud maps, and Pred1, Pred2, Pred3, Pred4, and Pred5 are the corresponding 5 predicted ground cloud maps of the output.

[0032] Furthermore, multiple real-time ground-based cloud images and multiple predicted ground-based cloud images... Figure 1 One-to-one correspondence, for example, Input1 is the real-time ground cloud map collected five minutes ago, Pred1 is the predicted ground cloud map five minutes later; Input2 is the real-time ground cloud map collected ten minutes ago, Pred2 is the predicted ground cloud map ten minutes later, and so on.

[0033] S103. Based on each predicted foundation cloud map, perform cloud map feature extraction processing to obtain the multimodal features corresponding to each predicted foundation cloud map.

[0034] S104. Based on each predicted ground cloud map and the corresponding multimodal features, photovoltaic power output is predicted using a pre-trained photovoltaic power prediction model to obtain the predicted photovoltaic power output value for the future period.

[0035] This method first predicts ground-based cloud images for future periods using a pre-trained cloud image prediction model. By combining local motion features and global evolution trends, it effectively captures dynamic changes in clouds and generates realistic predicted ground-based cloud images, achieving high-precision cloud image prediction. Next, cloud image feature extraction processing is performed to obtain multimodal features, which are then used to predict photovoltaic power output. By utilizing multimodal features, the prediction accuracy and robustness under abrupt weather changes are significantly enhanced, improving the precision of photovoltaic power output prediction.

[0036] Optionally, in an embodiment, the pre-trained cloud image prediction model is obtained by: acquiring a historical ground cloud image set for the photovoltaic station to be predicted; preprocessing the historical ground cloud image set to obtain a processed historical ground cloud image set; constructing samples based on the processed historical ground cloud image set to generate training samples; and training a preset generative neural network based on the training samples using a simulated annealing strategy and a teacher-forced strategy to obtain the pre-trained cloud image prediction model.

[0037] The historical foundation cloud image set contains multiple historical foundation cloud images for the photovoltaic (PV) station to be predicted. These historical foundation cloud images represent foundation cloud images collected for the PV station before the current time period. The historical foundation cloud image set undergoes preprocessing, specifically as follows:

[0038] First, to balance image clarity and memory usage, each historical ground cloud image in the historical ground cloud atlas was scaled to 64×64 pixels. Where H = W = 64.

[0039] Then normalize each historical foundation cloud map:

[0040]

[0041] By scaling and normalizing each historical foundation cloud map in the historical foundation cloud map atlas, the historical foundation cloud map atlas is preprocessed.

[0042] Based on the processed historical ground-based cloud atlas, multiple training samples are constructed. The specific process is as follows:

[0043] Sliding window construction: Set the input and output step sizes, let the input step size be M, the output step size be N, and the time granularity be 15 min, then the i-th training sample is:

[0044]

[0045] Optionally, in an embodiment, the preset generative neural network includes interconnected stacked ConvLSTM encoder layers and single-layer ConvLSTM decoders. The stacked ConvLSTM encoder layers are multi-layer network structures formed by stacking multiple ConvLSTM encoders based on a multi-layer state propagation mechanism.

[0046] During processing, the input sequence is fed into stacked ConvLSTM encoder layers, and then the final predicted image is output through a single-layer ConvLSTM decoder. By employing a generative neural network that combines local motion features and global evolution trends, it can effectively capture the dynamic changes of clouds and generate realistic predicted ground-based cloud maps.

[0047] Specifically, the model input is an image sequence in the form of [batch_size, 1, channels, height, width], and the output is an image in the form of [batch_size, channels, height, width]. In this embodiment, in order to balance training speed and computational cost, the shapes of the input and output are set to [128, 1, 3, 64, 64] and [128, 3, 64, 64], respectively.

[0048] Each ConvLSTM encoder consists of a single-step convolutional LSTM unit, enabling temporal modeling while preserving the spatial structure of the image. Four gates (input gate i) are generated by concatenating the current input with the hidden state from the previous time step and then performing a convolution operation. t Forgotten Gate t Output gate o t Candidate status This updates the hidden state and cell state, and the calculation formula is as follows:

[0049]

[0050] H t =o t ☉tanh(C t ).

[0051] The multi-layer state propagation mechanism employs a hierarchical heterogeneous memory mechanism, stacking multiple ConvLSTM encoders to form a multi-layer network structure with hidden state dimensions of 128, 128, and 64, respectively. Deeper temporal feature extraction and encoding are achieved through layer-by-layer propagation; the lower layers capture local short-term motion patterns, while the higher layers model global long-term evolutionary patterns. Each layer has independent hidden states and cell states. The first layer receives the actual input, and each subsequent layer receives the output of the previous layer as its input. If L layers of ConvLSTM are stacked, the inter-layer hidden state propagation formula is:

[0052]

[0053] In this embodiment, when constructing a multi-layer network structure, in order to balance model performance and complexity, a 3-layer structure is constructed as [128, 128, 64].

[0054] A single-layer ConvLSTM decoder transforms the output feature map of the last layer of the ConvLSTM encoder into an image with a specified number of channels (e.g., an RGB image) through a 1×1 convolution, and uses a Sigmoid activation function to constrain the output value range to [0,1] to adapt to pixel-level prediction tasks.

[0055] The model training process consists of two stages: the encoding stage and the decoding stage.

[0056] During the encoding phase, the model receives the input image sequence X = {X0, X1, ..., X...}. t-1}, and input to the decoder frame by frame, each step X t Combined with the output of the previous frame to predict the next frame. and with real image X t+1 Compare the calculated mean squared error loss (MSE):

[0057]

[0058] The decoding stage uses the last frame of the input image X t-1 Starting with {X}, recursively generate the prediction sequence {X}. t ,..,X N-1}, where N is the length of the target sequence, and the decoding loss is:

[0059]

[0060] The total loss is:

[0061] Specifically, during the decoding phase of the training process, a teacher forcing strategy is randomly adopted, which means that the real frame X from the previous time step is used as the training frame. t As input for the next step, it mitigates errors in the early training phase to accelerate convergence. A simulated annealing strategy is used to control the teacher-mandated ratio, initially set at 1 and gradually decreasing to 0, as shown in the formula:

[0062] r = max(0, 1 - α × epoch).

[0063] In this embodiment, the ratio α of the simulated annealing strategy is set to 0.03, which can achieve rapid convergence in the early stage of training while ensuring the training effect.

[0064] Model parameters are updated using the Adam optimizer, with ReduceLROnPlateau used to adaptively adjust the learning rate to improve convergence efficiency. Model weights are saved every few epochs and validated on a test set.

[0065] During the evaluation phase, since future real images are unavailable, the model can only predict the next N frame image sequence using an autoregression method based on its own generated historical frames. That is, the model starts from the last frame image X in the encoding phase. t-1 The prediction sequence is generated recursively by starting with the initial input and then using the predicted frame at each subsequent time step as the input for the next time step.

[0066] N represents the time scale to be predicted. The larger the value of N, the longer the prediction time scale; the smaller the value of N, the shorter the prediction time scale. In this embodiment, to balance prediction accuracy and time scale, N is set to 5 when the time granularity is 15 minutes.

[0067] The following evaluation metrics are calculated: mean squared error per frame and mean squared error per pixel. The mean squared error per frame is used to evaluate the error between the predicted image and the real image in each frame, and the mean squared error per pixel is used to evaluate the average error of each pixel.

[0068] Specifically:

[0069] Average mean square error per frame:

[0070]

[0071] Where B is the sample batch size. Let X represent the predicted image of the i-th sample in frame t. t (i) This corresponds to the real image. Ideally, B should be set to 128 and N to 5.

[0072] Average mean square error per pixel:

[0073]

[0074] Where H, W, and C represent the height, width, and number of channels of the image, respectively. In this embodiment, H = 64, W = 64, and C = 3.

[0075] Feature processing includes an image branch and a property branch. The image branch provides spatial visual features related to cloud distribution, structure, and motion, while the property branch uses an optical physics model to reflect the actual impact of cloud cover changes on radiation. Ultimately, the features extracted from both branches are fused into the model's input. The image branch primarily processes the original predicted ground-based cloud image to extract spatial and dynamic information influencing solar radiation variations.

[0076] Optionally, in an embodiment, for each predicted ground-based cloud image, cloud image feature extraction processing is performed based on the predicted ground-based cloud image to obtain the multimodal features corresponding to the predicted ground-based cloud image, including: performing local contrast enhancement processing on the predicted ground-based cloud image through contrast-limited adaptive histogram equalization to generate an enhanced cloud image corresponding to the predicted ground-based cloud image; performing grayscale processing on the predicted ground-based cloud image to generate a grayscale cloud image corresponding to the predicted ground-based cloud image; performing edge detection processing on the grayscale cloud image corresponding to the predicted ground-based cloud image using the Sobel operator to generate an edge intensity map corresponding to the predicted ground-based cloud image; calculating cloud cover based on the edge intensity map corresponding to the predicted ground-based cloud image; calculating the actual radiance value corresponding to the predicted ground-based cloud image under the cloud cover based on the cloud cover; and obtaining the multimodal features corresponding to the predicted ground-based cloud image based on the enhanced cloud image, the edge intensity map, and the actual radiance value.

[0077] Image enhancement involves improving the contrast and local feature structure of the predicted ground-based cloud image. Specifically, the input predicted ground-based cloud image has a resolution of H×W and includes three color channels: R, G, and B, providing the most intuitive information on cloud morphology, color, and density. To enhance cloud boundaries and differences in cloud thickness, the RGB image is converted to the LAB space, separating the luminance (L) and chroma (A / B) components.

[0078] The L component is processed using the Limiting Contrast Adaptive Histogram Equalization (CLAHE) method to improve the local contrast of the cloud boundaries. Specifically, CLAHE divides the original image into multiple tiles, performs local histogram balancing on each tile, and limits the contrast to prevent over-enhancement. Ideally, each tile is set to 8×8 pixels. Finally, the enhanced L component is merged back with the original A / B image and converted back to RGB to obtain the enhanced cloud image.

[0079] Edge detection captures the geometric features of cloud morphology and areas of abrupt change at cloud edges. Specifically, the predicted ground-based cloud image in RGB format is first converted to a grayscale image, which better preserves the brightness difference between cloud areas and clear sky areas: I gray =α·R+β·G+γ·B.

[0080] To better highlight the difference between clouds and clear skies, the values ​​of α, β, and γ were set to 0.299, 0.587, and 0.114, respectively.

[0081] Then use the Sobel operator to calculate the gradient image G of the image. x and G y And synthesize them into an edge intensity map:

[0082]

[0083] Gradient maps reflect the changing trends of cloud ladder edges, with strong gradients typically representing differences in cloud thickness and edge regions.

[0084] Optionally, in an embodiment, calculating cloud cover based on the edge intensity map corresponding to the predicted ground-based cloud map includes: determining the pixel category corresponding to each pixel in the edge intensity map corresponding to the predicted ground-based cloud map using the Otsu algorithm, wherein the pixel category is either a cloud area pixel or a non-cloud area pixel; and calculating the proportion of pixels with the corresponding pixel category as cloud area pixels in a preset effective field of view based on the pixel category corresponding to each pixel to obtain the cloud cover.

[0085] Cloud cover calculation quantifies the proportion and spatial distribution of clouds in an image. Specifically, the Otsu algorithm is used to binarize the edge intensity map to distinguish between "cloudy" and "non-cloudy" regions. An optimal threshold T is automatically selected by maximizing the inter-class variance, minimizing the variance when the image is divided into two classes.

[0086] T = argmax τ [ω0(τ)ω1(τ)(μ0(τ)-(μ1(τ)) 2 ];

[0087] Where: ω0, ω1 are the proportions of the two types of pixels; μ0, μ1 are the average gray levels of the two types of pixels.

[0088] The grayscale cloud image is compared with a threshold to generate a binary cloud mask:

[0089]

[0090] Where M(x,y) = 1 indicates that the pixel is a "cloud area", and M(x,y) = 0 indicates that the pixel is a "non-cloud area".

[0091] Since ground-based cloud images are typically fisheye images, to avoid invalid pixels at the edges affecting the statistical results, a circular field-of-view mask is used to perform statistics only on the effective area. The effective field-of-view area is defined by setting the center (x0, y0) and radius r:

[0092]

[0093] The final cloud cover calculation is performed only within the valid area. Cloud cover is defined as the percentage of "cloud pixels" within the valid area, and the calculation formula is as follows:

[0094]

[0095] Optionally, in an embodiment, calculating the actual radiation value corresponding to the predicted ground-based cloud map under the cloud cover includes: calculating the theoretical total radiation and solar apex angle of the photovoltaic station to be predicted under cloudless conditions using the Ineichen clear-sky model; calculating the optical thickness coefficient based on the cloud cover; calculating the atmospheric transmittance based on the optical thickness coefficient and the solar apex angle; correcting the atmospheric transmittance based on a preset adjustment factor to obtain the corrected atmospheric transmittance; calculating the scattered radiation value based on the preset adjustment factor and the theoretical total radiation; and obtaining the actual radiation value corresponding to the predicted ground-based cloud map by weighted calculation based on the scattered radiation value, the theoretical total radiation, and the corrected atmospheric transmittance.

[0096] The physical properties branch mainly involves modeling solar radiation using meteorological theories to estimate the actual radiation values ​​under different cloud cover conditions from a physical perspective.

[0097] Specifically, the theoretical total radiation (GHI) and solar zenith angle θ under cloudless conditions were calculated using the Ineichen clear-sky model. z Secondly, considering the nonlinear blocking effect of clouds on radiation, a cubic nonlinear relationship is constructed between the optical thickness coefficient τ and cloud cover:

[0098] T = a.CF 3 +b.

[0099] The optimal values ​​are a = 25.65 and b = 0.3.

[0100] Furthermore, by combining the solar zenith angle, atmospheric transmittance is obtained:

[0101]

[0102] The optimal value is c = 0.595.

[0103] Furthermore, to improve the model's response to rapid cloud changes, an external cloud cover rate adjustment factor α was introduced, and the corrected direct sunlight transmittance was:

[0104] T′=T·α.

[0105] Furthermore, considering the contribution of diffuse reflection from clouds, we assume it to be a linear function of the theoretical total radiation:

[0106] R diffuse =0.75·α·GHI.

[0107] The optimal value is α = 0.5.

[0108] The actual radiation value is obtained by weighted superposition of the direct and scattered components:

[0109] R actual = (1-L)·GHI·T′+L·R diffuse .

[0110] The optimal value is L = 0.488.

[0111] For each predicted ground-based cloud map, the enhanced cloud map, edge intensity map, and actual radiance value are calculated in the manner described above, thereby obtaining the multimodal features corresponding to each predicted ground-based cloud map.

[0112] Optionally, in an embodiment, the step of predicting photovoltaic output based on each predicted ground-based cloud map and the corresponding multimodal features of each predicted ground-based cloud map, and obtaining the predicted photovoltaic output value for the future period through a pre-trained photovoltaic power prediction model, includes: performing feature extraction processing on each predicted ground-based cloud map, the enhanced cloud map corresponding to each predicted ground-based cloud map, and the edge intensity map corresponding to each predicted ground-based cloud map to obtain a first feature vector corresponding to each predicted ground-based cloud map; performing feature extraction processing on the actual radiation value corresponding to each predicted ground-based cloud map to obtain a second feature vector corresponding to each predicted ground-based cloud map; fusing the first feature vector and the second feature vector corresponding to each predicted ground-based cloud map to obtain a fused feature corresponding to each predicted ground-based cloud map; and generating the predicted photovoltaic output value for the future period based on the fused features corresponding to each predicted ground-based cloud map.

[0113] The pre-trained photovoltaic power prediction model is a multimodal neural network structure that adopts a multi-input fusion architecture. It fuses and models image features and radiation data separately, and is ultimately used to predict actual photovoltaic output.

[0114] Specifically, the model consists of two parts: an image branch and a property branch. The image branch takes a 64×64×8 multi-channel cloud image as input (including the original predicted foundation cloud image, enhanced cloud image, and edge intensity image), extracts spatial features through a three-layer convolutional module, and compresses it into a 128-dimensional image feature vector through global average pooling, thus obtaining the first feature vector. The property branch takes the calculated actual radiation value as input, maps it to 64 dimensions through a fully connected layer, thus obtaining the second feature vector. The outputs of the two branches are concatenated and fused, and then further extracted through a fully connected layer to finally output a scalar, which is used for regression prediction of actual photovoltaic power output. The entire model uses mean squared error as the loss function and is trained using the Adam optimizer.

[0115] When predicting photovoltaic (PV) output using a pre-trained PV power prediction model, multiple predicted ground-based cloud maps and their corresponding multimodal features can be simultaneously input into the model to output multiple predicted PV output values. Alternatively, a single predicted ground-based cloud map and its corresponding multimodal features can be input into the model to output the predicted PV output value corresponding to that map.

[0116] Taking the example of the pre-trained cloud image prediction model outputting 5 frames of predicted ground-based images at a time, each of the 5 frames corresponds to a specific moment. When predicting photovoltaic (PV) output using the pre-trained PV power prediction model, a predicted PV output value can be derived based on each predicted ground-based image and its corresponding multimodal features. That is, each predicted ground-based image corresponds to a predicted PV output value, which represents the predicted PV output of the PV station at the moment corresponding to that predicted ground-based image. PV output refers to the electrical power output of a photovoltaic (PV) module or PV system within a specific time period.

[0117] like Figure 3 The diagram shown illustrates the photovoltaic output prediction results. This method integrates image features (enhanced cloud map, edge intensity map) and physical property parameters (actual radiation value), which retains the flexibility of data-driven approaches while introducing physical constraints, significantly enhancing the prediction accuracy and robustness under sudden weather changes.

[0118] Example 2

[0119] like Figure 4 As shown, this embodiment provides an ultra-short-term photovoltaic power output prediction device 200, including:

[0120] Image acquisition module 201 is used to acquire multiple real-time ground cloud images of the photovoltaic station to be predicted in the current time period;

[0121] The image prediction module 202 is used to predict the ground cloud map for future periods based on each real-time ground cloud map and a pre-trained cloud map prediction model, and generate multiple predicted ground cloud maps.

[0122] The feature extraction module 203 is used to perform cloud map feature extraction processing on each predicted foundation cloud map to obtain the multimodal features corresponding to each predicted foundation cloud map.

[0123] The photovoltaic power output prediction module 204 is used to predict photovoltaic power output based on each predicted ground cloud map and the corresponding multimodal features of each predicted ground cloud map, and to obtain the predicted photovoltaic power output value for the future period through a pre-trained photovoltaic power prediction model.

[0124] Optionally, in an embodiment, the pre-trained cloud image prediction model is obtained by: acquiring a historical ground cloud image set for the photovoltaic station to be predicted; preprocessing the historical ground cloud image set to obtain a processed historical ground cloud image set; constructing samples based on the processed historical ground cloud image set to generate training samples; and training a preset generative neural network based on the training samples using a simulated annealing strategy and a teacher-forced strategy to obtain the pre-trained cloud image prediction model.

[0125] Optionally, in an embodiment, the preset generative neural network includes interconnected stacked ConvLSTM encoder layers and single-layer ConvLSTM decoders. The stacked ConvLSTM encoder layers are multi-layer network structures formed by stacking multiple ConvLSTM encoders based on a multi-layer state propagation mechanism.

[0126] Optionally, in an embodiment, for each predicted ground-based cloud image, cloud image feature extraction processing is performed based on the predicted ground-based cloud image to obtain the multimodal features corresponding to the predicted ground-based cloud image, including: performing local contrast enhancement processing on the predicted ground-based cloud image through contrast-limited adaptive histogram equalization to generate an enhanced cloud image corresponding to the predicted ground-based cloud image; performing grayscale processing on the predicted ground-based cloud image to generate a grayscale cloud image corresponding to the predicted ground-based cloud image; performing edge detection processing on the grayscale cloud image corresponding to the predicted ground-based cloud image using the Sobel operator to generate an edge intensity map corresponding to the predicted ground-based cloud image; calculating cloud cover based on the edge intensity map corresponding to the predicted ground-based cloud image; calculating the actual radiance value corresponding to the predicted ground-based cloud image under the cloud cover based on the cloud cover; and obtaining the multimodal features corresponding to the predicted ground-based cloud image based on the enhanced cloud image, the edge intensity map, and the actual radiance value.

[0127] Optionally, in an embodiment, calculating cloud cover based on the edge intensity map corresponding to the predicted ground-based cloud map includes: determining the pixel category corresponding to each pixel in the edge intensity map corresponding to the predicted ground-based cloud map using the Otsu algorithm, wherein the pixel category is either a cloud area pixel or a non-cloud area pixel; and calculating the proportion of pixels with the corresponding pixel category as cloud area pixels in a preset effective field of view based on the pixel category corresponding to each pixel to obtain the cloud cover.

[0128] Optionally, in an embodiment, calculating the actual radiation value corresponding to the predicted ground-based cloud map under the cloud cover includes: calculating the theoretical total radiation and solar apex angle of the photovoltaic station to be predicted under cloudless conditions using the Ineichen clear-sky model; calculating the optical thickness coefficient based on the cloud cover; calculating the atmospheric transmittance based on the optical thickness coefficient and the solar apex angle; correcting the atmospheric transmittance based on a preset adjustment factor to obtain the corrected atmospheric transmittance; calculating the scattered radiation value based on the preset adjustment factor and the theoretical total radiation; and obtaining the actual radiation value corresponding to the predicted ground-based cloud map by weighted calculation based on the scattered radiation value, the theoretical total radiation, and the corrected atmospheric transmittance.

[0129] Optionally, in this embodiment, the photovoltaic output prediction module 204 includes:

[0130] The first feature extraction unit is used to perform feature extraction processing based on each of the predicted ground cloud maps, the enhanced cloud map corresponding to each of the predicted ground cloud maps, and the edge intensity map corresponding to each of the predicted ground cloud maps, to obtain a first feature vector corresponding to each of the predicted ground cloud maps;

[0131] The second feature extraction unit is used to perform feature extraction processing based on the actual radiation value corresponding to each of the predicted ground cloud maps to obtain the second feature vector corresponding to each of the predicted ground cloud maps.

[0132] The feature fusion unit is used to fuse the first feature vector and the second feature vector corresponding to each of the predicted ground cloud maps to obtain the fused feature corresponding to each of the predicted ground cloud maps.

[0133] The power output prediction unit is used to generate photovoltaic power output prediction values ​​for future periods based on the fusion characteristics corresponding to each of the predicted ground cloud maps.

[0134] In some embodiments, the ultra-short-term photovoltaic power output prediction device 200 of the present invention can be implemented in a combination of hardware and software. As an example, the ultra-short-term photovoltaic power output prediction device 200 of the present invention can be a processor in the form of a hardware decoding processor, which is programmed to execute the ultra-short-term photovoltaic power output prediction method of the present invention. For example, the processor in the form of a hardware decoding processor can be one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), or other electronic components.

[0135] The modules described in the embodiments of this invention can be implemented in software or hardware. The names of the modules are not, in some cases, limiting the scope of the module itself.

[0136] Example 3

[0137] like Figure 5 As shown, this embodiment provides an electronic 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 an ultra-short-term photovoltaic power output prediction method as described in Embodiment 1.

[0138] In other words, an electronic device according to an embodiment of the present invention may include, but is not limited to, a processor and a memory; the memory is used to store a computer program; the processor is used to execute an ultra-short-term photovoltaic power output prediction method shown in any embodiment of the present invention by calling the computer program.

[0139] In one alternative embodiment, an electronic device is provided, such as Figure 5 As shown, Figure 5 The illustrated electronic device 3000 includes a processor 3001 and a memory 3003. The processor 3001 and the memory 3003 are connected, for example, via a bus 3002. Optionally, the electronic device 3000 may further include a transceiver 3004, which can be used for data interaction between the electronic device and other electronic devices, such as sending and / or receiving data. It should be noted that in practical applications, the transceiver 3004 is not limited to one type, and the structure of the electronic device 3000 does not constitute a limitation on the embodiments of the present invention.

[0140] Processor 3001 may be a CPU (Central Processing Unit), a general-purpose processor, a DSP (Digital Signal Processor), an ASIC (Application Specific Integrated Circuit), an FPGA (Field Programmable Gate Array), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logic blocks, modules, and circuits described in conjunction with the disclosure of this invention. Processor 3001 may also be a combination that implements computational functions, such as including one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc.

[0141] Bus 3002 may include a path for transmitting information between the aforementioned components. Bus 3002 may be a PCI (Peripheral Component Interconnect) bus or an EISA (Extended Industry Standard Architecture) bus, etc. Bus 3002 can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 5 The bus 3002 is represented by only one thick line, but this does not mean that there is only one bus or one type of bus.

[0142] The memory 3003 may be ROM (Read Only Memory) or other types of static storage devices capable of storing static information and instructions, RAM (Random Access Memory) or other types of dynamic storage devices capable of storing information and instructions, or EEPROM (Electrically Erasable Programmable Read Only Memory), CD-ROM (Compact Disc Read Only Memory) or other optical disc storage, optical disc storage (including compressed optical discs, laser discs, optical discs, digital universal optical discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or any other medium capable of carrying or storing desired program code in the form of instructions or data structures and accessible by a computer, but not limited thereto.

[0143] The memory 3003 stores the application code (computer program) for executing the present invention, and its execution is controlled by the processor 3001. The processor 3001 executes the application code stored in the memory 3003 to implement the content shown in the foregoing method embodiments.

[0144] Among them, electronic devices can also be terminal devices, which can be any device that can install applications, including at least one of smartphones, tablets, laptops, desktop computers, smart speakers, smartwatches, smart TVs, and smart in-vehicle devices.

[0145] It should be noted that, Figure 5 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of the present invention.

[0146] Example 4

[0147] This embodiment provides a non-transitory computer-readable storage medium that stores computer instructions for causing a computer to execute an ultra-short-term photovoltaic power output prediction method as described in Embodiment 1.

[0148] Alternatively, the computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a compact disc read-only memory (CD-ROM), magnetic tape, a floppy disk, and an optical data storage device, etc.

[0149] In an exemplary embodiment, a computer program product or computer program is also provided, which includes computer instructions stored in a computer-readable storage medium. A processor of an electronic device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the electronic device to perform the aforementioned ultra-short-term photovoltaic power output prediction method.

[0150] Computer program code for performing the operations of this invention can be written in one or more programming languages ​​or a combination thereof. These programming languages ​​include object-oriented programming languages—such as Java, Smalltalk, and C++—and conventional procedural programming languages—such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0151] It should be understood that the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0152] The computer-readable storage medium provided in this invention can be, but is not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EEPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0153] The aforementioned computer-readable storage medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to perform the method shown in the above embodiments.

[0154] The above description is merely a preferred embodiment of the present invention and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of disclosure in this invention is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this invention.

[0155] It should be noted that the terms "first," "second," etc., used in the specification and claims of this application are used to distinguish similar objects and represent a limitation on a specific order or sequence. Where appropriate, the order of use for similar objects can be interchanged so that the embodiments of this application described herein can be implemented in an order other than that shown or described.

[0156] Those skilled in the art will recognize that this invention can be implemented as a system, method, or computer program product. Therefore, this invention can be specifically implemented in the following forms: it can be entirely hardware, entirely software (including firmware, resident software, microcode, etc.), or a combination of hardware and software, generally referred to herein as a "circuit," "module," or "system." Furthermore, in some embodiments, this invention can also be implemented as a computer program product contained in one or more computer-readable media, which includes computer-readable program code.

[0157] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A method for predicting ultra-short-term photovoltaic power output, characterized in that, include: Obtain multiple real-time ground-based cloud images of the photovoltaic station to be predicted during the current time period; Based on each real-time ground cloud map, a pre-trained cloud map prediction model is used to predict the ground cloud map for future periods, generating multiple predicted ground cloud maps. Based on each predicted foundation cloud map, cloud map feature extraction processing is performed to obtain the multimodal features corresponding to each predicted foundation cloud map; Based on each predicted ground cloud map and the corresponding multimodal features, photovoltaic power output is predicted using a pre-trained photovoltaic power prediction model to obtain the predicted photovoltaic power output value for future periods. For each predicted foundation cloud map, feature extraction processing is performed based on the predicted foundation cloud map to obtain the multimodal features corresponding to the predicted foundation cloud map, including: The predicted foundation cloud map is locally enhanced by limiting contrast adaptive histogram equalization to generate an enhanced cloud map corresponding to the predicted foundation cloud map. The predicted foundation cloud map is processed into grayscale to generate a grayscale cloud map corresponding to the predicted foundation cloud map; Based on the grayscale cloud map corresponding to the predicted foundation cloud map, edge detection processing is performed using the Sobel operator to generate the edge intensity map corresponding to the predicted foundation cloud map; Cloud cover is calculated based on the edge intensity map corresponding to the predicted foundation cloud map; Based on the cloud cover, calculate the actual radiation value corresponding to the predicted ground-based cloud map under the cloud cover. Based on the enhanced cloud map corresponding to the predicted ground cloud map, the edge intensity map corresponding to the predicted ground cloud map, and the actual radiation value corresponding to the predicted ground cloud map, the multimodal features corresponding to the predicted ground cloud map are obtained; The process of predicting photovoltaic output based on each predicted ground-based cloud map and its corresponding multimodal features, using a pre-trained photovoltaic power prediction model, to obtain predicted photovoltaic output values ​​for future periods includes: Feature extraction is performed on each of the predicted foundation cloud maps, the enhanced cloud map corresponding to each of the predicted foundation cloud maps, and the edge intensity map corresponding to each of the predicted foundation cloud maps to obtain a first feature vector corresponding to each of the predicted foundation cloud maps; Each predicted ground-based cloud map is processed by feature extraction based on the actual radiation value corresponding to it, to obtain a second feature vector corresponding to each predicted ground-based cloud map. The first feature vector and the second feature vector corresponding to each of the predicted ground cloud maps are fused to obtain the fused feature corresponding to each of the predicted ground cloud maps. Based on the fusion characteristics corresponding to each of the predicted ground cloud maps, the predicted photovoltaic output value for future time periods is generated. The pre-trained photovoltaic power prediction model is a multimodal neural network structure. The pre-trained photovoltaic power prediction model includes two parts: an image branch and a physical property branch. The image branch takes the original predicted foundation cloud map, enhanced cloud map, and edge intensity map as input, extracts spatial features through a three-layer convolutional module, and compresses them into a 128-dimensional image feature vector through global average pooling, which is the first feature vector. The physical property branch takes the calculated actual radiation value as input, and maps it to 64 dimensions through a fully connected layer, which is the second feature vector.

2. The ultra-short-term photovoltaic power output prediction method according to claim 1, characterized in that, The pre-trained cloud map prediction model was obtained in the following way: Obtain the historical ground-based cloud map atlas for the photovoltaic station to be predicted; The historical foundation cloud atlas is preprocessed to obtain the processed historical foundation cloud atlas; Based on the processed historical ground-based cloud atlas, sample construction is performed to generate training samples; Based on the training samples, a pre-set generative neural network is trained using a simulated annealing strategy and a teacher-forced strategy to obtain the pre-trained cloud map prediction model.

3. The ultra-short-term photovoltaic power output prediction method according to claim 2, characterized in that, The preset generative neural network includes interconnected stacked ConvLSTM encoder layers and single-layer ConvLSTM decoders. The stacked ConvLSTM encoder layers are multi-layer network structures formed by stacking multiple ConvLSTM encoders based on a multi-layer state propagation mechanism.

4. The ultra-short-term photovoltaic power output prediction method according to claim 1, characterized in that, The calculation of cloud cover based on the edge intensity map corresponding to the predicted ground-based cloud map includes: The Otsu algorithm is used to determine the pixel category of each pixel in the edge intensity map corresponding to the predicted ground cloud map, wherein the pixel category is either a cloud area pixel or a non-cloud area pixel. Based on the pixel category corresponding to each pixel, the proportion of pixels whose corresponding pixel category is cloud area pixels in the preset effective field of view is calculated to obtain the cloud amount.

5. The ultra-short-term photovoltaic power output prediction method according to claim 1, characterized in that, The step of calculating the actual radiation value corresponding to the predicted ground-based cloud map under the cloud cover includes: The theoretical total radiation and solar zenith angle of the photovoltaic station to be predicted under cloudless conditions were calculated using the Ineichen clear sky model. Calculate the optical thickness coefficient based on the cloud cover. Calculate the atmospheric transmittance based on the optical thickness coefficient and the solar apex angle; Based on a preset adjustment factor, the atmospheric transmittance is corrected to obtain the corrected atmospheric transmittance. The scattered radiation value is calculated based on the preset adjustment factor and the theoretical total radiation. The actual radiation value corresponding to the predicted ground-based cloud map is obtained by weighted calculation based on the scattered radiation value, the theoretical total radiation, and the corrected atmospheric transmittance.

6. A short-term photovoltaic power output prediction device, characterized in that, include: The image acquisition module is used to acquire multiple real-time ground cloud images of the photovoltaic station to be predicted during the current time period; The image prediction module is used to predict the ground cloud map for future periods based on each real-time ground cloud map and a pre-trained cloud map prediction model, generating multiple predicted ground cloud maps. The feature extraction module is used to perform cloud map feature extraction processing on each predicted foundation cloud map to obtain the multimodal features corresponding to each predicted foundation cloud map; The photovoltaic power output prediction module is used to predict photovoltaic power output based on each predicted ground cloud map and the corresponding multimodal features of each predicted ground cloud map, and to obtain the predicted photovoltaic power output value for the future period through a pre-trained photovoltaic power prediction model. For each predicted ground-based cloud image, feature extraction processing is performed based on the predicted ground-based cloud image to obtain the multimodal features corresponding to the predicted ground-based cloud image. This includes: performing local contrast enhancement processing on the predicted ground-based cloud image through contrast-limited adaptive histogram equalization to generate an enhanced cloud image corresponding to the predicted ground-based cloud image; performing grayscale processing on the predicted ground-based cloud image to generate a grayscale cloud image corresponding to the predicted ground-based cloud image; performing edge detection processing on the grayscale cloud image corresponding to the predicted ground-based cloud image using the Sobel operator to generate an edge intensity map corresponding to the predicted ground-based cloud image; calculating cloud cover based on the edge intensity map corresponding to the predicted ground-based cloud image; calculating the actual radiance value corresponding to the predicted ground-based cloud image under the cloud cover based on the cloud cover; and obtaining the multimodal features corresponding to the predicted ground-based cloud image based on the enhanced cloud image, the edge intensity map, and the actual radiance value. The photovoltaic output prediction module includes: The first feature extraction unit is used to perform feature extraction processing based on each of the predicted ground cloud maps, the enhanced cloud map corresponding to each of the predicted ground cloud maps, and the edge intensity map corresponding to each of the predicted ground cloud maps, to obtain a first feature vector corresponding to each of the predicted ground cloud maps; The second feature extraction unit is used to perform feature extraction processing based on the actual radiation value corresponding to each of the predicted ground cloud maps to obtain the second feature vector corresponding to each of the predicted ground cloud maps. The feature fusion unit is used to fuse the first feature vector and the second feature vector corresponding to each of the predicted ground cloud maps to obtain the fused feature corresponding to each of the predicted ground cloud maps. The output prediction unit is used to generate photovoltaic output prediction values ​​for future periods based on the fusion characteristics corresponding to each of the predicted ground cloud maps. The pre-trained photovoltaic power prediction model is a multimodal neural network structure. The pre-trained photovoltaic power prediction model includes two parts: an image branch and a physical property branch. The image branch takes the original predicted foundation cloud map, enhanced cloud map, and edge intensity map as input, extracts spatial features through a three-layer convolutional module, and compresses them into a 128-dimensional image feature vector through global average pooling, which is the first feature vector. The physical property branch takes the calculated actual radiation value as input, and maps it to 64 dimensions through a fully connected layer, which is the second feature vector.

7. An electronic device, characterized in that, The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement an ultra-short-term photovoltaic power output prediction method as described in any one of claims 1 to 5.

8. A non-transitory computer-readable storage medium storing computer instructions, characterized in that, The computer instructions are used to cause the computer to execute the ultra-short-term photovoltaic power output prediction method according to any one of claims 1 to 5.

Images