Scattered photovoltaic power station short-term power prediction method, device, medium and product
By combining the cloud map feature extraction model and the timing feature extraction model, the prediction misalignment caused by the movement of the scattered photovoltaic power stations in the cloud layer is solved, and a higher-precision photovoltaic power prediction is achieved.
Patent Information
- Application Number
- CN202510581592.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-07
- Publication Date
- 2025-08-08
AI Technical Summary
In the face of the sudden change in photovoltaic power generation caused by cloud movement, especially in dispersed rural areas, it is difficult to accurately capture the dynamic characteristics of clouds and effectively fuse multimodal data, resulting in insufficient prediction accuracy.
The pre-trained cloud map feature extraction model and timing feature extraction model are used to extract cloud map features and timing features from satellite cloud maps and power data respectively. The characteristics are aligned and fused through attention mechanisms and depth typical correlation methods to predict the power of the photovoltaic power station at the next moment.
It improves the perception ability of cloud movement and global feature capture ability, enhances the accuracy of power prediction of photovoltaic power plants, and solves the prediction error caused by cloud movement.
Smart Images
Figure CN120451817A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of photovoltaic power stations, and more specifically, to a method, equipment, medium and product for short-term power prediction of a decentralized photovoltaic power station. Background Art
[0002] Photovoltaic power generation, as a clean and sustainable form of energy, has been widely adopted and promoted. Particularly in rural areas, distributed photovoltaic power generation systems, due to their flexibility and low barriers to entry, have gradually become a vital component of rural energy supply. However, photovoltaic power generation systems in rural areas are typically dispersed, with small individual system capacities and significant influences from geographical and meteorological conditions. This poses numerous challenges to photovoltaic power generation prediction.
[0003] Photovoltaic power generation prediction is a critical component of PV power plant operation and management. Accurate power prediction not only contributes to the stable operation of the power grid but also improves the economic benefits of PV power generation. Traditional power prediction methods rely primarily on historical power and meteorological data, using statistical or machine learning models. However, these methods often exhibit insufficient prediction accuracy when faced with complex meteorological changes, particularly sudden changes in PV power caused by cloud movement.
[0004] Cloud movement is a significant factor influencing photovoltaic (PV) power generation. Cloud thickness, speed, and direction directly impact the irradiance of PV panels, leading to dramatic fluctuations in PV power generation. Traditional forecasting methods, lacking the ability to capture the dynamic characteristics of cloud cover, struggle to effectively address these sudden fluctuations. Furthermore, PV power generation systems in rural areas are dispersed, with individual systems having relatively small capacities. Traditional centralized forecasting methods struggle to fully utilize the environmental information from these distributed systems, further limiting forecast accuracy.
[0005] With the rapid development of satellite remote sensing and deep learning technologies, photovoltaic power generation forecasting using satellite imagery has become a research hotspot. Satellite imagery can provide large-scale, high-temporal and spatial resolution meteorological information, particularly information on cloud dynamics, providing a new data source for photovoltaic power generation forecasting. Furthermore, deep learning technologies, particularly convolutional neural networks (CNNs) and long short-term memory networks (LSTMs), have demonstrated powerful capabilities in image processing and time series data analysis, offering new solutions for multimodal data fusion and feature extraction.
[0006] However, existing photovoltaic power prediction methods based on satellite imagery and deep learning still have some shortcomings. First, traditional convolutional neural networks often ignore the dynamic changes in cloud layers when processing satellite imagery, making it difficult to effectively capture the direction and speed of cloud movement. Second, existing multimodal learning methods often lack consideration of cross-modal feature alignment when fusing image data and time series data, resulting in insufficient feature expression after fusion. In addition, for areas where photovoltaic power generation systems are more dispersed, such as rural areas, traditional photovoltaic power prediction methods find it difficult to fully utilize the environmental information of these decentralized power stations, resulting in insufficient accuracy in photovoltaic power generation predictions in rural areas. Summary of the Invention
[0007] The purpose of the present invention is to provide a method, device, medium and product for short-term power prediction of decentralized photovoltaic power stations, which solves the problem of inaccurate prediction caused by sudden changes in photovoltaic power due to cloud movement in decentralized photovoltaic power stations in rural areas.
[0008] The above technical objectives of the present invention are achieved through the following technical solutions:
[0009] A first aspect of the present invention provides a method for short-term power prediction of a distributed photovoltaic power station, the method comprising:
[0010] Obtain satellite cloud images and power data of photovoltaic power stations in a region at the current moment;
[0011] Call the pre-trained cloud image feature extraction model to extract cloud image features containing the attention weight distribution between cloud image feature information from the satellite cloud image;
[0012] Call the pre-trained time series feature extraction model to extract the feature vector containing the time series features between the power from the power data;
[0013] Based on the cloud map features and feature vectors, the power prediction results of the decentralized photovoltaic power station at the next moment are predicted.
[0014] In one embodiment, based on the cloud map features and feature vectors, the power prediction result of the decentralized photovoltaic power station at the next moment is predicted, specifically:
[0015] Based on the latitude and longitude information of the photovoltaic power station, the cloud map feature is clipped into an area centered on the photovoltaic power station, and the area is reduced to a regional cloud map feature of a fixed size;
[0016] The regional cloud features are upgraded through the convolution operation of the same fixed size, and the upgraded regional cloud features are mapped into features with the same dimension as the feature vector;
[0017] The feature vector is fused and spliced with the features to obtain the prediction result of the distributed photovoltaic power at the next moment.
[0018] In one embodiment, a deep canonical correlation method is used to align feature vectors with features, and an LF fusion algorithm is used to fuse the aligned feature vectors with features to obtain a prediction result of the distributed photovoltaic power at the next moment.
[0019] In one embodiment, the cloud map feature extraction model includes sequentially connecting a convolutional neural network, a long short-term memory neural network, and an attention mechanism network.
[0020] In one embodiment, the pre-trained cloud image feature extraction model is obtained by training the cloud image feature extraction model based on historical satellite cloud images in a time series;
[0021] In one embodiment, the temporal feature extraction model includes sequentially connecting a gated recurrent unit neural network and a fully connected network layer.
[0022] In one embodiment, the pre-trained time series feature extraction model is obtained by training the time series feature extraction model based on power data in a time series.
[0023] A second aspect of the present invention provides an electronic device including a memory and a processor;
[0024] a memory for storing a computer program, wherein the computer program includes program instructions;
[0025] The processor is configured to execute the program instructions so as to enable the electronic device to perform the steps of a method for short-term power prediction of a decentralized photovoltaic power station as provided in the first aspect of the present invention.
[0026] A third aspect of the present invention provides a computer program product comprising program instructions, which, when executed by an electronic device, enables the electronic device to execute the steps of a method for short-term power prediction of a distributed photovoltaic power station as provided in the first aspect of the present invention.
[0027] A fourth aspect of the present invention provides a computer-readable storage medium, which includes a computer program. When the computer program is executed by one or more processors, it implements a short-term power prediction method for a distributed photovoltaic power station as provided in the first aspect of the present invention.
[0028] Compared with the prior art, the present invention has the following beneficial effects:
[0029] The present invention provides a method, device, medium, and product for short-term power prediction of distributed photovoltaic power stations. First, to capture the dynamic characteristics of clouds within a region, a pre-trained cloud feature extraction model is used to extract cloud features from satellite cloud images, including the distribution of attention weights between cloud feature information. In other words, the cloud features extracted from satellite cloud images by the pre-trained cloud feature extraction model enhance the ability to perceive cloud motion characteristics and capture global features, enabling the cloud features to include high-dimensional features such as cloud thickness, movement direction, and speed. Second, a pre-trained time series feature extraction model is used to extract feature vectors from power data, including information such as power change trends. Finally, based on the cloud features and feature vectors, the power prediction result for the distributed photovoltaic power station at the next moment is output. This solves the problem of inaccurate predictions for distributed photovoltaic power stations in rural areas, which can be caused by sudden changes in photovoltaic power due to cloud movement. BRIEF DESCRIPTION OF THE DRAWINGS
[0030] The drawings described herein are used to provide a further understanding of the embodiments of the present invention, constitute a part of this application, and do not constitute a limitation of the embodiments of the present invention. In the drawings:
[0031] Figure 1 A schematic flow chart of a method for short-term power prediction of a distributed photovoltaic power station provided by an embodiment of the present invention;
[0032] Figure 2 A schematic diagram of a neural network architecture for implementing short-term power prediction provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0033] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with examples and drawings. The exemplary embodiments of the present invention and their descriptions are only used to explain the present invention and are not intended to limit the present invention.
[0034] It should be noted that the terms "include" or "may include" used in various embodiments of the present application indicate the presence of the claimed function, operation or element, and do not limit the addition of one or more functions, operations or elements. In addition, as used in various embodiments of the present application, the terms "include", "have" and their cognates are only intended to indicate specific features, numbers, steps, operations, elements, components or combinations of the foregoing items, and should not be understood as first excluding the presence of one or more other features, numbers, steps, operations, elements, components or combinations of the foregoing items or the possibility of adding one or more features, numbers, steps, operations, elements, components or combinations of the foregoing items.
[0035] It should be understood that terms such as "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of the technical features indicated. Thus, a feature specified as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, "plurality" means two or more, unless otherwise specifically defined.
[0036] A first aspect of the present invention provides a method for short-term power prediction of a distributed photovoltaic power station, the method comprising:
[0037] S101, obtaining satellite cloud images and power data of photovoltaic power stations in a region at the current moment.
[0038] In this embodiment, satellite cloud images can be generated using data collected by meteorological stations. These images contain information such as irradiance intensity, cloud thickness, cloud movement direction, and cloud movement speed. This is common knowledge in the art and is not subject to further limitations in this embodiment. Generally, satellite cloud images are collected every hour, while photovoltaic power station power generation data is collected every 15 minutes.
[0039] S102, calling a pre-trained cloud image feature extraction model to extract cloud image features containing attention weight distribution between cloud image feature information from the satellite cloud image.
[0040] In this embodiment, the cloud image feature extraction model includes a convolutional neural network, a long short-term memory neural network and an attention mechanism network connected in sequence. The pre-trained cloud image feature extraction model is obtained by training the cloud image feature extraction model based on historical satellite cloud images in a time series.
[0041] Specifically, in combination with the data collection content described in the above embodiment, in order to align the time features, the historical satellite cloud images collected in a time series are now divided into m data sets according to the time length of the prediction window, and the m data sets are input into the cloud image feature extraction model for training.
[0042] Please refer to Figure 2 , connect the convolutional neural network and the long short-term memory neural network in sequence to form a convolutional long short-term memory network, namely CNN-LSTM. The network operation process is as follows:
[0043] Where: ⊙ represents the convolution operation, is the Hadamard product, σ represents the sigmoid activation function, S t is the satellite cloud image input at time t, H t and Ct are the hidden state feature vector and the cell state feature vector, respectively, and b represents the bias. The learnable weight matrix W s ,W h and W c Respectively with S t ,H t and C t As a recursive structure, CNN-LSTM enables the model to learn the cloud motion model in a time series manner. Each time step generates a cloud map feature with the same size as the input image.
[0044] The Attention module is as follows:
[0045] Q=XW Q , K=XW K , V=XW V , where: X is the input sequence, that is, the H sequence composed of each moment of the first m moments is obtained by CNN-LSTM t-m+1 :H t-1 The sequence of W Q ,W K and W V They are the query matrix, key matrix, and value matrix of Ht-m+1:Ht-1, which are m latent feature inputs. After flattening, the m data are flattened into m 1-dimensional arrays with an array dimension of d. Softmax represents scaling, and Attention() represents the attention weight operation including the calculation of weights. Represents the transpose operation in matrix operations.
[0046] The attention mechanism proposed in this embodiment analyzes historical factors in pixel-level satellite cloud imagery observations and predictions for photovoltaic power plants in rural areas. This mechanism mimics the human attention mechanism, focusing varying amounts of attention on satellite cloud images over different time spans. This reduces computational effort and accelerates the convergence of the cloud image feature extraction model. Furthermore, the fixed parameters in the attention mechanism gradually converge to appropriate values as model training deepens. This allows the weight of historical cloud image influence within a small region to gradually converge, increasing the accuracy of satellite cloud image predictions for the next moment.
[0047] It can be seen that this embodiment can obtain the predicted cloud image at the current time t based on the input satellite cloud image, determine the movement direction of the cloud layer, establish the impact relationship of the cloud layer on the photovoltaic power output, and consider the movement speed of the cloud layer.
[0048] S103: Call a pre-trained time series feature extraction model to extract a feature vector containing time series features between powers from the power data.
[0049] In this embodiment, the temporal feature extraction model includes a gated recurrent unit neural network and a fully connected network layer connected in sequence. The pre-trained temporal feature extraction model is obtained by training the temporal feature extraction model based on power data in a time series.
[0050] Specifically, based on the historical time series power data P t-m+1:t , use data cleaning and enhancement methods to preprocess the data and remove abnormal data. Then use GRU gated recurrent unit neural network to extract time series trend features. The last step of hidden state output h t Use fully connected network layers to compress into feature vectors Finally, the obtained Align the cloud features extracted in step S102 The GRU gated recurrent unit neural network combines the input gate and forget gate into one update gate compared to the LSTM. This has fewer parameters than the LSTM update gate and is faster to train:
[0051] Where: P t Represents the time series data at time t. h t ,z t and r t They are hidden state, update gate and reset gate respectively. p and W h Respectively represent the corresponding P t and h t This architecture can effectively learn temporary dependencies in the data.
[0052] S104: Based on the cloud map features and the feature vectors, a power prediction result of the distributed photovoltaic power station at the next moment is predicted.
[0053] In this embodiment, before predicting the power prediction result of the decentralized photovoltaic power station at the next moment using the cloud map features and feature vectors, the following processing is performed on the dimensions of the cloud map features:
[0054] First, based on the latitude and longitude information of the photovoltaic power station, the cloud feature is clipped to the area centered on the photovoltaic power station, and the area is reduced to a regional cloud feature of a fixed size;
[0055] Secondly, the regional cloud features are upgraded in dimension through a convolution operation of the same fixed size, and the upgraded regional cloud features are mapped into features with the same dimension as the feature vector.
[0056] Finally, the feature vector is fused and spliced with the features to obtain the prediction result of the decentralized photovoltaic power at the next moment.
[0057] Specifically, the cloud features containing attention are obtained back, Clipped into an area centered on the target PV plant Then, the average pooling layer is used to reduce the cloud features of these areas to a 2×2 image; the Conv2d convolution operation is used to convert the 2×2 image from C to 2C dimensions; finally, the cloud features after dimensionality increase are flattened and mapped to a one-dimensional vector through a fully connected layer.
[0058] Finally, the deep canonical correlation method is used to align the feature vectors with the features, and the LF fusion algorithm is used to fuse the aligned feature vectors with the features to obtain the prediction result of the decentralized photovoltaic power at the next moment.
[0059] Specifically, this embodiment uses the DCCA-LF module to fuse the steps S102 and S103 to obtain a high-dimensional feature vector containing cloud data and power data. and Among them, DCCA performs feature data alignment, while LF performs the data fusion process.
[0060] (1)DCCA
[0061] Where: MLPhead(·) is a multi-layer perceptron used to decode the fusion features Output multi-step forecast Data encoder parameter θ S ,θ P By minimizing the MSE loss L pred and DCCA loss L DCCA Optimize.
[0062] Where: L DCCA is used as a regularization term, Calculating satellite cloud image features and photovoltaic output time series characteristics The correlation between them, and the regularization coefficient λ C Control the strength of the effect of feature correlation.
[0063] (2) Late fusion method
[0064] LF is a simple fusion method that concatenates two features to form a new feature vector E t LF Thanks to simple splicing, LF can avoid the problem of two features interfering with each other at this stage.
[0065] The objectives of DCCA are as follows:
[0066] Where: The data encoder parameters θ are optimized by minimizing the DCCA loss S ,θ P , the correlation between two eigenvectors is calculated as follows, argmax represents the parameter when the function reaches its maximum value.
[0067] in,
[0068] make is the new matrix of the training set of size n, where I is the identity matrix and the matrix T t The solution is given by:
[0069] T t The components and eigenvectors of and The relationship is
[0070]
[0071] Where (·)' represents the matrix transpose operation, and the regularization coefficients r1 and r2 are both 10 -3 .
[0072] It should be noted that the model training process involved in this embodiment adopts conventional neural network training technology, such as using a loss function to update the network parameters of the model, so this embodiment will not be described in detail.
[0073] The present application also provides an electronic device. The electronic device includes a processor, a memory, a communication interface, and at least one communication bus for connecting the processor, the memory, and the communication interface. The memory includes, but is not limited to, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (PROM), or a compact disc read-only memory (CD-ROM), and is used for storing relevant instructions and data.
[0074] The communication interface is used to receive and send data. The processor can be one or more CPUs. If the processor is a single CPU, the CPU can be a single-core CPU or a multi-core CPU. The processor in the electronic device is used to read one or more programs stored in the memory and perform the following operations: obtain satellite cloud images and power data of photovoltaic power plants in a region at the current moment; call a pre-trained cloud image feature extraction model to extract cloud image features containing the attention weight distribution between cloud image feature information from the satellite cloud image; call a pre-trained time series feature extraction model to extract feature vectors containing time series features between power from the power data; and predict the power prediction results of the decentralized photovoltaic power plant at the next moment based on the cloud image features and feature vectors.
[0075] It should be noted that the specific implementation of each operation can be as described above. Figure 1 The corresponding description of the method embodiment shown is that the electronic device can be used to execute a short-term power prediction method for a decentralized photovoltaic power station of the above method embodiment of the present application, which will not be described in detail here.
[0076] An embodiment of the present invention further provides a computer-readable storage medium, which is a memory device in a computer device and is used to store programs and data. It is understood that the computer-readable storage medium herein may include both built-in storage media in the computer device and, of course, extended storage media supported by the computer device. The computer-readable storage medium provides storage space that stores the terminal's operating system. Furthermore, this storage space also stores one or more instructions suitable for being loaded and executed by a processor. These instructions may be one or more computer programs (including program code). It should be noted that the computer-readable storage medium herein may be a high-speed RAM memory or a non-volatile memory, such as at least one disk storage device. The processor may load and execute the one or more instructions stored in the computer-readable storage medium to implement the corresponding steps of the method for short-term power prediction of a distributed photovoltaic power station described in the above embodiment. Those skilled in the art will appreciate that the embodiments of the present invention may be provided as a method, system, or computer program product. Therefore, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program codes.
[0077] The present application also provides a computer program product containing program instructions. This computer program product can be software or a program product containing program instructions that can be executed on a computing device or stored on any available medium. When executed on at least one electronic device, the computer program product causes the at least one electronic device to execute a method for short-term power forecasting of a decentralized photovoltaic power plant.
[0078] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.
Claims
1. A method for short-term power prediction of a distributed photovoltaic power station, characterized in that the method include: Obtain satellite cloud images and power data of photovoltaic power stations in a region at the current moment; Call the pre-trained cloud image feature extraction model to extract cloud image features containing the attention weight distribution between cloud image feature information from the satellite cloud image; Call the pre-trained time series feature extraction model to extract the feature vector containing the time series features between the power from the power data; Based on the cloud map features and feature vectors, the power prediction results of the decentralized photovoltaic power station at the next moment are predicted.
2. A method for short-term power prediction of a distributed photovoltaic power station according to claim 1, characterized in that: Based on the cloud map features and eigenvectors, the power prediction results of the decentralized photovoltaic power station at the next moment are predicted as follows: Based on the latitude and longitude information of the photovoltaic power station, the cloud map feature is clipped into an area centered on the photovoltaic power station, and the area is reduced to a regional cloud map feature of a fixed size; The regional cloud features are upgraded through the convolution operation of the same fixed size, and the upgraded regional cloud features are mapped into features with the same dimension as the feature vector; The feature vector is fused and spliced with the features to obtain the prediction result of the distributed photovoltaic power at the next moment.
3. A method for short-term power prediction of a distributed photovoltaic power station according to claim 2, characterized in that: The deep canonical correlation method is used to align the feature vectors with the features, and the LF fusion algorithm is used to fuse the aligned feature vectors with the features to obtain the prediction result of the decentralized photovoltaic power at the next moment.
4. The method for short-term power prediction of a distributed photovoltaic power station according to claim 1, characterized in that: The cloud map feature extraction model includes sequentially connecting a convolutional neural network, a long short-term memory neural network and an attention mechanism network.
5. A method for short-term power prediction of a distributed photovoltaic power station according to claim 4, characterized in that: The pre-trained cloud image feature extraction model is obtained by training the cloud image feature extraction model based on historical satellite cloud images in a time series.
6. A method for short-term power prediction of a distributed photovoltaic power station according to claim 1, characterized in that: The temporal feature extraction model includes sequentially connecting a gated recurrent unit neural network and a fully connected network layer.
7. The method for short-term power prediction of a distributed photovoltaic power station according to claim 1, characterized in that: The pre-trained time series feature extraction model is obtained by training the time series feature extraction model based on historical power data in a time series.
8. An electronic device, characterized in that: including memory and processor; a memory for storing a computer program, wherein the computer program includes program instructions; The processor is configured to execute the program instructions so that the electronic device performs the steps of the method for short-term power prediction of a decentralized photovoltaic power station as described in any one of claims 1 to 7.
9. A computer program product comprising program instructions, characterized in that When the program instructions are executed by an electronic device, the electronic device executes the steps of the method for short-term power prediction of a decentralized photovoltaic power station according to any one of claims 1 to 7.
10. A computer-readable storage medium, characterized in that The computer-readable storage medium includes a computer program, and when the computer program is executed by one or more processors, the computer program implements the method for short-term power prediction of a decentralized photovoltaic power station according to any one of claims 1 to 7.