Direct photovoltaic power prediction method and device, medium and product

Abstract

The invention discloses a direct photovoltaic power prediction method and device, a medium and a product, and relates to the field of power prediction, and the method comprises the steps: building a direct photovoltaic power prediction model based on a multi-head self-attention mechanism and a long short-term memory network; a feature extraction layer in the direct photovoltaic power prediction model is used for performing satellite cloud picture data serialization and space-time embedding on the satellite cloud picture data; the meteorological data are mapped into vectors with the same dimension as the spatio-temporal embedding dimension, and two vector matrixes are combined to obtain multi-source spatio-temporal information embedding; the encoder is used for obtaining a plurality of attention feature matrixes through embedding according to the multi-source spatio-temporal information, and operating the attention feature matrixes to obtain a feature matrix; the long-short-term memory network is used for performing power prediction according to the characteristic matrix; and performing real-time photovoltaic power prediction by using the trained direct photovoltaic power prediction model according to the real-time meteorological data and the real-time satellite cloud picture data. According to the invention, the prediction accuracy can be effectively improved.

Description

Technical Field

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

[0002] At present, the ultra-short-term prediction of photovoltaic power generation based on satellite cloud images can be divided into indirect prediction method and direct prediction method. Direct prediction is widely used because of its fast and accurate prediction. The direct prediction method is to extract the spatiotemporal characteristics of the historical satellite cloud image sequence, combine it with meteorological data, and directly establish a power mapping model.

[0003] Regarding the research on direct prediction methods, existing literature uses satellite observation data of different wavelengths to extract its spatiotemporal sequence features, and combines it with representative numerical weather forecast (NWP) data to establish a power prediction model to capture future atmospheric conditions. Cheng L et al. spliced ​​the optical flow between satellite cloud images with the original data in the channel dimension, extracted features through convolutional neural networks, and established an encoding and decoding structure based on the attention mechanism to predict photovoltaic power. Existing literature proposed a cloud amount prediction model based on a 3D long short-term memory network (Long Short-Term Memory, LSTM), using convolutional self-attention to effectively capture historical features and using LSTM to predict future cloud image generalizations. After extracting the gray-level co-occurrence matrix of the cloud image, the existing literature used different functions to describe the image texture features, extracted cloud occlusion features, and predicted future cloud occlusion features through the autoregressive integrated moving average model (ARIMA model) based on historical cloud occlusion features. Existing literature extracts global distribution information of clouds from satellite images as additional input, and combines meteorological and power inputs to train the prediction model. Existing literature uses historical satellite data to convert clear, cloudy and cloud shadow pixels into cloud index, and uses historical fitting functions from various stations to map cloud index to clear sky index. From the contents recorded in the above literature, it can be seen that in the direct prediction method, since the photovoltaic power mapping model is directly established based on historical data, there is no need to consider the error of cloud image prediction, and its prediction accuracy often depends on the accuracy and completeness of spatiotemporal feature extraction. How to effectively extract spatiotemporal characteristics from multi-source historical satellite cloud image sequences and historical meteorological sequences, while avoiding the errors caused by cloud image prediction in indirect prediction, is a technical problem that needs to be solved urgently.

[0004] Therefore, based on the above problems, there is an urgent need to provide a new direct photovoltaic power prediction method or system that can effectively extract spatiotemporal characteristics from multi-source historical satellite cloud image sequences and historical meteorological sequences, while avoiding the errors caused by cloud image prediction in indirect prediction. Summary of the invention

[0005] The purpose of this application is to provide a direct photovoltaic power prediction method, equipment, medium and product, which can effectively extract spatiotemporal characteristics from multi-source historical satellite cloud image sequences and historical meteorological sequences, while avoiding the errors caused by cloud image prediction in indirect prediction, and effectively improving the accuracy of prediction.

[0006] To achieve the above objectives, this application provides the following solutions:

[0007] In a first aspect, the present application provides a direct photovoltaic power prediction method, the direct photovoltaic power prediction method comprising:

[0008] Obtain historical meteorological data and historical satellite cloud image data and corresponding photovoltaic power; satellite cloud image data include: cloud type, cloud top height, cloud top temperature, cloud thickness and infrared band cloud image;

[0009] A direct photovoltaic power prediction model is constructed based on a multi-head self-attention mechanism and a long short-term memory network; the direct photovoltaic power prediction model includes: a feature extraction layer, an encoder and a long short-term memory network; the feature extraction layer is used to serialize and spatiotemporally embed satellite cloud image data; and the meteorological data is mapped into a vector with the same dimension as the spatiotemporal embedding, and the two vector matrices are combined to obtain multi-source spatiotemporal information embedding; the encoder is used to obtain multiple attention feature matrices according to the multi-source spatiotemporal information embedding, and average and reduce the dimension of the multiple attention feature matrices to obtain a feature matrix; the long short-term memory network is used to perform power prediction according to the feature matrix;

[0010] The direct photovoltaic power prediction model is trained using historical meteorological data, historical satellite cloud image data and corresponding photovoltaic power to obtain a trained direct photovoltaic power prediction model;

[0011] According to the real-time meteorological data and real-time satellite cloud image data, the real-time photovoltaic power prediction is carried out using the trained direct photovoltaic power prediction model.

[0012] Optionally, the process of satellite cloud image data serialization and spatiotemporal embedding specifically includes:

[0013] Perform image segmentation on the satellite cloud image data to obtain multiple sub-images;

[0014] Vectorize the subgraph to obtain the corresponding vector matrix;

[0015] Use linear projection method to map each vector matrix to obtain embedded vectors;

[0016] Add each embedding vector to the positional encoding to obtain spatial feature embedding;

[0017] Intercept two consecutive frames of satellite cloud image data to create a fused cloud image of the area above the station;

[0018] The time dimension t and channel dimension c of the fused cloud images of two consecutive frames above the station are spliced ​​into the same dimension c×t, and the time series features are embedded into the high-dimensional space to obtain time feature embedding.

[0019] Optionally, a Flatten operation is employed for vectorization.

[0020] Alternatively, using formula E i '=E i +PE i Add position coding;

[0021] Among them, E i ' is the embedding vector after adding position encoding, E i is the embedding vector, is the position information of the i-th embedding, D is the D dimension, is a real number.

[0022] Optionally, the feature extraction layer is also used to perform preliminary contrast enhancement on the two-dimensional matrix data corresponding to the cloud type, cloud top height, cloud top temperature and cloud thickness in the satellite cloud image data to obtain a multi-source cloud feature map;

[0023] A style transfer algorithm is adopted, with the infrared band cloud image in satellite cloud image data as the content image, retaining the cloud phase characteristics, and the multi-source cloud feature map as the style image, and the color feature distribution is transferred to the fused feature image to achieve cloud image enhancement.

[0024] In a second aspect, the present application provides a direct photovoltaic power prediction device, the direct photovoltaic power prediction device comprising:

[0025] The historical data acquisition module is used to obtain historical meteorological data and historical satellite cloud image data and the corresponding photovoltaic power; the satellite cloud image data includes: cloud type, cloud top height, cloud top temperature, cloud thickness and infrared band cloud image;

[0026] A model building module is used to build a direct photovoltaic power prediction model based on a multi-head self-attention mechanism and a long short-term memory network; the direct photovoltaic power prediction model includes: a feature extraction layer, an encoder and a long short-term memory network; the feature extraction layer is used to serialize and spatiotemporally embed satellite cloud image data; and the meteorological data is mapped into a vector with the same dimension as the spatiotemporal embedding, and the two vector matrices are combined to obtain multi-source spatiotemporal information embedding; the encoder is used to obtain multiple attention feature matrices according to the multi-source spatiotemporal information embedding, and average and reduce the dimension of the multiple attention feature matrices to obtain a feature matrix; the long short-term memory network is used to perform power prediction according to the feature matrix;

[0027] A model training module is used to train a direct photovoltaic power prediction model using historical meteorological data, historical satellite cloud image data, and corresponding photovoltaic power to obtain a trained direct photovoltaic power prediction model;

[0028] The model prediction module is used to perform real-time photovoltaic power prediction based on real-time meteorological data and real-time satellite cloud image data using a trained direct photovoltaic power prediction model.

[0029] In a third aspect, the present application provides a computer device, comprising: 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 the direct photovoltaic power prediction method.

[0030] In a fourth aspect, the present application provides a computer-readable storage medium having a computer program stored thereon, and the computer program implements the direct photovoltaic power prediction method when executed by a processor.

[0031] In a fifth aspect, the present application provides a computer program product, including a computer program, which implements the direct photovoltaic power prediction method when executed by a processor.

[0032] According to the specific embodiments provided in this application, this application has the following technical effects:

[0033] The present application provides a direct photovoltaic power prediction method, device, medium and product, which constructs a direct photovoltaic power prediction model based on a multi-head self-attention mechanism and a long short-term memory network, and uses the feature extraction layer in the direct photovoltaic power prediction model to perform satellite cloud image data serialization and spatiotemporal embedding of meteorological data and satellite cloud image data, thereby achieving a comprehensive fusion of multi-source data and providing stronger data support for power prediction. According to the collaborative work of the multi-head self-attention mechanism and the long short-term memory network, it can accurately identify and focus on the key features in the prediction process, thereby effectively improving the accuracy of the prediction. The present application can effectively extract spatiotemporal characteristics from multi-source historical satellite cloud image sequences and historical meteorological sequences, while avoiding the errors caused by cloud image prediction in indirect prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0035] Figure 1 This is a schematic flow chart of a direct photovoltaic power prediction method in an embodiment of the present application;

[0036] Figure 2 This is a schematic diagram of the principle of the direct photovoltaic power prediction model;

[0037] Figure 3 Schematic diagram of the process of spatiotemporal embedding;

[0038] Figure 4 Schematic diagram of power prediction deviation results under different prediction models. DETAILED DESCRIPTION

[0039] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0040] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0041] In an exemplary embodiment, Figure 1As shown, a direct photovoltaic power prediction method is provided, which is executed by a computer device, and can be executed by a computer device such as a terminal or a server alone, or by a terminal and a server together. In the embodiment of the present application, the method includes the following S101 to S104. Among them:

[0042] S101, obtaining historical meteorological data and historical satellite cloud image data and corresponding photovoltaic power; the satellite cloud image data includes: cloud type, cloud top height, cloud top temperature, cloud thickness and infrared band cloud image;

[0043] S102, a direct photovoltaic power prediction model is constructed based on multi-head self-attention mechanism and long short-term memory network, and Figure 2 As shown; the direct photovoltaic power prediction model includes: a feature extraction layer, an encoder and a long short-term memory network; the feature extraction layer is used to serialize and spatiotemporally embed satellite cloud image data; and the meteorological data is mapped into a vector with the same dimension as the spatiotemporal embedding, and the two vector matrices are combined to obtain multi-source spatiotemporal information embedding; the encoder is used to obtain multiple attention feature matrices according to the multi-source spatiotemporal information embedding, and average and reduce the dimension of the multiple attention feature matrices to obtain a feature matrix; the long short-term memory network is used to perform power prediction according to the feature matrix;

[0044] like Figure 3 As shown in the figure, the process of satellite cloud image data serialization and spatiotemporal embedding specifically includes:

[0045] S1, the satellite cloud image data is segmented (Patch Partitioning) to obtain multiple sub-images; a color image I is taken as input, and the image is represented as a three-dimensional tensor Where H, W and C represent the height, width and number of channels of the image respectively. The image is evenly divided into N non-overlapping sub-images of size P×P, denoted as where i∈{1,2,...,N}. Each subgraph The dimension is P×P×C, and it has a local information image in the spatial dimension (P, P).

[0046] Specifically, the satellite cloud image data is divided from 6×10×10 into 25 6×2×2 sub-images;

[0047] S2, vectorize the subgraph to obtain the corresponding vector matrix;

[0048] Each subgraph I pi Convert f to a one-dimensional vector V by vectorizing the operation i :

[0049]

[0050] Specifically, the Flatten operation is used for vectorization, and the sub-graphs are flattened, so that multiple sub-graphs form a 25×24 vector matrix.

[0051] S3, use the linear projection method to map each vector matrix to obtain the embedded vector;

[0052] Each vector V i After a linear projection layer L, it is parameterized as a weight matrix and the bias vector Output a D-dimensional embedding vector E i :D is 128;

[0053] E i =L(V i )=V i ×W+b,

[0054] The above operations are performed on all sub-images of the input image to create a set of embeddings {Ei} to represent the information of the original image.

[0055] S4, add each embedding vector to the position code to obtain spatial feature embedding;

[0056] The position embeddings can be fixed or learnable parameters, which are added with the linearly projected embeddings to form the final position-aware embedding E i ', that is, the embedding vector after adding the position encoding, which is used to capture the order information of the sequence.

[0057] Using Formula E i '=E i +PE i Add position coding;

[0058] Among them, E i ' is the embedding vector after adding position encoding, E i is the embedding vector, is the position information of the i-th embedding, D is the D dimension, is a real number.

[0059] S5, intercepting two consecutive frames of satellite cloud image data to obtain a fused cloud image of the area above the station;

[0060] S6, concatenate the time dimension t and channel dimension c of the fused cloud images of two consecutive frames of the area above the station into the same dimension c×t, embed the time series features into the high-dimensional space, and obtain the time feature embedding.

[0061] The feature extraction layer is also used to perform preliminary contrast enhancement on the two-dimensional matrix data corresponding to cloud type, cloud top height, cloud top temperature and cloud thickness in the satellite cloud image data to obtain a multi-source cloud feature map;

[0062] A style transfer algorithm is adopted, with the infrared band cloud image in satellite cloud image data as the content image, retaining the cloud phase characteristics, and the multi-source cloud feature map as the style image, and the color feature distribution is transferred to the fused feature image to achieve cloud image enhancement.

[0063] The multi-head self-attention mechanism is called self-attention. In the prediction process, the correlation between inputs can be represented by attention. Self-attention is the basis for calculating correlation. It converts the input X into three different representations of Qi, Ki, and Vi through linear transformation:

[0064] Q i ,K i ,V i =X·W Q ,X·W K ,X·W V ;

[0065] By calculating Qi and The dot product between them is used to obtain the similarity score, which prevents the gradient from becoming unstable in the future, and the similarity score is factored Scaling, where d k represents the dimension of the key vector. The scaled similarity scores are normalized using the softmax function to obtain a probability distribution. The probability distribution represents the attention weights assigned to different elements in the key vector.

[0066]

[0067] In addition, multi-head attention is an extension of self-attention, which concatenates the attention outputs of individual heads to obtain different representation subspaces.

[0068] MultiHeadAttn(Q,K,V)=Concat(h1,...,h H )W o ;

[0069] where h1,...,h H For multiple heads, W o It is a learnable parameter matrix. The forward layer is used for feature conversion between self-attention layers to avoid the problem of lack of feature expression in linear calculation.

[0070] S103, using historical meteorological data, historical satellite cloud image data and corresponding photovoltaic power to train a direct photovoltaic power prediction model to obtain a trained direct photovoltaic power prediction model;

[0071] S104, performing real-time photovoltaic power prediction using a trained direct photovoltaic power prediction model according to real-time meteorological data and real-time satellite cloud image data.

[0072] In this application, image information is serialized by converting it into an embedded representation. The serialized data format increases the flexibility of the model in processing different image sizes and resolutions. At the same time, in order to capture the relative or absolute position information of the sub-image in the original image, a position embedding PE is introduced and added to the embedding of the linear projection to form the final position-aware embedding.

[0073] The self-attention mechanism can capture the long-range dependencies between different blocks in the image sequence and effectively obtain the dynamic features in the image time series. The self-attention mechanism is used to mine the spatiotemporal features of historical satellite cloud images. In the prediction process, the correlation between the inputs is represented by attention. i and The dot product between them is used to obtain the similarity score to prevent the gradient from becoming unstable in the future. In addition, multi-head attention is an extension of self-attention, which connects the attention outputs of a single head to obtain different representation subspaces.

[0074] The following is an explanation through a specific embodiment. In this implementation, the measured power, NWP data and satellite cloud image data of a photovoltaic power station (102.07° east longitude, 38.35° north latitude) from September 2017 to December 2018 for a total of 113 days are used for experimental analysis. The data set is in days, and each month contains about one week of data. In order to verify the effectiveness of the direct photovoltaic power prediction method proposed in this application, the root mean square error RMSE and the mean absolute error MAE are selected as evaluation indicators.

[0075] In order to verify the effectiveness of the prediction model proposed in this application, different models are used to predict the scheme of combining the fusion feature map with meteorological data proposed in this application. Figure 4 It can be seen that the CNN and MLP models have large fluctuations in the previous predictions. Since the convolution is very sensitive to the local field of view, the wrong cloud features are extracted, resulting in deviations in the model predictions. The GAN model generates prediction values ​​based on noise, which is more random, so the prediction error is larger. LSTM and CNN-LSTM can fit the actual power better, but due to the lack of capture of the global state, the prediction accuracy is not as good as this application. This application can not only obtain the global key features, but also improve the ability to predict time series features by connecting the LSTM model, so the prediction effect is the best.

[0076] In order to more intuitively display the prediction results and further verify the effectiveness of the method proposed in this application, the direct prediction model of this application was used to test the data of the four seasons of spring, summer, autumn and winter. The prediction results of the four seasons under fluctuating weather are shown in Table 1. It can be seen from the power prediction results in Table 1 that compared with using only meteorological data as input, the power prediction accuracy will be significantly improved after adding cloud map features. This is because clouds are the main reason for determining the strength of solar radiation reaching photovoltaic power stations. After adding cloud maps, the prediction model will become more sensitive to sudden changes in power, especially in cloudy weather. After using the cloud feature map, since the cloud features contained in the cloud feature map are more complete than the features extracted from the infrared cloud map, higher accuracy will be obtained. The fused cloud map of this application not only contains texture structures such as cloud phases, but also strengthens the characteristics of cloud type, cloud top height, cloud optical thickness, and cloud top temperature. The fused map has high integrity and highlights the features that have a greater impact on power, which has the most obvious improvement in power prediction effect. Comparative analysis of Table 1 shows that the fusion feature map proposed in this application is compared with the single meteorological data input, and the root mean square error is reduced by 2.95% and the average absolute error is reduced by 1.92% on the scale of 1h ahead prediction. Compared with the scheme of combining infrared cloud map with meteorological data, the root mean square error is reduced by 1.30% and the average absolute error is reduced by 0.70%. Compared with the input scheme of cloud feature map combined with meteorological data, the root mean square error is reduced by 0.88% and the average absolute error is reduced by 1.18%.

[0077] Table 1

[0078]

[0079] Based on the same inventive concept, the embodiment of the present application also provides a direct photovoltaic power prediction device for implementing the direct photovoltaic power prediction method involved above. The implementation scheme for solving the problem provided by the device is similar to the implementation scheme recorded in the above method, so the specific limitations in one or more direct photovoltaic power prediction device embodiments provided below can refer to the limitations of the direct photovoltaic power prediction method above, and will not be repeated here.

[0080] In an exemplary embodiment, a direct photovoltaic power prediction device is provided, comprising:

[0081] The historical data acquisition module is used to obtain historical meteorological data and historical satellite cloud image data and the corresponding photovoltaic power; the satellite cloud image data includes: cloud type, cloud top height, cloud top temperature, cloud thickness and infrared band cloud image;

[0082] A model building module is used to build a direct photovoltaic power prediction model based on a multi-head self-attention mechanism and a long short-term memory network; the direct photovoltaic power prediction model includes: a feature extraction layer, an encoder and a long short-term memory network; the feature extraction layer is used to serialize and spatiotemporally embed satellite cloud image data; and the meteorological data is mapped into a vector with the same dimension as the spatiotemporal embedding, and the two vector matrices are combined to obtain multi-source spatiotemporal information embedding; the encoder is used to obtain multiple attention feature matrices according to the multi-source spatiotemporal information embedding, and average and reduce the dimension of the multiple attention feature matrices to obtain a feature matrix; the long short-term memory network is used to perform power prediction according to the feature matrix;

[0083] A model training module is used to train a direct photovoltaic power prediction model using historical meteorological data, historical satellite cloud image data, and corresponding photovoltaic power to obtain a trained direct photovoltaic power prediction model;

[0084] The model prediction module is used to perform real-time photovoltaic power prediction based on real-time meteorological data and real-time satellite cloud image data using a trained direct photovoltaic power prediction model.

[0085] In an exemplary embodiment, a computer device is provided, which may be a server or a terminal. The computer device includes a processor, a memory, an input / output interface (I / O for short) and a communication interface. The processor, the memory and the input / output interface are connected via a system bus, and the communication interface is connected to the system bus via the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The input / output interface of the computer device is used to exchange information between the processor and an external device. The communication interface of the computer device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, a direct photovoltaic power prediction method is implemented.

[0086] In an exemplary embodiment, a computer-readable storage medium is provided, storing a computer program, and when the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0087] In an exemplary embodiment, a computer program product is provided, including a computer program, and when the computer program is executed by a processor, the steps in the above method embodiments are implemented.

[0088] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data must comply with relevant regulations.

[0089] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to the memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM may be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM).

[0090] The database involved in each embodiment provided in this application may include at least one of a relational database and a non-relational database. The non-relational database may include a distributed database based on blockchain, etc., but is not limited thereto. The processor involved in each embodiment provided in this application may be a general-purpose processor, a central processing unit, a graphics processor, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., but is not limited thereto.

[0091] In this application, all actions to obtain signals, information or data are carried out in compliance with the relevant data protection laws and policies of the country where they are located and with the authorization given by the owner of the corresponding device.

[0092] The technical features of the above embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0093] This article uses specific examples to illustrate the principles and implementation methods of this application. The description of the above embodiments is only used to help understand the method and core ideas of this application. At the same time, for those skilled in the art, according to the ideas of this application, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. A direct photovoltaic power prediction method, characterized in that: The direct photovoltaic power prediction method comprises: Obtain historical meteorological data and historical satellite cloud image data and corresponding photovoltaic power; satellite cloud image data include: cloud type, cloud top height, cloud top temperature, cloud thickness and infrared band cloud image; A direct photovoltaic power prediction model is constructed based on a multi-head self-attention mechanism and a long short-term memory network; the direct photovoltaic power prediction model includes: a feature extraction layer, an encoder and a long short-term memory network; the feature extraction layer is used to serialize and spatiotemporally embed satellite cloud image data; and the meteorological data is mapped into a vector with the same dimension as the spatiotemporal embedding, and the two vector matrices are combined to obtain multi-source spatiotemporal information embedding; the encoder is used to obtain multiple attention feature matrices according to the multi-source spatiotemporal information embedding, and average and reduce the dimension of the multiple attention feature matrices to obtain a feature matrix; the long short-term memory network is used to perform power prediction according to the feature matrix; The direct photovoltaic power prediction model is trained using historical meteorological data, historical satellite cloud image data and corresponding photovoltaic power to obtain a trained direct photovoltaic power prediction model; According to the real-time meteorological data and real-time satellite cloud image data, the real-time photovoltaic power prediction is carried out using the trained direct photovoltaic power prediction model.

2. The direct photovoltaic power prediction method according to claim 1, characterized in that: The process of satellite cloud image data serialization and spatiotemporal embedding specifically includes: Perform image segmentation on the satellite cloud image data to obtain multiple sub-images; Vectorize the subgraph to obtain the corresponding vector matrix; Use linear projection method to map each vector matrix to obtain embedded vectors; Add each embedding vector to the positional encoding to obtain spatial feature embedding; Intercept two consecutive frames of satellite cloud image data to create a fused cloud image of the area above the station; The time dimension t and channel dimension c of the fused cloud images of two consecutive frames above the station are spliced ​​into the same dimension c×t, and the time series features are embedded into the high-dimensional space to obtain time feature embedding.

3. The direct photovoltaic power prediction method according to claim 2, characterized in that: Use Flatten operation for vectorization.

4. The direct photovoltaic power prediction method according to claim 2, characterized in that: Using Formula E i '=E i +PE i Add position coding; Among them, E i ' is the embedding vector after adding position encoding, E i is the embedding vector, is the position information of the i-th embedding, D is the D dimension, is a real number.

5. The direct photovoltaic power prediction method according to claim 1, characterized in that: The feature extraction layer is also used to perform preliminary contrast enhancement on the two-dimensional matrix data corresponding to cloud type, cloud top height, cloud top temperature and cloud thickness in the satellite cloud image data to obtain a multi-source cloud feature map; A style transfer algorithm is adopted, with the infrared band cloud image in satellite cloud image data as the content image, retaining the cloud phase characteristics, and the multi-source cloud feature map as the style image, and the color feature distribution is transferred to the fused feature image to achieve cloud image enhancement.

6. A direct photovoltaic power prediction device, characterized in that: The direct photovoltaic power prediction device comprises: The historical data acquisition module is used to obtain historical meteorological data and historical satellite cloud image data and the corresponding photovoltaic power; the satellite cloud image data includes: cloud type, cloud top height, cloud top temperature, cloud thickness and infrared band cloud image; A model building module is used to build a direct photovoltaic power prediction model based on a multi-head self-attention mechanism and a long short-term memory network; the direct photovoltaic power prediction model includes: a feature extraction layer, an encoder and a long short-term memory network; the feature extraction layer is used to serialize and spatiotemporally embed satellite cloud image data; and the meteorological data is mapped into a vector with the same dimension as the spatiotemporal embedding, and the two vector matrices are combined to obtain multi-source spatiotemporal information embedding; the encoder is used to obtain multiple attention feature matrices according to the multi-source spatiotemporal information embedding, and average and reduce the dimension of the multiple attention feature matrices to obtain a feature matrix; the long short-term memory network is used to perform power prediction according to the feature matrix; A model training module is used to train a direct photovoltaic power prediction model using historical meteorological data, historical satellite cloud image data, and corresponding photovoltaic power to obtain a trained direct photovoltaic power prediction model; The model prediction module is used to perform real-time photovoltaic power prediction based on real-time meteorological data and real-time satellite cloud image data using a trained direct photovoltaic power prediction model.

7. A computer device comprising: 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 the direct photovoltaic power prediction method according to any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the direct photovoltaic power prediction method according to any one of claims 1 to 5 is implemented.

9. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the direct photovoltaic power prediction method according to any one of claims 1 to 5 is implemented.

Images