Solar photovoltaic power generation prediction methods and distributed photovoltaic power generation charging pile systems

Abstract

This invention relates to a solar photovoltaic (PV) power generation prediction method and a distributed PV power generation charging pile system, belonging to the field of PV power generation technology. It includes: acquiring satellite remote sensing cloud images and historical PV sequence data; inputting the satellite remote sensing cloud images into a complex cloud feature capture module to obtain remote sensing cloud image features; inputting the historical PV sequence data into a PV perception attention module to obtain PV perception attention features; fusing the remote sensing cloud image features and the PV perception attention features to obtain PV multimodal fusion features, followed by multi-dimensional information capture to obtain PV three-dimensional attention features, and finally, obtaining the predicted PV power generation through a PV power generation prediction module. The distributed PV power generation charging pile system includes a PV power generation module, an energy storage module, a power monitoring module, a power distribution control module, a charging pile module, a communication module, and a power management module. This invention can improve the prediction accuracy of the model and its robustness in complex scenarios.

Description

Technical Field

[0001] This invention belongs to the field of photovoltaic power generation technology, specifically relating to a solar photovoltaic power generation prediction method and a distributed photovoltaic power generation charging pile system. Background Technology

[0002] Solar photovoltaic (PV) power generation forecasting technology can be traced back to the rise of the new energy industry in the 20th century. Early PV power generation forecasting methods mainly relied on historical power generation data and statistical analysis of simple meteorological parameters, which had obvious limitations. Early models assumed a linear relationship between power generation variables, while PV power generation is affected by nonlinear factors such as nonlinear coupling of irradiance and temperature and sudden changes in cloud cover. At the data acquisition level, traditional methods heavily rely on limited parameters such as irradiance and cloud cover provided by a single ground meteorological station. At the same time, the update frequency of meteorological data is generally low, making it impossible to capture sudden changes in irradiance caused by short-term cloud movement.

[0003] With the development of computer analysis technology, early researchers mostly used machine learning algorithms such as linear regression and support vector machines to predict photovoltaic power generation. However, these methods have limited ability to model the complex nonlinear relationships in photovoltaic power generation. At the same time, traditional models lack the ability to model spatiotemporal correlations, cannot analyze cloud migration trajectories through satellite remote sensing images, and cannot build short-term solar photovoltaic power generation prediction models. As a result, existing short-term solar photovoltaic power generation prediction methods suffer from low accuracy and poor generalization ability.

[0004] Therefore, this invention proposes a solar photovoltaic power generation prediction method and a distributed photovoltaic power generation charging pile system to solve the problems of low power generation prediction accuracy and poor generalization ability of existing solar photovoltaic power generation prediction models. Summary of the Invention

[0005] To address the problems of low accuracy and insufficient multimodal information fusion capability of existing photovoltaic power generation prediction methods under complex weather conditions, this invention provides a solar photovoltaic power generation prediction method and a distributed photovoltaic power generation charging pile system.

[0006] To achieve the above objectives, the present invention employs the following technical solution:

[0007] This invention provides a method for predicting solar photovoltaic power generation, comprising the following steps:

[0008] S1. Acquire satellite remote sensing cloud images and historical photovoltaic sequence data;

[0009] S2. Construct a complex cloud feature capture module to capture satellite remote sensing cloud images of relevant areas and corresponding times. The data is input into the complex cloud feature capture module, which extracts features from both global and local information of the image, and outputs remote sensing cloud image features. ;

[0010] S3. Construct a photovoltaic sensing attention module to process historical photovoltaic sequence data. The input is fed into the photovoltaic sensing attention module, where it undergoes spatial and frequency domain feature extraction operations to output photovoltaic sensing attention features. ;

[0011] S4. Construct a photovoltaic multimodal feature fusion module to integrate remote sensing cloud image features. Photovoltaic perception attention characteristics The input is fed into the photovoltaic multimodal feature fusion module, and after feature fusion operation, the photovoltaic multimodal fused features are output. ;

[0012] S5. Construct a 3D attention module for photovoltaics to fuse multimodal features. The input is fed into the photovoltaic 3D attention module for multi-dimensional information capture, and the photovoltaic 3D attention features are output. ;

[0013] S6. Construct a photovoltaic power generation prediction module, incorporating photovoltaic three-dimensional attention features. Inputting the data into the photovoltaic power generation prediction module yields the predicted photovoltaic power generation. .

[0014] Furthermore, step S2 specifically includes:

[0015] The complex cloud feature capture module includes a 1D convolutional layer, a reshaping layer, a layer normalization layer, and a cloud convolutional attention module; the kernel size of the 1D convolutional layer is [missing information]. The cloud convolutional attention module includes a first branch, a second branch, and a third branch in parallel.

[0016] S21. Develop satellite remote sensing cloud images of the relevant area and corresponding time. The input is fed into the complex cloud feature capture module, and the first convolutional feature is obtained after passing through a 1D convolutional layer. The first convolutional feature The first convolutional features are reshaped by the remodeling layer. Spatial Dimensions Merged into sequence length Obtain the first reshaping feature The first reshaping feature After performing layer normalization operations, the first layer normalized features are obtained. The formula is expressed as follows:

[0017] ,

[0018] ,

[0019] ,

[0020] in, Indicates the altitude of cloud layers in satellite remote sensing images. Indicates the width of a satellite remote sensing cloud image. This indicates the number of channels in a satellite remote sensing cloud image. This indicates the number of channels in a satellite remote sensing cloud image after processing. Indicates the kernel size as 1D convolutional layer operations, This indicates a reshaping operation. This represents the transpose of a tensor matrix. Presentation layer normalization operation;

[0021] S22. Normalize the first layer of features. The input is fed into the cloud convolutional attention module. In the first branch, the first layer of normalized features... The feature tensor of the first half of the channels is extracted through slicing to obtain the first slice feature. The first slice feature Expanding the channel through a linear layer The first extended feature is obtained by multiplying the original feature by 1. First extended feature After activation function Processing yields global features. The formula is expressed as follows:

[0022] ,

[0023] ,

[0024] ,

[0025] in, Indicates the feature Slice the first half of the channel count. and These represent the first and second learnable parameters in the linear layer, respectively. express Activation function operations;

[0026] S23. In the second branch, features Feature extraction after slicing The feature tensor of the last half of the channels is used to output the second slice feature. The second slice feature Expanding the channel through a linear layer This yields the second extended feature. The second extended feature After convolution kernel size is The 1D convolutional layer has a kernel size of [missing information]. 1D depthwise separable convolutional layers are used to obtain local features. The local features After layer normalization, the second layer normalized features are obtained. The formula is expressed as follows:

[0027] ,

[0028] ,

[0029] ,

[0030] ,

[0031] in, Indicates the feature Slicing is performed on the last half of the channel count. and These represent the third and fourth learnable parameters in the linear layer, respectively. Indicates the kernel size as 1D convolutional layer, Indicates the kernel size as 1D depth-separable convolutional layers;

[0032] S24. In the third branch, the feature The channel is expanded through a linear layer to... This yields the third extended feature. Third extended feature The spatial dimensions were restored after the reshaping process. The second reshaping feature is obtained. The second reshaping feature After convolution kernel size is 2D depthwise separable convolutional layers, activation functions and the kernel size is The 2D convolutional layer is processed to obtain the second convolutional feature. The formula is expressed as follows:

[0033] ,

[0034] ,

[0035] ,

[0036] in, and These represent the fifth and sixth learnable parameters in the linear layer, respectively. Indicates features Spatial dimension restored to Operation; Indicates the kernel size as 2D convolutional layers, Indicates the kernel size as 2D depth-separable convolutional layers;

[0037] S25. Global features With the second layer of normalized features Perform product fusion to obtain the first fusion feature. ; the first fusion feature With the second convolution feature By splicing and merging the features, we can obtain the cloud convolutional attention features. The satellite remote sensing cloud image Convolutional attention features of clouds By stitching and fusion, the features of the remote sensing cloud image are obtained. The formula is expressed as follows:

[0038] ,

[0039] ,

[0040] ,

[0041] in, This indicates a feature concatenation operation. This indicates a product fusion operation.

[0042] Furthermore, step S3 specifically includes:

[0043] The photovoltaic sensing attention module includes a photovoltaic frequency domain attention module and a photovoltaic spatial attention module;

[0044] S31. Historical Data of Photovoltaic Sequences After convolution kernel size 1D convolutional layers, and after Activation function processing, outputting third convolutional features The formula is expressed as follows:

[0045] ,

[0046] in, Indicates the kernel size as 1D convolutional layer; the third convolutional feature The low-frequency features are obtained by global low-frequency feature capture through the photovoltaic frequency domain attention module. High-frequency features are obtained by calculating using the residual method. ; for low-frequency characteristics and high frequency characteristics Weighted fusion is performed to obtain the photovoltaic frequency domain attention features. The formula is expressed as follows:

[0047] ,

[0048] ,

[0049] ,

[0050] in, Represented as features The sequence length, ; and These represent the first and second learnable weight parameters obtained through backpropagation during model training. The multiplication operation represents the matrix element multiplication; the photovoltaic frequency domain attention feature After convolution kernel size is The 1D convolutional layer yields the fourth convolutional feature. The fourth convolutional feature After slicing, the feature tensors of the first half of the channels are extracted to obtain the features of the third slice. The fourth convolutional feature After slicing, half of the channel-count feature tensors are extracted. This yields the fourth slice feature tensor. The formula is expressed as follows:

[0051] ,

[0052] ,

[0053] ,

[0054] in, Representation of features The sequence length, Representation of features The number of channels; Specifically, it refers to... Slice along the first half of the channel dimension; Specifically, it refers to... Slice along the last half of the channel dimension;

[0055] S32. In the photovoltaic spatial attention module, the third slice feature After global average pooling, the kernel size is... 1D depthwise separable convolution and Activation function to obtain fifth convolutional features The fifth convolutional feature Features of the third slice By splicing and merging, the second fusion feature is obtained. After global average pooling, the kernel size is... 1D depthwise separable convolution and Activation function to obtain the sixth convolutional feature The sixth convolutional feature With the second fusion feature By splicing and merging, the third fusion feature is obtained. The third fusion feature Features of the third slice By adding each element, we obtain the first comprehensive feature. The formula is expressed as follows:

[0056] ,

[0057] ,

[0058] ,

[0059] ,

[0060] ,

[0061] in, express Activation function This indicates a global average pooling operation. This indicates an element-wise addition operation. and Indicates the kernel size as The 1D depthwise separable convolution and convolution kernel size are The 1D depthwise separable convolution; similarly, the fourth slice feature After the above third slice feature Following the same steps, the second comprehensive feature is obtained. The first comprehensive feature Second comprehensive features By splicing and merging, photovoltaic sensing attention can be obtained. The formula is expressed as: .

[0062] Furthermore, step S4 specifically includes:

[0063] The remote sensing cloud image features Photovoltaic perception attention characteristics Input into the photovoltaic multimodal feature fusion module, remote sensing cloud image features After convolution kernel size is 2D transposed convolutional layers for features Perform high-resolution restoration and apply ReLU activation function for non-linear processing to output the seventh convolutional feature. Photovoltaic sensing attention features With the seventh convolution feature Perform feature concatenation along the channel dimension to obtain the concatenated features. The splicing features are processed by a convolution kernel with a size of [missing value]. 2D depthwise separable convolution, then using The activation function is subjected to nonlinear weighting to obtain the photovoltaic multimodal fusion characteristics. The formula is expressed as follows:

[0064] ,

[0065] ,

[0066] ,

[0067] in, Represents the ReLU activation function. Indicates the kernel size as 2D transposed convolutional layer, Indicates the kernel size as 2D depthwise separable convolution.

[0068] Furthermore, step S5 specifically includes:

[0069] The photovoltaic multimodal fusion feature The input is fed into the photovoltaic 3D attention module to fuse photovoltaic multimodal features. Use along the channel dimension The operation divides the data into 4 sub-features along the channels, with each sub-feature having 10 channels. ,in Features The number of channels is expressed by the following formula:

[0070] ,

[0071] in, This indicates a segmentation feature operation along the channel. These represent the first initial sub-feature, the second initial sub-feature, the third initial sub-feature, and the fourth initial sub-feature, respectively; the first initial sub-feature The first sub-feature is obtained by preserving sub-features through skip connections. ; the second initial sub-feature After convolution kernel size is 2D convolutional layers and using The activation function performs non-linear processing, followed by a convolution kernel with a size of [missing value]. The 2D depthwise separable convolutional layer yields the second sub-feature. ; the second sub-feature With the third initial sub-feature By splicing and blending, a fourth fusion feature is obtained. The fourth fusion feature After convolution kernel size is 2D convolutional layers and using The activation function performs non-linear processing, followed by a convolution kernel with a size of [missing value]. The 2D depthwise separable convolutional layer yields the third sub-feature. The third sub-feature With the fourth initial sub-feature By splicing and merging, the fifth fusion feature is obtained. The fifth fusion feature After convolution kernel size is 2D convolutional layers and using The activation function performs non-linear processing, followed by a convolution kernel with a size of [missing value]. The 2D depthwise separable convolutional layer yields the fourth sub-feature. The formula is expressed as follows:

[0072] ,

[0073] ,

[0074] ,

[0075] ,

[0076] ,

[0077] in, Indicates the kernel size as 2D convolutional layers, and These represent the kernel size as follows: The 2D convolutional layer and the kernel size are 2D depthwise separable convolutional layers Indicates the kernel size as 2D depthwise separable convolutional layers; sub-features The components are spliced ​​and merged, and the kernel size is [missing information]. 2D convolutional layers are used to obtain photovoltaic 3D attention features. The formula is expressed as follows:

[0078] ,

[0079] in, Indicates the kernel size as 2D convolutional layers.

[0080] Furthermore, step S6 specifically includes:

[0081] Photovoltaic three-dimensional attention features The input is fed into the photovoltaic power generation prediction module to generate photovoltaic 3D attention features. Flattened into a one-dimensional vector, then passed through a fully connected layer and Activation function to calculate predicted photovoltaic power generation The formula is expressed as follows:

[0082] ,

[0083] in, This indicates a fully connected operation. This represents the vector flattening operation. and These represent the third and fourth learnable weight parameters obtained by backpropagation of the model, respectively.

[0084] Furthermore, the mean squared error loss is used as the loss function, and the calculation process of this loss function is as follows:

[0085] ,

[0086] in, This represents the calculated value of the loss function. The number of data samples, This is expressed as the actual photovoltaic power value. The sample is The photovoltaic power value predicted by the model at that time.

[0087] The present invention also provides a distributed photovoltaic power generation charging pile system, including a photovoltaic power generation module, an energy storage module, a power monitoring module, a power distribution control module, a charging pile module, a communication module, and a power management module;

[0088] The output terminal of the photovoltaic power generation module is connected to the input terminal of the energy storage module, and the output terminal of the energy storage module is connected to...

[0089] The input terminal of the power monitoring module is connected to the power distribution control module; the output terminal of the power monitoring module is connected to the input terminal of the power distribution control module; the output terminal of the energy storage module is connected to the input terminal of the charging pile module; the input terminal of the communication module is connected to the output terminals of the power monitoring module and the power distribution control module respectively; the output terminal of the communication module is connected to the cloud; the output terminal of the power distribution control module is connected to the user's power system; the input terminal of the charging pile module is also connected to the output terminal of the municipal power grid; the input terminal of the power consumption management module is connected to the output terminals of the power distribution control module and the municipal power grid; the output terminal of the power consumption management module is connected to the input terminal of the charging pile module and the power system; the power distribution control module is used to execute the solar photovoltaic power generation prediction method.

[0090] The advantages of this invention are:

[0091] This invention fully utilizes multi-source heterogeneous information such as satellite remote sensing meteorological images and photovoltaic system sensing data. Through a complex cloud feature capture module, it performs multi-scale feature analysis on satellite remote sensing images, deeply mining the dynamic correlation between cloud morphology evolution and light intensity, significantly improving the ability to extract key features under complex meteorological conditions. The designed photovoltaic sensing attention module employs a spatiotemporal frequency domain joint modeling mechanism to effectively capture the long-term and short-term temporal dependence features of photovoltaic power generation data, enhancing the dynamic perception capability of the photovoltaic system's operating status. The proposed photovoltaic multimodal feature fusion module integrates and supplements heterogeneous data features. The photovoltaic 3D attention module in the method fully captures the deep-level information of multimodal data, significantly improving the robustness of the model in complex scenarios. Attached Figure Description

[0092] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0093] Figure 1 This is a flowchart of the steps of the method of the present invention;

[0094] Figure 2 This is a system block diagram of the distributed photovoltaic power generation charging pile system proposed in this invention;

[0095] Figure 3 This is a schematic diagram of the distributed photovoltaic power generation charging pile system proposed in this invention. Detailed Implementation

[0096] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0097] Example 1

[0098] In this embodiment, as Figure 1 As shown, this invention provides a method for predicting solar photovoltaic power generation, the specific steps of which include:

[0099] S1. Acquire satellite remote sensing cloud images and historical photovoltaic sequence data.

[0100] S2. Construct a complex cloud feature capture module to capture satellite remote sensing cloud images of relevant areas and corresponding times. The data is input into the complex cloud feature capture module, which extracts features from both global and local information of the image, and outputs remote sensing cloud image features. .

[0101] Specifically, the complex cloud feature capture module includes a 1D convolutional layer, a reshaping layer, a layer normalization layer, and a cloud convolutional attention module; the kernel size of the 1D convolutional layer is [missing information]. The cloud convolutional attention module includes a first branch, a second branch, and a third branch in parallel.

[0102] S21. Develop satellite remote sensing cloud images of the relevant area and corresponding time. The input is fed into the complex cloud feature capture module, and the first convolutional feature is obtained after passing through a 1D convolutional layer. The first convolutional feature The first convolutional features are reshaped by the remodeling layer. Spatial Dimensions Merged into sequence length Obtain the first reshaping feature The first reshaping feature After performing layer normalization operations, the first layer normalized features are obtained. The formula is expressed as follows:

[0103] ,

[0104] ,

[0105] ,

[0106] in, Indicates the altitude of cloud layers in satellite remote sensing images. Indicates the width of a satellite remote sensing cloud image. This indicates the number of channels in a satellite remote sensing cloud image. This indicates the number of channels in a satellite remote sensing cloud image after processing. Indicates the kernel size as 1D convolutional layer operations, This indicates a reshaping operation. This represents the transpose of a tensor matrix. Presentation layer normalization operation;

[0107] S22. Normalize the first layer of features. The input is fed into the cloud convolutional attention module. In the first branch, the first layer of normalized features... The feature tensor of the first half of the channels is extracted through slicing to obtain the first slice feature. The first slice feature Expanding the channel through a linear layer The first extended feature is obtained by multiplying the original feature by 1. First extended feature After activation function Processing yields global features. The formula is expressed as follows:

[0108] ,

[0109] ,

[0110] ,

[0111] in, Indicates the feature Slice the first half of the channel count. and These represent the first and second learnable parameters in the linear layer, respectively. express Activation function operations;

[0112] S23. In the second branch, features Feature extraction after slicing The feature tensor of the last half of the channels is used to output the second slice feature. The second slice feature Expanding the channel through a linear layer This yields the second extended feature. The second extended feature After convolution kernel size is The 1D convolutional layer has a kernel size of [missing information]. 1D depthwise separable convolutional layers are used to obtain local features. The local features After layer normalization, the second layer normalized features are obtained. The formula is expressed as follows:

[0113] ,

[0114] ,

[0115] ,

[0116] ,

[0117] in, Indicates the feature Slicing is performed on the last half of the channel count. and These represent the third and fourth learnable parameters in the linear layer, respectively. Indicates the kernel size as 1D convolutional layer, Indicates the kernel size as 1D depth-separable convolutional layers;

[0118] S24. In the third branch, the feature The channel is expanded through a linear layer to... This yields the third extended feature. Third extended feature The spatial dimensions were restored after the reshaping process. The second reshaping feature is obtained. The second reshaping feature After convolution kernel size is 2D depthwise separable convolutional layers, activation functions and the kernel size is The 2D convolutional layer is processed to obtain the second convolutional feature. The formula is expressed as follows:

[0119] ,

[0120] ,

[0121] ,

[0122] in, and These represent the fifth and sixth learnable parameters in the linear layer, respectively. Indicates features Spatial dimension restored to Operation; Indicates the kernel size as 2D convolutional layers, Indicates the kernel size as 2D depth-separable convolutional layers;

[0123] S25. Global features With the second layer of normalized features Perform product fusion to obtain the first fusion feature. ; the first fusion feature With the second convolution feature By splicing and merging the features, we can obtain the cloud convolutional attention features. The satellite remote sensing cloud image Convolutional attention features of clouds By stitching and fusion, the features of the remote sensing cloud image are obtained. The formula is expressed as follows:

[0124] ,

[0125] ,

[0126] ,

[0127] in, This indicates a feature concatenation operation. This indicates a product fusion operation.

[0128] S3. Construct a photovoltaic sensing attention module to process historical photovoltaic sequence data. The input is fed into the photovoltaic sensing attention module, where it undergoes spatial and frequency domain feature extraction operations to output photovoltaic sensing attention features. .

[0129] Specifically, the photovoltaic sensing attention module includes a photovoltaic frequency domain attention module and a photovoltaic spatial attention module;

[0130] S31. Historical Data of Photovoltaic Sequences After convolution kernel size 1D convolutional layers, and after Activation function processing, outputting third convolutional features The formula is expressed as follows:

[0131] ,

[0132] in, Indicates the kernel size as 1D convolutional layer; the third convolutional feature The low-frequency features are obtained by global low-frequency feature capture through the photovoltaic frequency domain attention module. High-frequency features are obtained by calculating using the residual method. ; for low-frequency characteristics and high frequency characteristics Weighted fusion is performed to obtain the photovoltaic frequency domain attention features. The formula is expressed as follows:

[0133] ,

[0134] ,

[0135] ,

[0136] in, Represented as features The sequence length, ; and These represent the first and second learnable weight parameters obtained through backpropagation during model training. The multiplication operation represents the matrix element multiplication; the photovoltaic frequency domain attention feature After convolution kernel size is The 1D convolutional layer yields the fourth convolutional feature. The fourth convolutional feature After slicing, the feature tensors of the first half of the channels are extracted to obtain the features of the third slice. The fourth convolutional feature After slicing, half of the channel-numbered feature tensors are extracted, resulting in the fourth slice feature. The formula is expressed as follows:

[0137] ,

[0138] ,

[0139] ,

[0140] in, Representation of features The sequence length, Representation of features The number of channels; Specifically, it refers to... Slice along the first half of the channel dimension; Specifically, it refers to... Slice along the last half of the channel dimension;

[0141] S32. In the photovoltaic spatial attention module, the third slice feature After global average pooling, the kernel size is... 1D depthwise separable convolution and Activation function to obtain fifth convolutional features The fifth convolutional feature Features of the third slice By splicing and merging, the second fusion feature is obtained. After global average pooling, the kernel size is... 1D depthwise separable convolution and Activation function to obtain the sixth convolutional feature The sixth convolutional feature With the second fusion feature By splicing and merging, the third fusion feature is obtained. The third fusion feature Features of the third slice By adding each element, we obtain the first comprehensive feature. The formula is expressed as follows:

[0142] ,

[0143] ,

[0144] ,

[0145] ,

[0146] ,

[0147] in, express Activation function This indicates a global average pooling operation. This indicates an element-wise addition operation. and Indicates the kernel size as The 1D depthwise separable convolution and convolution kernel size are The 1D depthwise separable convolution; similarly, the fourth slice feature After the above third slice feature Following the same steps, the second comprehensive feature is obtained. The first comprehensive feature Second comprehensive features By splicing and merging, photovoltaic sensing attention can be obtained. The formula is expressed as: .

[0148] S4. Construct a photovoltaic multimodal feature fusion module to integrate remote sensing cloud image features. Photovoltaic perception attention characteristics The input is fed into the photovoltaic multimodal feature fusion module, and after feature fusion operation, the photovoltaic multimodal fused features are output. .

[0149] Specifically, the remote sensing cloud image features Photovoltaic perception attention characteristics Input into the photovoltaic multimodal feature fusion module, remote sensing cloud image features After convolution kernel size is 2D transposed convolutional layers for features Perform high-resolution restoration and apply ReLU activation function for non-linear processing to output the seventh convolutional feature. Photovoltaic sensing attention features With the seventh convolution feature Perform feature concatenation along the channel dimension to obtain the concatenated features. The splicing features are processed by a convolution kernel with a size of [missing value]. 2D depthwise separable convolution, then using The activation function is subjected to nonlinear weighting to obtain the photovoltaic multimodal fusion characteristics. The formula is expressed as follows:

[0150] ,

[0151] ,

[0152] ,

[0153] in, Represents the ReLU activation function. Indicates the kernel size as 2D transposed convolutional layer, Indicates the kernel size as 2D depthwise separable convolution.

[0154] S5. Construct a 3D attention module for photovoltaics to fuse multimodal features. The input is fed into the photovoltaic 3D attention module for multi-dimensional information capture, and the photovoltaic 3D attention features are output. .

[0155] Specifically, the photovoltaic multimodal fusion feature The input is fed into the photovoltaic 3D attention module to fuse photovoltaic multimodal features. Use along the channel dimension The operation divides the data into 4 sub-features along the channels, with each sub-feature having 10 channels. ,in Features The number of channels is expressed by the following formula:

[0156] ,

[0157] in, This indicates a segmentation feature operation along the channel. These represent the first initial sub-feature, the second initial sub-feature, the third initial sub-feature, and the fourth initial sub-feature, respectively; the first initial sub-feature The first sub-feature is obtained by preserving sub-features through skip connections. ; the second initial sub-feature After convolution kernel size is 2D convolutional layers and using The activation function performs non-linear processing, followed by a convolution kernel with a size of [missing value]. The 2D depthwise separable convolutional layer yields the second sub-feature. ; the second sub-feature With the third initial sub-feature By splicing and blending, a fourth fusion feature is obtained. The fourth fusion feature After convolution kernel size is 2D convolutional layers and using The activation function performs non-linear processing, followed by a convolution kernel with a size of [missing value]. The 2D depthwise separable convolutional layer yields the third sub-feature. The third sub-feature With the fourth initial sub-feature By splicing and merging, the fifth fusion feature is obtained. The fifth fusion feature After convolution kernel size is 2D convolutional layers and using The activation function performs non-linear processing, followed by a convolution kernel with a size of [missing value]. The 2D depthwise separable convolutional layer yields the fourth sub-feature. The formula is expressed as follows:

[0158] ,

[0159] ,

[0160] ,

[0161] ,

[0162] ,

[0163] in, Indicates the kernel size as 2D convolutional layers, and These represent the kernel size as follows: The 2D convolutional layer and the kernel size are 2D depthwise separable convolutional layers Indicates the kernel size as 2D depthwise separable convolutional layers; sub-features The components are spliced ​​and merged, and the kernel size is [missing information]. 2D convolutional layers are used to obtain photovoltaic 3D attention features. The formula is expressed as follows:

[0164] ,

[0165] in, Indicates the kernel size as 2D convolutional layers.

[0166] S6. Construct a photovoltaic power generation prediction module, incorporating photovoltaic three-dimensional attention features. Inputting the data into the photovoltaic power generation prediction module yields the predicted photovoltaic power generation. .

[0167] Specifically, the photovoltaic three-dimensional attention features The input is fed into the photovoltaic power generation prediction module to generate photovoltaic 3D attention features. Flattened into a one-dimensional vector, then passed through a fully connected layer and Activation function to calculate predicted photovoltaic power generation The formula is expressed as follows:

[0168] ,

[0169] in, This indicates a fully connected operation. This represents the vector flattening operation. and These represent the third and fourth learnable weight parameters obtained by backpropagation of the model, respectively.

[0170] Specifically, the mean squared error loss is used as the loss function, and the calculation process of this loss function is as follows:

[0171] ,

[0172] in, This represents the calculated value of the loss function. The number of data samples, This is expressed as the actual photovoltaic power value. The sample is The photovoltaic power value predicted by the model at that time.

[0173] Example 2

[0174] To verify the effectiveness of the solar photovoltaic power generation prediction method proposed in this invention in the field of photovoltaic power generation prediction, the proposed solar photovoltaic power generation prediction method was compared with existing photovoltaic power generation prediction methods under the same experimental conditions. The photovoltaic power generation prediction results of each model in the comparative experiment were compared and analyzed to verify the effectiveness of the solar photovoltaic power generation prediction method proposed in this invention in the field of photovoltaic power generation prediction.

[0175] In the comparative experiments of this invention, three mainstream photovoltaic power generation time series prediction model methods were selected as comparative experimental models with the proposed method. The three comparative models used in the experiment are as follows: 1. The LSTM model is based on a recurrent neural network architecture, which can effectively capture long-term dependencies in time series data, but it is prone to gradient vanishing problem when processing long sequences and has low computational efficiency; 2. The Transformer model adopts a self-attention mechanism, which is good at modeling global dependencies in sequence data and supports parallel computing, but its computational complexity increases quadratically with the sequence length and has high requirements for the amount and quality of training data; 3. The ViT model applies the Transformer architecture to image patch sequence modeling, which performs well in global feature extraction, but lacks the ability to finely model local spatial details.

[0176] In the comparative experiment of the proposed solar photovoltaic power generation prediction method, three experimental indicators were used to evaluate the model's performance: RMSE (Root Mean Square Error Percentage), which measures the absolute error of the model's prediction by calculating the root mean square of the squared difference between the predicted and actual values; a smaller RMSE value indicates higher prediction accuracy. MAE (Mean Absolute Error Percentage), which calculates the average of the absolute errors between the predicted and actual values, directly reflects the actual scale of the prediction error. MAPE (Mean Absolute Percentage Error), which calculates the average of the absolute values ​​of the relative errors between the predicted and actual values. RMSE and MAE are used to directly evaluate the model's prediction error at different time steps (0–4 hours), reflecting the model's ability to capture short- and medium-term photovoltaic power fluctuations. MAPE is used to measure the relative error stability of the model under different weather conditions (sunny, cloudy, overcast).

[0177] The dataset used in the experiment is an existing dataset - the multi-modal photovoltaic power prediction dataset. This dataset is jointly collected through satellite remote sensing technology and ground monitoring equipment, covering the centralized photovoltaic power station in Australia (BP Station) and 9 distributed photovoltaic sites in China (S1 - S9 Stations). The data content includes time series data such as photovoltaic power, global horizontal irradiance (GHI), and solar zenith angle with high spatio-temporal resolution, as well as dual-channel remote sensing images of infrared spectrum (IR) and short-wave radiation (SWR) provided by the Japanese Himawari-8 satellite. The dataset is divided into three types of scenarios according to weather types: sunny days (CSI > 0.9), cloudy days (0.3 ≤ CSI ≤ 0.9), and overcast days (0 < CSI < 0.3), covering multi-season continuous observation data from 2020 - 2022 (BP Station) and 2022 - 2023 (Chinese sites). In the experiment, the dataset is divided into a training set, a test set, and a validation set in a ratio of 8:1:1. Among them, the BP Station contains 66,365 samples, and each Chinese site contains 31,395 samples.

[0178] The experimental results of the solar photovoltaic power generation prediction method proposed in this invention and the experimental comparison models on the data of BP Station and S_{1}-S_{9} Stations are shown in Table 1 and Table 2.

[0179] Table 1 Comparison results of models in the BP Station dataset

[0180]

[0181] Table 2 Comparison results of models in the S_{1}-S_{9} Stations dataset

[0182]

[0183] In the data of the BP Station, the RMSE, MAE, and MAPE indicators of the proposed method are 19.13%, 9.82%, and 8.47% respectively. All three indicators are the best among the same type of models. Compared with the sub-optimal model Transformer, the performance improvements are 0.93%, 0.29%, and 0.79% respectively; in the data of the S_{1}-S_{9} Stations, the RMSE, MAE, and MAPE indicators of the proposed method are 58.07%, 36.54%, and 27.69%, and all indicators are the best among the comparison models.

[0184] Example 3

[0185] As Figures 2-3 shown, this example also proposes a distributed photovoltaic power generation charging pile system, including a photovoltaic power generation module, an energy storage module, a power monitoring module, a power distribution control module, a charging pile module, a communication module, and an electricity consumption management module.

[0186] The output of the photovoltaic power generation module is connected to the input of the energy storage module, and the output of the energy storage module is connected to the input of the power monitoring module. The output of the power monitoring module is connected to the input of the power distribution control module, and the output of the energy storage module is connected to the input of the charging pile module. The input of the communication module is connected to the outputs of both the power monitoring module and the power distribution control module, and the output of the communication module is connected to the cloud. The output of the power distribution control module is connected to the user's power system, and the input of the charging pile module is also connected to the output of the municipal power grid. The input of the power management module is connected to both the output of the power distribution control module and the output of the municipal power grid, and the output of the power management module is connected to both the input of the charging pile module and the power system. The power distribution control module is used to execute the solar photovoltaic power generation prediction method to predict the power generation for better subsequent allocation. It features high efficiency, accuracy, and intelligence, providing a foundation and guarantee for the formulation of the entire power generation plan and operation mode.

[0187] Introducing power generation forecasting has the following advantages:

[0188] Improving the economic efficiency of power plants: By accurately predicting future power generation, power dispatch can be rationally arranged to avoid power waste or shortages and improve the economic efficiency of power plants.

[0189] Optimize power plant operation and maintenance: Based on power generation forecasts, equipment maintenance plans can be adjusted in a timely manner to improve equipment lifespan and efficiency.

[0190] A photovoltaic (PV) power generation module consists of several photovoltaic panels installed on the roof. The module converts solar energy into electrical energy, which is then stored in an energy storage module. This energy storage module supplies power to the charging pile module and the user's electrical system. The charging pile module charges new energy vehicles. The electrical system includes lighting, air conditioning, and heating systems, providing lighting and heating for the user's home. For rural pitched roof tiled houses, the PV panels are installed directly on the sunny side of the roof, utilizing the roof angle and fixed to the sides. For rural single-story houses, a non-penetrating roof-penetrating counterweight method is used to fix the PV panels.

[0191] The power monitoring module monitors the real-time energy storage of the energy storage module and has minimum and maximum energy storage thresholds. When the energy storage reaches the minimum threshold, the power monitoring module transmits this information to the power management module, which then sends a control command to the distribution control module. The distribution control module disconnects the energy storage device's power supply and simultaneously activates the mains power supply. Conversely, when the energy storage reaches the maximum threshold, the power monitoring module transmits this information to the power management module, which then sends a control command to the distribution control module. The distribution control module disconnects the mains power supply and simultaneously activates the energy storage device's power supply. The purpose of setting the minimum threshold is to ensure that even in the event of a mains power failure, when the mains power cannot supply power to the user's home electrical system and charging pile module, the energy storage at this minimum threshold can provide emergency power, guaranteeing the normal operation of the charging pile module and the user's electrical system. The maximum threshold provides a node where the energy storage module can be used for power supply.

[0192] The power distribution control module can predict solar power generation and, through the power distribution control module and the power consumption management module, can uniformly allocate power consumption for charging piles and the power system in users' homes, thereby achieving effective energy management.

[0193] The communication module is either a Bluetooth wireless communication module or a WIFI wireless communication module. The communication module transmits the power information of the energy storage module of the photovoltaic energy storage charging pile system, as well as the power consumption information of the charging pile and the user's power system, to the cloud. Users can view this information in real time through a mobile app, so that they can understand the status of the charging pile and the power consumption information of the power system from the cloud.

[0194] The distributed photovoltaic power generation charging pile system proposed in this embodiment is a distributed solar power generation-new energy vehicle charging system, which has the following practical significance:

[0195] 1. Save building energy consumption; 2. Supplement the power grid's energy structure and alleviate grid pressure during peak electricity consumption periods; 3. Save energy and reduce emissions, thereby reducing carbon emissions; 4. Increase farmers' income.

[0196] Competitive advantage analysis:

[0197] The rural market is extremely vast; this is the first time that solar power generation and charging piles have been combined. The charging piles have high power and short charging time; the price is low, making them easily acceptable to farmers and easy to promote.

[0198] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting solar photovoltaic power generation, characterized in that, Includes the following steps: S1. Acquire satellite remote sensing cloud images and historical photovoltaic sequence data; S2. Construct a complex cloud feature capture module, which includes a 1D convolutional layer, a reshaping layer, a layer normalization layer, and a cloud convolutional attention module; the kernel size of the 1D convolutional layer is [missing information]. The cloud convolutional attention module includes parallel first, second, and third branches; it integrates satellite remote sensing cloud images of relevant regions and corresponding times. The data is input into the complex cloud feature capture module, which extracts features from both global and local information of the image, and outputs remote sensing cloud image features. ; S3. Construct a photovoltaic sensing attention module, which includes a photovoltaic frequency domain attention module and a photovoltaic spatial attention module; process historical photovoltaic sequence data. The input is fed into the photovoltaic sensing attention module, where it undergoes spatial and frequency domain feature extraction operations to output photovoltaic sensing attention features. ; S4. Construct a photovoltaic multimodal feature fusion module to integrate remote sensing cloud image features. Photovoltaic perception attention characteristics The input is fed into the photovoltaic multimodal feature fusion module, and after feature fusion operation, the photovoltaic multimodal fused features are output. ; S5. Construct a 3D attention module for photovoltaics to fuse multimodal features. The input is fed into the photovoltaic 3D attention module for multi-dimensional information capture, and the photovoltaic 3D attention features are output. ; S6. Construct a photovoltaic power generation prediction module, incorporating photovoltaic three-dimensional attention features. Inputting the data into the photovoltaic power generation prediction module yields the predicted photovoltaic power generation. .

2. The solar photovoltaic power generation prediction method according to claim 1, characterized in that, The 1D convolutional layer, reshaping layer, and layer normalization layer in step S2 specifically include: Satellite remote sensing cloud images of the relevant area and the corresponding time. The input is fed into the complex cloud feature capture module, and the first convolutional feature is obtained after passing through a 1D convolutional layer. The first convolutional feature The first convolutional features are reshaped by the remodeling layer. Spatial Dimensions Merged into sequence length Obtain the first reshaping feature The first reshaping feature After performing layer normalization operations, the first layer normalized features are obtained. The formula is expressed as follows: ， ， ， in, Indicates the altitude of cloud layers in satellite remote sensing images. Indicates the width of a satellite remote sensing cloud image. This indicates the number of channels in a satellite remote sensing cloud image. This indicates the number of channels in a satellite remote sensing cloud image after processing. Indicates the kernel size as 1D convolutional layer operations, This indicates a reshaping operation. This represents the transpose of a tensor matrix. Presentation layer normalization operation.

3. The solar photovoltaic power generation prediction method according to claim 2, characterized in that, The cloud convolutional attention module in step S2 specifically includes: Normalize the first layer of features The input is fed into the cloud convolutional attention module. In the first branch, the first layer of normalized features... The feature tensor of the first half of the channels is extracted through slicing to obtain the first slice feature. The first slice feature Expanding the channel through a linear layer The first extended feature is obtained by multiplying the original feature by 1. First extended feature After activation function Processing yields global features. The formula is expressed as follows: ， ， ， in, Indicates the feature Slice the first half of the channel count. and Let these represent the first and second learnable parameters in the linear layer, respectively. express Activation function operations; In the second branch, features Feature extraction after slicing The feature tensor of the last half of the channels is used to output the second slice feature. The second slice feature Expanding the channel through a linear layer This yields the second extended feature. The second extended feature After convolution kernel size is The 1D convolutional layer has a kernel size of [missing information]. 1D depthwise separable convolutional layers are used to obtain local features. The local features After layer normalization, the second layer normalized features are obtained. The formula is expressed as follows: ， ， ， ， in, Indicates the feature Slicing is performed on the last half of the channel count. and These represent the third and fourth learnable parameters in the linear layer, respectively. Indicates the kernel size as 1D convolutional layer, Indicates the kernel size as 1D depth-separable convolutional layers; In the third branch, features The channel is expanded through a linear layer to... This yields the third extended feature. Third extended feature The spatial dimensions were restored after the reshaping process. The second reshaping feature is obtained. The second reshaping feature After convolution kernel size is 2D depthwise separable convolutional layers, activation functions and the kernel size is The 2D convolutional layer is processed to obtain the second convolutional feature. The formula is expressed as follows: ， ， ， in, and These represent the fifth and sixth learnable parameters in the linear layer, respectively. Indicates features Spatial dimension restored to Operation; Indicates the kernel size as 2D convolutional layers, Indicates the kernel size as 2D depth-separable convolutional layers; global features With the second layer of normalized features Perform product fusion to obtain the first fusion feature. ; the first fusion feature With the second convolution feature By splicing and merging the features, we can obtain the cloud convolutional attention features. The satellite remote sensing cloud image Convolutional attention features of clouds By stitching and fusion, the features of the remote sensing cloud image are obtained. The formula is expressed as follows: ， ， ， in, This indicates a feature concatenation operation. This indicates a product fusion operation.

4. The solar photovoltaic power generation prediction method according to claim 3, characterized in that, The photovoltaic frequency domain attention module in step S3 specifically includes: Historical data of photovoltaic series After convolution kernel size 1D convolutional layer, and after Activation function processing, outputting third convolutional features The formula is expressed as follows: ， in, Indicates the kernel size as 1D convolutional layer; the third convolutional feature The low-frequency features are obtained by global low-frequency feature capture through the photovoltaic frequency domain attention module. High-frequency features are obtained by calculating using the residual method. ; for low-frequency characteristics and high frequency characteristics Weighted fusion is performed to obtain the photovoltaic frequency domain attention features. The formula is expressed as follows: ， ， ， in, Represented as features The sequence length, ; and These represent the first and second learnable weight parameters obtained through backpropagation during model training. The multiplication operation represents the matrix element multiplication; the photovoltaic frequency domain attention feature After convolution kernel size is The 1D convolutional layer yields the fourth convolutional feature. The fourth convolutional feature After slicing, the feature tensors of the first half of the channels are extracted to obtain the features of the third slice. The fourth convolutional feature After slicing, half of the channel-numbered feature tensors are extracted to obtain the fourth slice feature. The formula is expressed as follows: ， ， ， in, Representation of features The sequence length, Representation of features The number of channels; Specifically, it refers to... Slice along the first half of the channel dimension; Specifically, it refers to... Slice along the last half of the channel dimension.

5. The solar photovoltaic power generation prediction method according to claim 4, characterized in that, The photovoltaic spatial attention module in step S3 specifically includes: The third slice feature After global average pooling, the kernel size is... 1D depthwise separable convolution and Activation function to obtain fifth convolutional features The fifth convolutional feature Features of the third slice By splicing and merging, the second fusion feature is obtained. After global average pooling, the kernel size is... 1D depthwise separable convolution and Activation function to obtain the sixth convolutional feature The sixth convolutional feature With the second fusion feature By splicing and merging, the third fusion feature is obtained. The third fusion feature Features of the third slice By adding each element, we obtain the first comprehensive feature. The formula is expressed as follows: ， ， ， ， ， in, express Activation function This indicates a global average pooling operation. This indicates an element-wise addition operation. and Indicates the kernel size as The 1D depthwise separable convolution and convolution kernel size are The 1D depthwise separable convolution; similarly, the fourth slice feature After the above third slice feature Following the same steps, the second comprehensive feature is obtained. The first comprehensive feature Second comprehensive features By splicing and merging, photovoltaic sensing attention can be obtained. The formula is expressed as: .

6. The solar photovoltaic power generation prediction method according to claim 5, characterized in that, Step S4 specifically includes: The remote sensing cloud image features Photovoltaic perception attention characteristics Input into the photovoltaic multimodal feature fusion module, remote sensing cloud image features After convolution kernel size is 2D transposed convolutional layers for features Perform high-resolution restoration and apply ReLU activation function for non-linear processing to output the seventh convolutional feature. Photovoltaic sensing attention features With the seventh convolution feature Perform feature concatenation along the channel dimension to obtain the concatenated features. The splicing features are processed by a convolution kernel with a size of [missing value]. 2D depthwise separable convolution, then using The activation function is subjected to nonlinear weighting to obtain the photovoltaic multimodal fusion characteristics. The formula is expressed as follows: ， ， ， in, Represents the ReLU activation function. Indicates the kernel size as 2D transposed convolutional layer, Indicates the kernel size as 2D depthwise separable convolution.

7. The solar photovoltaic power generation prediction method according to claim 6, characterized in that, Step S5 specifically includes: The photovoltaic multimodal fusion feature The input is fed into the photovoltaic 3D attention module to fuse photovoltaic multimodal features. Use along the channel dimension The operation divides the data into 4 sub-features along the channels, with each sub-feature having 10 channels. ,in Features The number of channels is expressed by the following formula: ， in, This indicates a segmentation feature operation along the channel. These represent the first initial sub-feature, the second initial sub-feature, the third initial sub-feature, and the fourth initial sub-feature, respectively; the first initial sub-feature The first sub-feature is obtained by preserving sub-features through skip connections. ; the second initial sub-feature After convolution kernel size is 2D convolutional layers and using The activation function performs non-linear processing, followed by a convolution kernel with a size of [missing value]. The 2D depthwise separable convolutional layer yields the second sub-feature. ; the second sub-feature With the third initial sub-feature By splicing and blending, a fourth fusion feature is obtained. The fourth fusion feature After convolution kernel size is 2D convolutional layers and using The activation function performs non-linear processing, followed by a convolution kernel with a size of [missing value]. The 2D depthwise separable convolutional layer yields the third sub-feature. The third sub-feature With the fourth initial sub-feature By splicing and merging, the fifth fusion feature is obtained. The fifth fusion feature After convolution kernel size is 2D convolutional layers and using The activation function performs non-linear processing, followed by a convolution kernel with a size of [missing value]. The 2D depthwise separable convolutional layer yields the fourth sub-feature. The formula is expressed as follows: ， ， ， ， ， in, Indicates the kernel size as 2D convolutional layers, and These represent the kernel size as follows: The 2D convolutional layer and the kernel size are 2D depthwise separable convolutional layers Indicates the kernel size as 2D depthwise separable convolutional layers; sub-features The components are spliced ​​and merged, and the kernel size is [missing information]. 2D convolutional layers are used to obtain photovoltaic 3D attention features. The formula is expressed as follows: ， in, Indicates the kernel size as 2D convolutional layers.

8. The solar photovoltaic power generation prediction method according to claim 7, characterized in that, Step S6 specifically includes: Photovoltaic three-dimensional attention features The input is fed into the photovoltaic power generation prediction module to generate photovoltaic 3D attention features. Flattened into a one-dimensional vector, then passed through a fully connected layer and Activation function to calculate predicted photovoltaic power generation The formula is expressed as follows: ， in, This indicates a fully connected operation. This represents the vector flattening operation. and These represent the third and fourth learnable weight parameters obtained by backpropagation of the model, respectively.

9. The solar photovoltaic power generation prediction method according to claim 8, characterized in that, The mean squared error loss is used as the loss function, and the calculation process of this loss function is as follows: ， in, This represents the calculated value of the loss function. The number of data samples, This is expressed as the actual photovoltaic power value. The sample is The photovoltaic power value predicted by the model at that time.

10. A distributed photovoltaic power generation charging pile system, characterized in that, It includes photovoltaic power generation modules, energy storage modules, power monitoring modules, power distribution control modules, charging pile modules, communication modules, and power management modules; The output terminal of the photovoltaic power generation module is connected to the input terminal of the energy storage module, and the output terminal of the energy storage module is connected to the input terminal of the power monitoring module; the output terminal of the power monitoring module is connected to the input terminal of the power distribution control module, the output terminal of the energy storage module is connected to the input terminal of the charging pile module, the input terminal of the communication module is connected to the output terminals of the power monitoring module and the power distribution control module respectively, and the output terminal of the communication module is connected to the cloud; the output terminal of the power distribution control module is connected to the user's power system, the input terminal of the charging pile module is also connected to the output terminal of the municipal power grid, the input terminal of the power management module is connected to the output terminal of the power distribution control module and the output terminal of the municipal power grid, and the output terminal of the power management module is connected to the input terminal of the charging pile module and the power system; wherein, the power distribution control module is used to execute the solar photovoltaic power generation prediction method according to any one of claims 1-9.

Images