A Photovoltaic Power Prediction Method Based on Multi-View Adaptive Feature Fusion

Abstract

This invention discloses a photovoltaic power generation prediction method based on multi-view adaptive feature fusion. This method can mine the potential correlations between information from four perspectives: global, local, temporal, and spatial, and use them for photovoltaic power generation prediction. The collected data is divided into global and local perspectives, and future total features and historical total features are extracted from temporal and spatial perspectives, respectively. In order to better fuse feature information, a parameter-sharing feature extraction module is constructed to extract common features, and an attention mechanism is introduced to assign appropriate attention weights to the local features of each time window, so that the entire model can adaptively learn the most relevant information from global and local features and fully fuse them. At the same time, consistency constraints and independence constraints are used to construct a loss function to strengthen the model, thereby accurately predicting photovoltaic power generation.

Description

Technical Field

[0001] This invention relates to the field of lighting equipment, and more particularly to a photovoltaic power generation prediction method based on multi-view adaptive feature fusion. Background Technology

[0002] Solar energy is a clean, safe, and sustainable power generation resource. After installing photovoltaic (PV) power generation systems near their power consumption sites, users can not only generate and consume their own electricity but also transmit surplus power to the national grid, balancing the power distribution system. This allows for full utilization of local solar resources, replacing and reducing fossil fuel consumption. However, PV power generation has several drawbacks: First, it relies on sunlight and only generates electricity during the day, making its generation intermittent. Second, PV power output fluctuates frequently due to changes in solar irradiance, temperature, and other meteorological factors. These two points indicate that PV power generation is significantly intermittent and volatile. This unstable power generation method and insufficient controllability of output power mean that large-scale PV grid integration can impact the safe and stable operation of the power grid. Therefore, accurate PV power generation forecasting is a key technology for solving this problem and a crucial guarantee for the safe and reliable operation of the power grid.

[0003] Currently, many scholars have conducted extensive research on the prediction of photovoltaic (PV) power generation, with methods mainly falling into two categories: direct prediction and indirect prediction. Direct prediction uses historical PV power plant data and meteorological data as input to build a mathematical model to predict PV power generation, ultimately calculating the predicted power output. Indirect prediction first predicts solar radiation from historical meteorological data, then, based on the PV power plant's power model, further builds a mathematical model to predict the power generation. Many researchers both domestically and internationally have conducted extensive academic research on PV power generation prediction based on artificial intelligence technologies such as machine learning and deep learning. Common machine learning prediction methods include the Autoregressive Integrated Moving Average (ARIMA) model; Support Vector Machine (SVM); and Markov Chain (MC), among others. These models can use windows of past information to predict the output of the next time step and learn the mapping relationship between input and output from a large number of historical samples. However, these models also ignore the mutual influence between features, i.e. the spatial correlation between features. Therefore, the prediction accuracy of photovoltaic power generation still needs to be improved. Summary of the Invention

[0004] To address the aforementioned problems, this invention provides a photovoltaic power generation prediction method based on multi-view adaptive feature fusion. The method extracts feature representations from the collected data from both global and local perspectives, following temporal and spatial perspectives respectively. Considering the potential common features in both global and local information, a shared feature extraction module is used to extract common features from both. An attention fusion module adaptively assigns appropriate attention weights to the local features of each time window. The resulting fused features are then processed by a multilayer perceptron to obtain the photovoltaic power generation prediction result.

[0005] To achieve the above objectives, this invention provides a photovoltaic power generation prediction method based on multi-view adaptive feature fusion, comprising:

[0006] S1: The number of features for local information is defined as follows: The time elapsed in each time window is Data granularity is Then the size of a time window is , The number of time windows determines the local information within each time window. The input is sent to the temporal feature extraction module, which uses a GRU gated recurrent unit for temporal feature extraction. The calculation formula is as follows:

[0007]

[0008] In the formula, Represents local information at time t , This indicates the hidden state at the previous moment. This indicates that the door is being reset. Indicates an update to the door. This represents the activation function. Represents the weight matrix. Represents the bias vector;

[0009] S2: According to the above , Calculate hidden state The calculation formula is as follows:

[0010]

[0011] In the formula, Indicates the candidate hidden state. This represents the hidden state at time t obtained after processing by the GRU module. This represents the activation function. This indicates element-wise operations. Represents the weight matrix. Represents the bias vector;

[0012] The hidden state Temporal characteristics of local information;

[0013] S3: The temporal features of the local information The data is input into the spatial feature extraction module, which obtains the historical total features of the local information. The temporal features of the local information of the temporal feature extraction module for w time windows are given. The spatial feature extraction module uses a graph convolutional network for spatial feature extraction, and its calculation formula is as follows:

[0014]

[0015] In the formula, Represents the adjacency matrix. Represents the identity matrix. Represented by the adjacency matrix With the identity matrix The reachability matrix obtained by addition, k This represents the number of layers in a graph convolutional network. The graph convolutional network represents the first... k Layer weight matrix, This represents the activation function. express The angle matrix; Indicates the first i The initial input to the graph convolutional network with each time window is also the output of the temporal feature module. , Indicates the first i The first time window k The feature matrix output by the layer graph convolution is defined as the historical feature. ,So for w Historical total features of local information from spatial feature extraction modules for each time window;

[0016] S4: Defines the number of features for global information as follows The time required to collect global information is Data granularity is global information The input is fed into the graph convolutional network of the spatial feature extraction module, and then... k The future total features obtained after layer learning are global information. The calculation formula is as follows:

[0017]

[0018] In the formula, Represents the adjacency matrix. Represents the identity matrix. Represented by the adjacency matrix With the identity matrix The reachability matrix obtained by addition, k This represents the number of layers in a graph convolutional network. The graph convolutional network represents the first... k Layer weight matrix, This represents the activation function. express The angle matrix, This represents the initial input to the graph convolutional network, which is also global information. , For the first k The feature matrix output by the graph convolutional network of layers is defined as the future total feature matrix. ;

[0019] S5: Transfer the local information and global information The input is fed into the parameter shared space extraction module to obtain shared historical features and shared future features. The parameter shared space extraction module uses a shared graph convolutional network for shared feature extraction, and its calculation formula is as follows:

[0020]

[0021] In the formula, Indicates the first k The weight matrix of the layer shared parameter module, , The meaning is the same as in step S3. and The meaning is the same as in step S4. This represents the activation function. Indicates the first i The first time window k The historical feature matrix output by the layer-shared graph convolutional network is defined as the shared historical feature. ; Indicates the first k The future feature matrix output by the layer-shared graph convolutional network is defined as the shared future feature matrix. ;

[0022] S6: The historical overall characteristics of the local information The future overall characteristics of global information Shared Future Features and shared historical features Input is fed into the attention fusion module; leveraging future total features Historical total characteristics for each time window The attention weights are assigned using the following formula:

[0023]

[0024] In the formula, Represents the attention function. They represent the general characteristics of history. Attention weights The specific calculation formula for the function is as follows:

[0025]

[0026] In the formula, Indicates the first i Historical characteristics of each time window Attention value in each dimension express a time window The matrix of attention values ​​formed Represents the weight matrix. The term tanh represents the bias term, and tanh(∙) represents the activation function. Represent the attention vector; then apply the softmax activation function to the attention value matrix. Normalize by column to obtain This leads to the final attention matrix. The calculation formula is as follows:

[0027]

[0028] In the formula, Indicates the first i The first window j Attention coefficients of historical features across all dimensions. Indicates the first i The attention vector of the total historical features of each window is The function's output, express The attention matrix is ​​composed of;

[0029] S7: Will Calculate the average value for each column separately. And combine them into a blend of historical features Then the attention matrix With integration of historical characteristics The historical feature vector after attention fusion is obtained by taking the inner product of the columns. ;

[0030] S8: Based on steps S6 and S7, the shared historical features are calculated. The corresponding attention matrix Then, the shared historical feature vector after fusing attention weights is obtained. ;

[0031] S9: Will , The vector is obtained by taking the mean of each column. , will the , , , The splicing is the final fusion feature. ;

[0032] S10: The fusion features Z The data is input into a multilayer perceptron to obtain the final output value of the photovoltaic power generation prediction model. The calculation formula is as follows:

[0033]

[0034] In the formula, This represents the predicted value of photovoltaic power generation. MLP This represents a multilayer perceptron, which consists of multiple fully connected layers;

[0035] S11: Overall objective function for constructing a photovoltaic power generation prediction model L :

[0036]

[0037] In the formula, The mean squared error (MSE) is used as the loss function for the prediction task. For consistency constraints, The multiplier corresponding to the consistency constraint. For independence constraints, The multiplier corresponding to the independence constraint;

[0038] S12: Make predictions and compare experimental results. Select different time granularities to predict the photovoltaic power generation for two consecutive days, and compare the predictions with the actual power to draw a curve. Use evaluation indicators to compare and evaluate the model's performance. The evaluation indicators are MAE, MAPE, and RMSE, and the formulas are as follows:

[0039]

[0040] In the formula, and These represent the predicted and actual values ​​of photovoltaic power generation, respectively. n The number of samples;

[0041] The local information includes at least one of the following: solar irradiance, temperature, humidity, soil moisture, CO2, air pressure, wind direction, instantaneous wind speed, 2-minute average wind speed, 10-minute average wind speed, evaporation, PM2.5, and cumulative daily radiation from historical meteorological data, as well as power plant power generation, DC voltage, DC current, electricity consumption voltage, battery voltage, and battery power from power plant power generation data.

[0042] The global information includes at least one of the following: future solar irradiance, temperature, humidity, air pressure, wind speed, wind direction, precipitation, and cloud cover from weather forecast data.

[0043] In the preferred embodiment, the loss function The calculation formula is:

[0044]

[0045] In the formula, and These represent the predicted and actual values ​​of photovoltaic power generation, respectively. n This represents the number of samples.

[0046] In the preferred embodiment, the consistency constraint is:

[0047]

[0048] In the formula, Indicates the number of time windows. This indicates that the KL divergence is used to measure the difference between two characteristic distributions.

[0049] In the preferred embodiment, the independence constraint is:

[0050]

[0051] In the formula, This indicates that the Hilbert-Schmidt independence index is used to measure the independence between the two characteristics.

[0052] The beneficial effects of this invention are as follows: It uses a series of deep learning techniques to extract temporal and spatial feature information from both global and local perspectives from existing data; it designs a parameter-sharing feature extraction module to extract common features, and uses an attention mechanism to adaptively fuse the feature representation with the learned weights, ultimately extracting the most relevant information from the network model—something traditional graph networks struggle to achieve; it introduces consistency and independence constraints to construct the loss function, and experiments verify the effectiveness of these two constraints, thus demonstrating that the model can solve the problem of low accuracy in photovoltaic power generation prediction. Attached Figure Description

[0053] Figure 1This is a flowchart of the photovoltaic power generation prediction method of the present invention;

[0054] Figure 2 This is a structural diagram of the GRU (Gate Circulation Unit) model of the present invention;

[0055] Figure 3 This is the attention weight allocation diagram of the present invention;

[0056] Figure 4 This is a comparison chart of the predicted power and the actual power using the method proposed in this invention. Detailed Implementation Specific Implementation

[0058] Feature extraction is performed on the raw data from both global and local perspectives. The collected data is divided into two categories: dynamic time-series data that changes over time (considered local information) and non-time-series data that does not change over time (considered global information). This invention uses historical meteorological data, which includes features such as solar irradiance, temperature, humidity, soil moisture, CO2, air pressure, wind direction, instantaneous wind speed, 2-minute average wind speed, 10-minute average wind speed, evaporation, PM2.5, and cumulative daily radiation; and power plant generation data, which includes features such as power plant output, DC voltage, DC current, electricity consumption voltage, battery voltage, and battery power. These two types of data are considered local information. Weather forecast data, which includes features such as future solar irradiance, temperature, humidity, air pressure, wind speed, wind direction, precipitation, and cloud cover, is used as global information. Different feature extraction methods are then applied to the data from these two perspectives to obtain their respective feature representations after information fusion.

[0059] like Figure 1 As shown, for historical meteorological data and power plant power generation data of local information, a series of deep learning techniques are used for feature extraction according to the perspective of "time series first, then space". The specific method is as follows:

[0060] S1: The number of features for local information is defined as follows: The time elapsed in each time window is Data granularity is Then the size of a time window is , The number of time windows determines the local information within each time window. The input is fed to the temporal feature extraction module, which typically employs an RNN, LSTM, or GRU gated recurrent unit network model. In this invention, the temporal feature extraction module uses a GRU gated recurrent unit for temporal feature extraction. Figure 2 As shown, the calculation formula is as follows:

[0061]

[0062] In the formula, Represents local information at time t , This indicates the hidden state at the previous moment. This indicates that the door is being reset. Indicates an update to the door. This represents the activation function. Represents the weight matrix. Represents the bias vector;

[0063] S2: According to the above , Calculate hidden state The calculation formula is as follows:

[0064]

[0065] In the formula, Indicates the candidate hidden state. This represents the hidden state at time t obtained after processing by the GRU module. This represents the activation function. This indicates element-wise operations. Represents the weight matrix. Represents the bias vector;

[0066] The hidden state Temporal characteristics of local information;

[0067] S3: The temporal features of the local information Input to the spatial feature extraction module, The temporal feature extraction module extracts local information for w time windows. The temporal features are further aggregated with spatial information to obtain the historical total features of the local information. This module typically uses a GNN, GAT, or GCN graph convolutional network model. The spatial feature extraction module uses a GCN graph convolutional network for spatial feature extraction, and its calculation formula is as follows:

[0068]

[0069] In the formula, Represents the adjacency matrix. Represents the identity matrix. Represented by the adjacency matrix With the identity matrix The reachability matrix obtained by addition, k This represents the number of layers in a graph convolutional network. The graph convolutional network represents the first... k Layer weight matrix, This represents the activation function. express The angle matrix; Indicates the first i The initial input to the graph convolutional network with each time window is also the output of the temporal feature module. , Indicates the first i The first time window k The feature matrix output by the layer graph convolution is defined as the historical feature. ,So for w Historical total features of local information from spatial feature extraction modules for each time window;

[0070] S4: Defines the number of features for global information as follows The time required to collect global information is Data granularity is global information The input is fed into the graph convolutional network of the spatial feature extraction module, and then... k The future total features obtained after layer learning are global information. The calculation formula is as follows:

[0071]

[0072] In the formula, Represents the adjacency matrix. Represents the identity matrix. Represented by the adjacency matrix With the identity matrix The reachability matrix obtained by addition, k This represents the number of layers in a graph convolutional network. The graph convolutional network represents the first... k Layer weight matrix, This represents the activation function. express The angle matrix, This represents the initial input to the graph convolutional network, which is also global information. , For the first k The feature matrix output by the graph convolutional network of layers is defined as the future total feature matrix. ;

[0073] Furthermore, a parameter-sharing spatial feature extraction module is designed to extract common feature representations from both perspectives. Since there are potential correlations between the features from the global and local perspectives, it is also necessary to extract their shared common features. This allows the model to be more adaptively applied to subsequent prediction tasks, specifically as follows:

[0074] S5: Transfer the local information and global information The input is fed into the parameter shared space extraction module to obtain shared historical features and shared future features. The parameter shared space extraction module uses a shared graph convolutional network for shared feature extraction, and its calculation formula is as follows:

[0075]

[0076] In the formula, Indicates the first k The weight matrix of the layer shared parameter module, , The meaning is the same as in step S3. and The meaning is the same as in step S4. This represents the activation function. Indicates the first i The first time window k The historical feature matrix output by the layer-shared graph convolutional network is defined as the shared historical feature. ; Indicates the first k The future feature matrix output by the layer-shared graph convolutional network is defined as the shared future feature matrix. ;

[0077] like Figure 3 As shown, an attention mechanism is then introduced. Based on the historical total features and shared historical features obtained in steps S3 and S5, different attention weights are assigned according to the time window. These are then weighted and fused to obtain fused features, which are used for the final photovoltaic power generation prediction. The attention mechanism allows the model to focus on information more critical to the current task from a large amount of input information, reducing attention to other information and filtering irrelevant information. Specifically:

[0078] S6: The historical overall characteristics of the local information The future overall characteristics of global information Shared Future Features and shared historical features Input is fed into the attention fusion module; leveraging future total features Historical total characteristics for each time window The attention weights are assigned using the following formula:

[0079]

[0080] In the formula, Represents the attention function. They represent the general characteristics of history. Attention weights The specific calculation formula for the function is as follows:

[0081]

[0082] In the formula, Indicates the first i Historical characteristics of each time window Attention value in each dimension express a time window The matrix of attention values ​​formed Represents the weight matrix. The term tanh represents the bias term, and tanh(∙) represents the activation function. Represent the attention vector; then apply the softmax activation function to the attention value matrix. Normalize by column to obtain This leads to the final attention matrix. The calculation formula is as follows:

[0083]

[0084] In the formula, Indicates the first i The first window j The attention coefficient of the total historical features in each dimension; the larger the value, the more attention the corresponding dimension's attention coefficient. i The first window j The more important the features in each dimension; Indicates the first i The attention vector of the total historical features of each window is The function's output, express The attention matrix is ​​composed of;

[0085] S7: Will Calculate the average value for each column separately. And combine them into a blend of historical features Then the attention matrix With integration of historical characteristics The historical feature vector after attention fusion is obtained by taking the inner product of the columns. ;

[0086] S8: Based on steps S6 and S7, the shared historical features are calculated. The corresponding attention matrix Then, the shared historical feature vector after fusing attention weights is obtained. ;

[0087] S9: Will , The vector is obtained by taking the mean of each column. , will the , , , The splicing is the final fusion feature. ;

[0088] S10: The fusion features Z The data is input into a multilayer perceptron to obtain the final output value of the photovoltaic power generation prediction model. The calculation formula is as follows:

[0089]

[0090] In the formula, This represents the predicted value of photovoltaic power generation. MLP This represents a multilayer perceptron, which consists of multiple fully connected layers;

[0091] S11: Overall objective function for constructing a photovoltaic power generation prediction model L :

[0092]

[0093] In the formula, The mean squared error (MSE) is used as the loss function for the prediction task. For consistency constraints, The multiplier corresponding to the consistency constraint. For independence constraints, The multiplier corresponding to the independence constraint;

[0094] The loss function The calculation formula is:

[0095]

[0096] In the formula, and These represent the predicted and actual values ​​of photovoltaic power generation, respectively. n This represents the number of samples.

[0097] In order to share the feature representation generated by the parameter sharing feature extraction module and Further decoupling between them, using KL Divergence is used to measure this constraint, and the independence constraint is:

[0098]

[0099] In the formula, Indicates the number of time windows. This indicates that the KL divergence is used to measure the difference between two characteristic distributions.

[0100] To ensure Different information can be captured between them. The Hilbert-Schmidt Independence Criterion (HSIC) is used to measure and enhance the difference between the two features. The independence constraint formula is as follows:

[0101]

[0102] In the formula, This indicates that the Hilbert-Schmidt independence index is used to measure the independence between the two features. An objective function with both consistency and independence constraints can adaptively fuse feature information during model training, thereby achieving more accurate photovoltaic power generation prediction.

[0103] S12: Make predictions and compare experimental results. This invention uses historical meteorological data, power plant power generation data, and weather forecast data from April 1, 2022 to June 30, 2022, provided by the Dalian Institute of Chemical Physics, Chinese Academy of Sciences, as our dataset. Three time granularities are set: 10 / 30 / 60 minutes. The method proposed in this invention is used to predict the photovoltaic power generation for two consecutive days, June 29 and 30, 2022, and compares the predicted power with the actual power. Figure 4 As shown, the prediction method performs well at different time granularities. To compare and evaluate the model's performance, MAE, MAPE, and RMSE are selected as evaluation metrics, as shown in the following formulas:

[0104]

[0105] In the formula, and These represent the predicted and actual values ​​of photovoltaic power generation, respectively. n The sample size is used to compare the model proposed in this invention, its three variant models (i.e., models without consistency constraints, independence constraints, or neither), and two other existing models, namely SVM and ARIMA. The evaluation index calculation results are shown in Table 1. It can be seen that the prediction results of the model proposed in this invention have smaller values ​​compared with SVM and ARIMA. The model with both consistency constraints and independence constraints has the best prediction performance, which verifies the effectiveness of the method proposed in this invention and shows superior prediction performance for the photovoltaic power generation prediction problem.

[0106] Table 1 Calculation results of three evaluation indicators

[0107]

[0108] The key points of this invention are: (1) constructing global and local feature representations of existing data from four perspectives: global, local, temporal, and spatial; (2) constructing a parameter-sharing feature extraction module to extract common features, and introducing an attention fusion module to adaptively allocate attention weights to local features in different time windows, thereby generating fused features for photovoltaic power generation prediction; (3) in order to enhance the performance of the model, designing a loss function with consistency constraints and independence constraints to make more accurate photovoltaic power generation prediction.

[0109] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented in various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0110] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0111] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0112] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0113] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0114] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A photovoltaic power generation prediction method based on multi-view adaptive feature fusion, characterized in that, It includes: S1: The number of features for local information is defined as follows: The time elapsed in each time window is Data granularity is Then the size of a time window is , The number of time windows determines the local information within each time window. The input is sent to the temporal feature extraction module, which uses a GRU gated recurrent unit for temporal feature extraction. The calculation formula is as follows: In the formula, Represents local information at time t , This indicates the hidden state at the previous moment. This indicates that the door is being reset. Indicates an update to the door. This represents the activation function. Represents the weight matrix. Represents the bias vector; S2: According to the above , Calculate hidden state The calculation formula is as follows: In the formula, Indicates the candidate hidden state. This represents the hidden state at time t obtained after processing by the GRU module. This represents the activation function. This indicates element-wise operations. Represents the weight matrix. Represents the bias vector; The hidden state Temporal characteristics of local information; S3: The temporal features of the local information The data is input into the spatial feature extraction module to obtain the historical total features of local information. The temporal features of the local information of the temporal feature extraction module for w time windows are given. The spatial feature extraction module uses a graph convolutional network for spatial feature extraction, and its calculation formula is as follows: In the formula, Represents the adjacency matrix. Represents the identity matrix. Represented by the adjacency matrix With the identity matrix The reachability matrix obtained by addition, k This represents the number of layers in a graph convolutional network. The graph convolutional network represents the first... k Layer weight matrix, This represents the activation function. express The angle matrix; Indicates the first i The initial input to the graph convolutional network with each time window is also the output of the temporal feature module. , Indicates the first i The first time window k The feature matrix output by the layer graph convolution is defined as the historical feature. ,So for w Historical total features of local information from spatial feature extraction modules for each time window; S4: Defines the number of features for global information as follows The time required to collect global information is Data granularity is global information The input is fed into the graph convolutional network of the spatial feature extraction module, and then... k The future total features obtained after layer learning are global information. The calculation formula is as follows: In the formula, Represents the adjacency matrix. Represents the identity matrix. Represented by the adjacency matrix With the identity matrix The reachability matrix obtained by addition, k This represents the number of layers in a graph convolutional network. The graph convolutional network represents the first... k Layer weight matrix, This represents the activation function. express The angle matrix, This represents the initial input to the graph convolutional network, which is also global information. , For the first k The feature matrix output by the graph convolutional network of layers is defined as the future total feature matrix. ; S5: Transfer the local information and global information The input is fed into the parameter shared space extraction module to obtain shared historical features and shared future features. The parameter shared space extraction module uses a shared graph convolutional network for shared feature extraction, and its calculation formula is as follows: In the formula, Indicates the first k The weight matrix of the layer shared parameter module, , The meaning is the same as in step S3. and The meaning is the same as in step S4. This represents the activation function. Indicates the first i The first time window k The historical feature matrix output by the layer-shared graph convolutional network is defined as the shared historical feature. ; Indicates the first k The future feature matrix output by the layer-shared graph convolutional network is defined as the shared future feature matrix. ; S6: The historical overall characteristics of the local information The future overall characteristics of global information Shared Future Features and shared historical features Input is fed into the attention fusion module; leveraging future total features Historical characteristics for each time window The attention weights are assigned using the following formula: In the formula, Represents the attention function. They represent the general characteristics of history. Attention weights The specific calculation formula for the function is as follows: In the formula, Indicates the first i Historical characteristics of each time window Attention value in each dimension express a time window The matrix of attention values ​​formed Represents the weight matrix. The term tanh represents the bias term, and tanh(∙) represents the activation function. Represent the attention vector; then apply the softmax activation function to the attention value matrix. Normalize by column to obtain This leads to the final attention matrix. The calculation formula is as follows: In the formula, Indicates the first i The first window j Attention coefficients of historical features across all dimensions. Indicates the first i The attention vector of the total historical features of each window is The function's output, express The attention matrix is ​​composed of; S7: Will Calculate the average value for each column separately. And combine them into a blend of historical features Then the attention matrix With integration of historical characteristics The historical feature vector after attention fusion is obtained by taking the inner product of the columns. ; S8: Based on steps S6 and S7, the shared historical features are calculated. The corresponding attention matrix Then, the shared historical feature vector after fusing attention weights is obtained. ; S9: Will , The vector is obtained by taking the mean of each column. , will the , , , The splicing results in the final fusion feature. ; S10: The fusion features Z The data is input into a multilayer perceptron to obtain the final output value of the photovoltaic power generation prediction model. The calculation formula is as follows: In the formula, This represents the predicted value of photovoltaic power generation. MLP This represents a multilayer perceptron, which consists of multiple fully connected layers; S11: Overall objective function for constructing a photovoltaic power generation prediction model L : In the formula, The mean squared error (MSE) is used as the loss function for the prediction task. For consistency constraints, The multiplier corresponding to the consistency constraint. For independence constraints, The multiplier corresponding to the independence constraint; S12: Make predictions and compare experimental results. Select different time granularities to predict the photovoltaic power generation for two consecutive days, and compare the predictions with the actual power to draw a curve. Use evaluation indicators to compare and evaluate the model's performance. The evaluation indicators are MAE, MAPE, and RMSE, and the formulas are as follows: In the formula, and These represent the predicted and actual values ​​of photovoltaic power generation, respectively. n The number of samples; The local information includes at least one of the following: solar irradiance, temperature, humidity, soil moisture, CO2, air pressure, wind direction, instantaneous wind speed, 2-minute average wind speed, 10-minute average wind speed, evaporation, PM2.5, and cumulative daily radiation from historical meteorological data, as well as power plant power generation, DC voltage, DC current, electricity consumption voltage, battery voltage, and battery power from power plant power generation data. The global information includes at least one of the following: future solar irradiance, temperature, humidity, air pressure, wind speed, wind direction, precipitation, and cloud cover from weather forecast data.

2. The photovoltaic power generation prediction method based on multi-view adaptive feature fusion according to claim 1, characterized in that, The loss function The calculation formula is: In the formula, and These represent the predicted and actual values ​​of photovoltaic power generation, respectively. n This represents the number of samples.

3. The photovoltaic power generation prediction method based on multi-view adaptive feature fusion according to claim 1, characterized in that, The consistency constraint is: In the formula, Indicates the number of time windows. This indicates that the KL divergence is used to measure the difference between two characteristic distributions.

4. The photovoltaic power generation prediction method based on multi-view adaptive feature fusion according to claim 1, characterized in that, The independence constraint is: In the formula, This indicates that the Hilbert-Schmidt independence index is used to measure the independence between the two characteristics.

Images