Photovoltaic output prediction method and system based on improved MTGNN model

Abstract

The invention discloses a photovoltaic output prediction method and system based on an improved MTGNN model, and the method comprises the steps: collecting historical photovoltaic output data and corresponding photovoltaic output feature data, carrying out the feature extraction of a timestamp part of the photovoltaic output data, carrying out the splicing, and obtaining a photovoltaic feature data set after the time dimension is expanded; using the photovoltaic feature data sets to train an improved MTGNN model, wherein the improved MTGNN model comprises a self-adaptive graph learning module, a time convolution module and a graph convolution module introducing a screening attention mechanism; obtaining photovoltaic output characteristic data in real time so as to obtain a corresponding normalized photovoltaic characteristic data set; and outputting a prediction model from the input model. According to the method, the information integration capability is improved by using the adaptive adjacency matrix and the attention screening mechanism, and the method has better tracking capability for non-stationary data.

Description

Technical Field

[0001] The present invention belongs to the technical field of photovoltaic prediction, and more specifically, relates to a photovoltaic output prediction method and system based on an improved MTGNN model. Background Art

[0002] Photovoltaic power generation, as a relatively mature new energy power generation method, plays a key role in carbon emission reduction. However, photovoltaic output is closely related to light intensity, weather and other conditions, and has strong randomness and intermittency. In order to ensure the safe operation of the power system, accurate prediction of photovoltaic output is of great significance to the analysis of the new energy carrying capacity of the distribution network.

[0003] CN117081049A proposes a photovoltaic power station short-term power intelligent prediction method based on a multivariate time series graph neural network (Time-Extrapolator Multivariate Time Series Forecasting With Graph Neural Networks, TXP-MTGNN) model with multi-time series optimization and time extrapolation convolution module. The method comprises: preprocessing missing data and abnormal data in active power data and meteorological factor data to obtain target active power data and target meteorological factor data; constructing a data set of the TXP-MTGNN photovoltaic power station short-term power prediction model according to the target active power data and the target meteorological factor data; constructing the TXP-MTGNN photovoltaic power station short-term power prediction model, inputting the training data set into the TXP-MTGNN photovoltaic power station short-term power prediction model for training; inputting the verification data set into the first prediction model to obtain a first prediction result and verifying the accuracy of the first prediction result; and performing value supplementation processing on the second prediction result according to the characteristics of the photovoltaic power station itself. However, photovoltaic data usually contains a year-month-day-hour-minute-second timestamp, and the patent does not take into account the characteristics of the timestamp, which will result in the photovoltaic data time series being not tightly connected. Summary of the invention

[0004] In order to solve the deficiencies in the prior art, the present invention provides a photovoltaic output prediction method and system for multiple data time series based on an improved and optimized multivariate time series graph neural network.

[0005] The present invention adopts the following technical solution.

[0006] The first aspect of the present invention proposes a photovoltaic output prediction method based on an improved MTGNN model, comprising: Collect historical photovoltaic output data and corresponding photovoltaic output feature data, extract the timestamp features of the photovoltaic output feature data, normalize them, and then splice them with the corresponding photovoltaic output feature data to obtain the photovoltaic feature data set after expanding the time dimension and normalize them; use these photovoltaic feature data sets to train an improved MTGNN model, which includes adding an adaptive graph learning module, a time convolution module, and a graph convolution module that introduces a screening attention mechanism; obtain photovoltaic output feature data in real time to obtain the corresponding photovoltaic feature data set, input it into the adaptive graph learning module, and establish an adaptive adjacency matrix by calculating the mutual information matrix; perform 1*1 convolution on the photovoltaic feature data set and then input it into the time convolution module to output the initial graph structure transformed from the photovoltaic feature data set and the node features as the original state information, input the original state information and the adaptive adjacency matrix into the graph convolution module that introduces the screening attention mechanism, and output the predicted photovoltaic output data.

[0007] Preferably, the photovoltaic output characteristic data includes humidity, component temperature, global radiation, direct radiation, scattered radiation, ambient temperature and air pressure.

[0008] Preferably, the features are extracted from the timestamp of the photovoltaic output characteristic data, specifically: The timestamp of the photovoltaic output characteristic data is extracted into five features: minutes per hour, hours per day, days per week, days per month, and days per year, and all are normalized to between [-1,0]; The normalization formula is:

[0009] Where: represents the sth feature extracted from the timestamp, s={1,2,3,4,5}, , , , , Respectively represent the five features extracted: minutes per hour, hours per day, days per week, days per month, and days per year; Represents the data corresponding to the sth feature extracted by the timestamp, Indicates the maximum value of the data corresponding to the sth feature extracted by the timestamp, , , , , They are 60, 7, 24, 30, and 365 respectively.

[0010] Preferably, the normalized photovoltaic feature data set is input into an adaptive graph learning module, and an adaptive adjacency matrix is ​​established by calculating the mutual information matrix, specifically: Each data in the photovoltaic characteristic data set is taken as a node, and the mutual information of each pair of data is taken as an edge, so as to convert the photovoltaic characteristic data set into a graph form; The mutual information calculation formula is:

[0011] Where: and are two data in the photovoltaic characteristic data set, Representation data and data The mutual information of Respectively represent data and data The probability distribution of p ( x,y ) indicates data and data The joint probability distribution of Calculate the mutual information value between all the data in the photovoltaic characteristic data set and get the mutual information matrix M Im ;matrix M Im The element in the p-th row and q-th column of is the mutual information between the p-th element and the q-th element of the photovoltaic characteristic data set; The connection relationship between each two nodes of the graph converted from the photovoltaic feature data set is calculated based on mutual information, and the connection relationship is formed into an adaptive adjacency matrix for:

[0012] Where: and To initialize the embedding features; and represents the weight matrix obtained by embedding features of nodes; and is a model parameter; α is a parameter that controls the saturation rate of the activation function tanh; relu and tanh are activation functions; represents the Hadamard product; The mutual information matrix calculated for each pair of data.

[0013] Preferably, the graph convolution module is divided into an information transmission part and an information selection part. The hidden information of each layer is obtained in the information transmission part, and an attention mechanism is added to filter the hidden information. The hidden information of each layer is spliced ​​in the information selection part. The formula for obtaining the hidden information of each layer in the information transmission part is:

[0014]

[0015]

[0016] Where: represents the hidden information of the kth layer calculated without the attention mechanism. represents the learnable weight matrix, β is the set hyperparameter, k represents the number of transfer layers of the graph convolution module, σ represents the sigmoid activation function, Represents the original status information, represents the hidden information of k-1 layers, represents the degree matrix, represents the adjacency matrix, represents the identity matrix, A ij express The jth element in the i-th row of express The ith element of the ith row of .

[0017] Preferably, the adding of the attention mechanism to filter the hidden information is specifically as follows: Through the softmax activation function, the proportional coefficient is calculated:

[0018] Where: softmax is the activation function for normalization; Perform Hadamard product operation on the scale coefficient and the hidden state information of the k-1 layer, and then use the sigmoid activation function to obtain the output hidden state information of the kth layer. The formula is:

[0019] Where: is the ratio coefficient, H (k) is the output hidden state of the kth layer filtered by the attention mechanism, H (k-1) Represents the output hidden state of the k-1th layer.

[0020] Preferably, the information selection part concatenates the output state information of each layer, specifically:

[0021] Where: H (k-1) represents the output hidden state of the k-1th layer, H (0) Equal to the original state information H in ， Hout Represents the output of the graph convolution module, represents splicing, W out It is the weight matrix of the PV prediction results and the output hidden state, and the weight matrix is ​​continuously updated during the propagation process.

[0022] Preferably, the time convolution module adopts hole convolution, and the photovoltaic feature data set is convolved through the filter of the time convolution module after 1*1 convolution processing to obtain the data time feature, and then the data time feature is filtered and screened by tanh and sigmoid activation functions respectively, and the screened result is passed to the graph convolution module after a Hadamard product operation.

[0023] Preferably, the convolution is performed by the filter of the time convolution module, specifically: In the temporal convolution module, iterative calculations are performed through filters. Four filters are used, namely, filters with convolutions of 1*2, 1*3, 1*6 and 1*7. The initial expansion factor of the iteration is set to 1 and the expansion factor of each layer increases exponentially, with an exponent of a, and a>1. The convolution kernel size is q, and the receptive field of the atrous convolution of the nth layer iteration is:

[0024] The formula for convolution is:

[0025] Where: represents the temporal convolution module l The input sequence of the layer, Indicates that the convolution size is 1× h The filter h ={2,3,6,7}; Indicates l The convolution size of the layer is 1× h The output sequence of the filter elements; is the void coefficient; represents the u-th element of the convolution kernel, represents the interval of the dilated convolution, express No. elements; According to the largest filter, the outputs of all filters are truncated to the same length and cascaded in the channel dimension. The formula is expressed as:

[0026] Where: Indicates l-1 layer input, * indicates convolution, and concat indicates concatenation operation.

[0027] The second aspect of the present invention proposes a photovoltaic output prediction system based on an improved MTGNN model using the method described in the first aspect of the present invention, comprising an acquisition module, a timestamp feature extraction module, a model building module, and a prediction module, characterized in that: Acquisition module: used to collect historical photovoltaic output data and obtain the corresponding photovoltaic output characteristic data; obtain photovoltaic output characteristic data and its corresponding timestamp in real time; The timestamp feature extraction module is used to extract features from the timestamp of the photovoltaic output feature data, normalize the extracted timestamp features and splice them with the corresponding photovoltaic output feature data to obtain a photovoltaic feature data set after the time dimension is expanded, and normalize each photovoltaic feature data set; Model building module: used to take all photovoltaic feature data sets and corresponding historical photovoltaic output data as training sets, and input them into the improved MTGNN model for training. The improved MTGNN model includes an adaptive graph learning module, a temporal convolution module, and a graph convolution module that introduces a screening attention mechanism; Prediction module: used to obtain the corresponding normalized photovoltaic feature data set; input it into the adaptive graph learning module, and establish an adaptive adjacency matrix by calculating the mutual information matrix; after performing 1*1 convolution on the photovoltaic feature data set, input it into the time convolution module to capture the temporal periodicity of the data and output the original state information, and input the original state information and the adaptive adjacency matrix into the graph convolution module that introduces the screening attention mechanism to output the predicted photovoltaic output data.

[0028] The beneficial effects of the present invention are that, compared with the prior art, (1) the present invention extracts the timestamp into five features, splices them with the original features, supplements the time series features, makes the time series connection of photovoltaic feature data closer, effectively improves the prediction accuracy, and improves the generalization ability of the model.

[0029] (2) Based on the normalized timestamp features, the present invention establishes a graph structure and introduces a mutual information matrix into the graph learning module to establish an adaptive adjacency matrix under different influence relationships between data. The present invention uses an asymmetric adaptive adjacency matrix to improve the differences in the degree of influence between different data. The use of a screening attention mechanism allows the model to pay more attention to the output of the graph convolution hidden layer, improves the ability to integrate information, and has better tracking capabilities for non-stationary data. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 Schematic diagram for extracting timestamp features; Figure 2 This is the framework diagram of the graph convolution module; Figure 3 Diagram of the multi-data time series model framework for improved and optimized multivariate time series graphical neural networks. DETAILED DESCRIPTION

[0031] In order to make the purpose, technical scheme and advantages of the present invention clearer, the technical scheme of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. The embodiments described in this application are only embodiments of a part of the present invention, rather than all embodiments. Based on the spirit of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work belong to the protection scope of the present invention.

[0032] Embodiment 1 of the present invention proposes a photovoltaic output prediction method based on an improved MTGNN model, comprising: Collect historical photovoltaic output data, obtain the corresponding photovoltaic output characteristic data, extract features from the timestamps of the photovoltaic output characteristic data, normalize the extracted timestamp features and splice them with the corresponding photovoltaic output characteristic data to obtain a photovoltaic characteristic data set after expanding the time dimension, and normalize each photovoltaic characteristic data set; An improved MTGNN model is trained using these photovoltaic feature datasets. The model includes an adaptive graph learning module, a temporal convolution module, and a graph convolution module that introduces a screening attention mechanism. Acquire photovoltaic output characteristic data in real time to obtain the corresponding normalized photovoltaic characteristic data set; It is input into the adaptive graph learning module, and an adaptive adjacency matrix is ​​established by calculating the mutual information matrix; The photovoltaic feature dataset is subjected to 1*1 convolution and then input into the temporal convolution module to output the original state information. The original state information is the initial graph structure and node features transformed from the photovoltaic feature dataset. The original state information and the adaptive adjacency matrix are input into the graph convolution module that introduces the screening attention mechanism to output the predicted photovoltaic output data.

[0033] Preferably, the photovoltaic output characteristic data includes humidity, component temperature, global radiation, direct radiation, scattered radiation, ambient temperature and air pressure.

[0034] Preferably, the feature extraction from the timestamp of the photovoltaic output characteristic data is specifically as follows: The timestamp of the photovoltaic output characteristic data is extracted into five features: minutes per hour, hours per day, days per week, days per month, and days per year, and all are normalized to between [-1,0]; The normalization formula is:

[0035] Where: represents the sth feature extracted from the timestamp, s={1,2,3,4,5}, , , , , Respectively represent the five features extracted: minutes per hour, hours per day, days per week, days per month, and days per year; Represents the data corresponding to the sth feature extracted by the timestamp, Indicates the maximum value of the data corresponding to the sth feature extracted by the timestamp, , , , , They are 60, 7, 24, 30, and 365 respectively.

[0036] like Figure 1 As shown in the figure, the extracted timestamp features are spliced ​​with the original photovoltaic output feature data to obtain the photovoltaic feature data set:

[0037] Where: X Indicates photovoltaic output characteristic data; X new Represents the photovoltaic feature dataset; concat is a concatenation operation.

[0038] Specifically, this embodiment selects data from a photovoltaic power station in northwest my country, which has an environmental collection system. The data time range used by the training model of the present invention is from January 1, 2019 to June 5, 2019, and the data sampling time interval is 15 minutes. 80% is selected as the training set, 10% as the validation set, and 10% as the test set.

[0039] Preferably, the normalized photovoltaic feature data set is input into an adaptive graph learning module, and an adaptive adjacency matrix is ​​established by calculating a mutual information matrix, specifically: Each data in the photovoltaic characteristic data set is taken as a node, and the mutual information of each pair of data is taken as an edge, so as to convert the photovoltaic characteristic data set into a graph form; The mutual information calculation formula is:

[0040] Where: and are two data in the photovoltaic characteristic data set, Representation data and data The mutual information of Respectively represent data and data The probability distribution of p ( x,y ) indicates data and data The joint probability distribution of Calculate the mutual information value between all the data in the photovoltaic characteristic data set and get the mutual information matrix M Im ;matrix M Im The element in the p-th row and q-th column of is the mutual information between the p-th element and the q-th element of the photovoltaic characteristic data set; The connection relationship between each two nodes of the graph converted from the photovoltaic feature data set is calculated based on mutual information, and the connection relationship is formed into an adaptive adjacency matrix for:

[0041] Where: and To initialize the embedding features; and represents the weight matrix obtained by embedding features of nodes; and is a model parameter; α is a parameter that controls the saturation rate of the activation function tanh; relu and tanh are activation functions; represents the Hadamard product; The mutual information matrix calculated for each pair of data.

[0042] Preferably, if Figure 2 As shown, the graph convolution module is divided into an information transmission part and an information selection part. In the information transmission part, the hidden information of each layer is obtained, and the attention mechanism is added to filter the hidden information. In the information selection part, the hidden information of each layer is spliced; The formula for obtaining the hidden information of each layer in the information transmission part is:

[0043]

[0044]

[0045] Where: represents the hidden information of the kth layer calculated without the attention mechanism. represents the learnable weight matrix, β is the set hyperparameter, k represents the number of transfer layers of the graph convolution module, σ represents the sigmoid activation function, Represents the original status information, represents the hidden information of k-1 layers, represents the degree matrix, represents the adjacency matrix, represents the identity matrix, A ij express The jth element in the i-th row of express The ith element of the ith row of .

[0046] Preferably, the adding of the attention mechanism to filter the hidden information is specifically as follows: Through the softmax activation function, the proportional coefficient is calculated:

[0047] Where: softmax is the activation function for normalization; Perform Hadamard product operation on the scale coefficient and the hidden state information of the k-1 layer, and then use the sigmoid activation function to obtain the output hidden state information of the kth layer. The formula is:

[0048] Where: is the ratio coefficient, H (k) is the output hidden state of the kth layer filtered by the attention mechanism, H (k-1) Represents the output hidden state of the k-1th layer.

[0049] Preferably, the information selection part concatenates the output state information of each layer, specifically:

[0050] Where: H (k-1) represents the output hidden state of the k-1th layer, H (0) Equal to the original state information H in ， H out Represents the output of the graph convolution module, represents splicing, W out It is the weight matrix of the PV prediction results and the output hidden state, and the weight matrix is ​​continuously updated during the propagation process.

[0051] Preferably, the time convolution module adopts hole convolution, and the photovoltaic feature data set is convolved through the filter of the time convolution module after 1*1 convolution processing to obtain the data time feature, and then the data time feature is filtered and screened by tanh and sigmoid activation functions respectively, and the screened result is passed to the graph convolution module after a Hadamard product operation.

[0052] Preferably, the convolution is performed by the filter of the time convolution module, specifically: In the temporal convolution module, iterative calculations are performed through filters. Four filters are used, namely, filters with convolutions of 1*2, 1*3, 1*6 and 1*7. The initial expansion factor of the iteration is set to 1 and the expansion factor of each layer increases exponentially, with an exponent of a, and a>1. The convolution kernel size is q, and the receptive field of the atrous convolution of the nth layer iteration is:

[0053] The formula for convolution is:

[0054] Where: represents the temporal convolution module l The input sequence of the layer, Indicates that the convolution size is 1× h The filter h ={2,3,6,7}; Indicates l The convolution size of the layer is 1× h The output sequence of the filter elements; is the void coefficient; represents the u-th element of the convolution kernel, represents the interval of the dilated convolution, express No. elements; According to the largest filter, the outputs of all filters are truncated to the same length and cascaded in the channel dimension. The formula is expressed as:

[0055] Where: Indicates l -1 layer input, * indicates convolution, and concat indicates concatenation operation.

[0056] Embodiment 2 of the present invention proposes a photovoltaic output prediction system based on an improved MTGNN model using the method described in Embodiment 1 of the present invention, including an acquisition module, a timestamp feature extraction module, a model building module, and a prediction module, characterized in that: Acquisition module: used to collect historical photovoltaic output data and obtain the corresponding photovoltaic output characteristic data; obtain photovoltaic output characteristic data and its corresponding timestamp in real time; The timestamp feature extraction module is used to extract features from the timestamp of the photovoltaic output feature data, normalize the extracted timestamp features and splice them with the corresponding photovoltaic output feature data to obtain a photovoltaic feature data set after the time dimension is expanded, and normalize each photovoltaic feature data set; Model building module: used to take all photovoltaic feature data sets and corresponding historical photovoltaic output data as training sets, and input them into the improved MTGNN model for training. The improved MTGNN model includes an adaptive graph learning module, a temporal convolution module, and a graph convolution module that introduces a screening attention mechanism; Prediction module: used to obtain the corresponding normalized photovoltaic feature data set; input it into the adaptive graph learning module, and establish an adaptive adjacency matrix by calculating the mutual information matrix; after performing 1*1 convolution on the photovoltaic feature data set, input it into the time convolution module to capture the temporal periodicity of the data and output the original state information, and input the original state information and the adaptive adjacency matrix into the graph convolution module that introduces the screening attention mechanism to output the predicted photovoltaic output data.

[0057] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents, and any modifications or equivalent replacements that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A photovoltaic output prediction method based on an improved MTGNN model, characterized in that: include: Collect historical photovoltaic output data and corresponding photovoltaic output feature data, extract the timestamp features of the photovoltaic output feature data, normalize them, and then splice them with the corresponding photovoltaic output feature data to obtain the photovoltaic feature data set after expanding the time dimension and normalize them; use these photovoltaic feature data sets to train an improved MTGNN model, which includes adding an adaptive graph learning module, a time convolution module, and a graph convolution module that introduces a screening attention mechanism; obtain photovoltaic output feature data in real time to obtain the corresponding photovoltaic feature data set, input it into the adaptive graph learning module, and establish an adaptive adjacency matrix by calculating the mutual information matrix; perform 1*1 convolution on the photovoltaic feature data set and then input it into the time convolution module to output the initial graph structure transformed from the photovoltaic feature data set and the node features as the original state information, input the original state information and the adaptive adjacency matrix into the graph convolution module that introduces the screening attention mechanism, and output the predicted photovoltaic output data.

2. A photovoltaic output prediction method based on an improved MTGNN model according to claim 1, characterized in that: include: The photovoltaic output characteristic data includes humidity, component temperature, total radiation, direct radiation, scattered radiation, ambient temperature and air pressure.

3. A photovoltaic output prediction method based on an improved MTGNN model according to claim 2, characterized in that: include: Extract features from the timestamp of photovoltaic output characteristic data, specifically: The timestamp of the photovoltaic output characteristic data is extracted into five features: minutes per hour, hours per day, days per week, days per month, and days per year, and all are normalized to between [-1,0]; The normalization formula is: Where: represents the sth feature extracted from the timestamp, s={1,2,3,4,5}, , , , , Respectively represent the five features extracted: minutes per hour, hours per day, days per week, days per month, and days per year; Represents the data corresponding to the sth feature extracted by the timestamp, Indicates the maximum value of the data corresponding to the sth feature extracted by the timestamp, , , , , They are 60, 7, 24, 30, and 365 respectively.

4. The photovoltaic output prediction method based on the improved MTGNN model according to claim 1 is characterized in that: include: The normalized photovoltaic feature data set is input into the adaptive graph learning module, and the adaptive adjacency matrix is ​​established by calculating the mutual information matrix, which is as follows: Each data in the photovoltaic characteristic data set is taken as a node, and the mutual information of each pair of data is taken as an edge, so as to convert the photovoltaic characteristic data set into a graph form; The mutual information calculation formula is: Where: and are two data in the photovoltaic characteristic data set, Representation data and data The mutual information of Respectively represent data and data The probability distribution of p ( x,y ) indicates data and data The joint probability distribution of Calculate the mutual information value between all the data in the photovoltaic characteristic data set and get the mutual information matrix M Im ;matrix M Im The element in the p-th row and q-th column of is the mutual information between the p-th element and the q-th element of the photovoltaic characteristic data set; The connection relationship between each two nodes of the graph converted from the photovoltaic feature data set is calculated based on mutual information, and the connection relationship is formed into an adaptive adjacency matrix for: Where: and To initialize the embedding features; and represents the weight matrix obtained by embedding features of nodes; and is a model parameter; α is a parameter that controls the saturation rate of the activation function tanh; relu and tanh are activation functions; represents the Hadamard product; The mutual information matrix calculated for each pair of data.

5. A photovoltaic output prediction method based on an improved MTGNN model according to claim 4, characterized in that: include: The graph convolution module is divided into an information transmission part and an information selection part. The hidden information of each layer is obtained in the information transmission part, and the attention mechanism is added to filter the hidden information. The hidden information of each layer is spliced ​​in the information selection part. The formula for obtaining the hidden information of each layer in the information transmission part is: Where: represents the hidden information of the kth layer calculated without the attention mechanism. represents the learnable weight matrix, β is the set hyperparameter, k represents the number of transfer layers of the graph convolution module, σ represents the sigmoid activation function, Represents the original status information, represents the hidden information of k-1 layers, represents the degree matrix, represents the adjacency matrix, represents the identity matrix, A ij express The jth element in the i-th row of express The ith element of the ith row of .

6. A photovoltaic output prediction method based on an improved MTGNN model according to claim 5, characterized in that: include: The attention mechanism is added to filter the hidden information, specifically: Through the softmax activation function, the proportional coefficient is calculated: Where: softmax is the activation function for normalization; The Hadamard product operation is performed on the scale coefficient and the hidden state information of the k-1 layer, and then the output hidden state information of the kth layer is obtained through the sigmoid activation function. The formula is: Where: is the ratio coefficient, H (k) is the output hidden state of the kth layer filtered by the attention mechanism, H (k-1) Represents the output hidden state of the k-1th layer.

7. A photovoltaic output prediction method based on an improved MTGNN model according to claim 6, characterized in that: include: The information selection part concatenates the output status information of each layer, specifically: Where: H (k-1) represents the output hidden state of the k-1th layer, H (0) Equal to the original state information H in ， H out Represents the output of the graph convolution module, represents splicing, W out It is the weight matrix of the PV prediction results and the output hidden state, and the weight matrix is ​​continuously updated during the propagation process.

8. The photovoltaic output prediction method based on the improved MTGNN model according to claim 1 is characterized in that: include: The temporal convolution module adopts hole convolution. After the photovoltaic feature data set is processed by 1*1 convolution, it is convolved through the filter of the temporal convolution module to obtain the data time feature. The data time feature is filtered and screened by tanh and sigmoid activation functions respectively. The screened result is subjected to a Hadamard product operation and then passed to the graph convolution module.

9. A photovoltaic output prediction method based on an improved MTGNN model according to claim 8, characterized in that: include: The convolution is performed through the filter of the temporal convolution module, specifically: In the temporal convolution module, iterative calculations are performed through filters. Four filters are used, namely, filters with convolutions of 1*2, 1*3, 1*6 and 1*7. The initial expansion factor of the iteration is set to 1 and the expansion factor of each layer increases exponentially, with an exponent of a, and a>1. The convolution kernel size is q, and the receptive field of the atrous convolution of the nth layer iteration is: The formula for convolution is: Where: represents the temporal convolution module l The input sequence of the layer, Indicates that the convolution size is 1× h The filter, h ={2,3,6,7}; Indicates l The convolution size of the layer is 1× h The output sequence of the filter elements; is the void coefficient; represents the convolution kernel u elements, represents the interval of the dilated convolution, express No. elements; According to the largest filter, the outputs of all filters are truncated to the same length and cascaded in the channel dimension. The formula is expressed as: Where: Indicates l -1 layer input, * indicates convolution, and concat indicates concatenation operation.

10. A photovoltaic output prediction system based on an improved MTGNN model using the method according to any one of claims 1 to 9, comprising an acquisition module, a timestamp feature extraction module, a model building module, and a prediction module, characterized in that: Acquisition module: used to collect historical photovoltaic output data and obtain the corresponding photovoltaic output characteristic data; obtain photovoltaic output characteristic data and its corresponding timestamp in real time; The timestamp feature extraction module is used to extract features from the timestamp of the photovoltaic output feature data, normalize the extracted timestamp features and splice them with the corresponding photovoltaic output feature data to obtain a photovoltaic feature data set after the time dimension is expanded, and normalize each photovoltaic feature data set; Model building module: used to take all photovoltaic feature data sets and corresponding historical photovoltaic output data as training sets, and input them into the improved MTGNN model for training. The improved MTGNN model includes an adaptive graph learning module, a temporal convolution module, and a graph convolution module that introduces a screening attention mechanism; Prediction module: used to obtain the corresponding normalized photovoltaic feature data set; input it into the adaptive graph learning module, and establish an adaptive adjacency matrix by calculating the mutual information matrix; after performing 1*1 convolution on the photovoltaic feature data set, input it into the time convolution module to capture the temporal periodicity of the data and output the original state information, and input the original state information and the adaptive adjacency matrix into the graph convolution module that introduces the screening attention mechanism to output the predicted photovoltaic output data.

Images